WEBVTT
00:00:00.047 --> 00:00:03.657
So Mike, we're gonna talk today about how you can't always trust what you can see online.
00:00:03.777 --> 00:00:07.466
That feels like a kind of running theme through Ctrl Alt Speech.
00:00:07.746 --> 00:00:09.787
Um, but
00:00:09.884 --> 00:00:13.064
But you can trust what you can trust when you listen to us.
00:00:13.105 --> 00:00:14.064
You can trust us
00:00:14.467 --> 00:00:15.836
you can, you can.
00:00:16.137 --> 00:00:17.339
That's a, high bar.
00:00:17.339 --> 00:00:19.000
It means we cannot make any mistakes today.
00:00:20.109 --> 00:00:20.579
Yes
00:00:20.579 --> 00:00:23.710
Um, not that I think anybody is fact-checking us.
00:00:23.943 --> 00:00:36.296
but you might not be able to trust everything, but there are some good things about the internet, and I wanted to ask you, what have you been watching or enjoying that makes you feel like the internet is worthwhile sticking with?
00:00:36.723 --> 00:00:37.253
Yes.
00:00:37.264 --> 00:00:49.371
Well, you know, I think one of the things that I think lots of us think a fun outgrowth of the internet is the rise of memes and meme culture and being able to share memes and things like that.
00:00:49.790 --> 00:00:56.546
So I found this thing which is brand new, it's only been out for a few days, called Meme Museum.
00:00:56.866 --> 00:01:08.754
So it's meme-museum.at, and it is built on the AT Protocol, which, you know, I am a supporter of the AT Protocol, ding, due
00:01:08.840 --> 00:01:09.620
please go bold.
00:01:09.843 --> 00:01:10.944
to, to Blue Sky.
00:01:11.243 --> 00:01:12.474
This is not Blue Sky.
00:01:12.483 --> 00:01:13.974
This is the AT Protocol.
00:01:14.343 --> 00:01:21.293
and there's actually a really interesting article from the guy who built it, this guy Joe Grmusica, I think is, is the name.
00:01:21.661 --> 00:01:27.590
and was just saying he had this idea of something he thought would be cool, which was a tool to help you manage your memes.
00:01:27.921 --> 00:01:35.861
So if you have, memes and you wanna be able to share them on social media, there are people who have, like, folders full of memes and things like that.
00:01:36.203 --> 00:01:38.784
But this is built into the AT Protocol.
00:01:39.180 --> 00:01:56.963
but he wrote a really interesting article about how it was just something that he had this thought and he could build it, and then it goes on about how nice it is to have that protocol that you can just build something and be able to, you know, build something into, a social media system, and it's a cool tool for managing all your memes.
00:01:57.043 --> 00:01:57.403
And I was
00:01:57.439 --> 00:01:58.019
I love that
00:01:58.161 --> 00:01:59.010
I like memes.
00:01:59.370 --> 00:02:05.667
I like being able to build on open protocols, and so this made me happy about the internet this week.
00:02:06.417 --> 00:02:07.037
How about you?
00:02:07.048 --> 00:02:10.354
mean, I, I've actually worked with Joe quite closely on a couple of projects.
00:02:10.784 --> 00:02:13.594
Yeah, he's done some great work, and, he's a really fun guy.
00:02:13.625 --> 00:02:17.170
So it's no surprise that he's done something cool, on the out protocol.
00:02:17.411 --> 00:02:20.131
And, and just before we move on, what, what is in your meme museum?
00:02:20.290 --> 00:02:22.871
I think people are gonna be wondering, what are you collecting?
00:02:23.067 --> 00:02:23.407
I am...
00:02:23.599 --> 00:02:27.366
I, well, I mean, I've just been playing around, so there's not that much in there yet.
00:02:27.387 --> 00:02:29.537
I need to, I need to start adding stuff to it.
00:02:29.787 --> 00:02:34.486
I have not been as good about some other people with, with my meme game.
00:02:34.747 --> 00:02:47.902
So maybe this will be my inspiration to up my meme game,'cause I'm-- I feel like memes are a little bit more of a millennial thing, and I am solidly in the Gen X group.
00:02:47.913 --> 00:02:54.836
But I feel sort of millennial at heart at times, and so this is my inspiration to maybe up my meme game
00:02:54.933 --> 00:02:55.283
Yeah.
00:02:55.343 --> 00:02:55.633
Nice.
00:02:55.633 --> 00:03:02.456
If you've got any, ideas, listeners, of memes that Mike should be curating in his museum, get in touch, podcast.ctrl-alt-speech.com.
00:03:02.877 --> 00:03:07.776
You have full unfettered rights to send him memes of all kinds.
00:03:09.396 --> 00:03:10.187
That's what I heard.
00:03:10.409 --> 00:03:10.800
Ben?
00:03:10.800 --> 00:03:11.599
What have you done?
00:03:13.056 --> 00:03:16.973
Um, well, I, My kind of fun thing this week, Mike, is, something for my son.
00:03:16.973 --> 00:03:28.790
Obviously, he made an appearance on the podcast a few weeks ago while he was having his bath and going to bed, and so You know, it's not the only interruption that I've had from him while sat at this very desk.
00:03:28.881 --> 00:03:41.591
He likes to, normally after a nap, come into my office, look at, you know, what I'm up to, and start bashing the keyboard, and that's, that's fairly problematic when you've got kind of documents and, emails up and things like that.
00:03:41.841 --> 00:03:44.131
You know, he manages to close a lot of windows.
00:03:44.270 --> 00:03:49.234
He manages to, you know, he managed to send an email at one point, which I was fairly horrified at.
00:03:49.614 --> 00:03:58.824
Um, so I've, I, I, I've found a tool, a website called tinyfingers.net, which is specifically created for this purpose.
00:03:59.093 --> 00:04:14.491
It is a full-screen website where babies can come along, and they can smash the keyboard to their heart's content, and it brings up little, you know, kind of emojis and emoticons and, has a nice kind of snowscape pattern as well, and I've used it a couple of times.
00:04:14.491 --> 00:04:21.721
It works incredibly well for keeping him happy and, uh, s- avoids any kind of professional, issues on my side.
00:04:21.721 --> 00:04:23.141
So great all round
00:04:23.428 --> 00:04:23.889
Yeah.
00:04:23.992 --> 00:04:30.543
I mean, I'm looking at it right now and it says 1,072 kids are smashing keyboards right now.
00:04:30.903 --> 00:04:39.528
Now I would be remiss if I didn't question the trust and safety implications of a website like this.
00:04:39.528 --> 00:04:39.538
But
00:04:39.788 --> 00:04:43.428
the kind of screen time social media ban, impetus
00:04:43.536 --> 00:04:44.257
Exactly.
00:04:44.257 --> 00:04:52.112
Like, does this comply with, uh, you know, is this site covered by the, bans on social media for kids?
00:04:52.362 --> 00:04:59.293
I mean, this is definitely for kids under the age of 16, and so I, I have to wonder about some of that.
00:05:00.536 --> 00:05:01.877
Look, look, don't get me started.
00:05:12.242 --> 00:05:20.826
Hello, and welcome to Ctrl-Alt-Speech, the podcast where we make sense of the major debates shaping online speech, platform power, content moderation, and the future of the internet.
00:05:21.185 --> 00:05:37.153
It's August the 13th, 2026, and this week we're discussing why everyone is freaking out about Anthropic's watermarking announcement, what's next for AI ethics, and for our Patreon supporters, how misinformation took hold in a tiny Spanish enclave a few weeks ago.
00:05:37.622 --> 00:05:44.252
Ctrl-Alt-Speech is the weekly moral audit of the internet, co-hosted f- fairly grand.
00:05:45.403 --> 00:05:46.043
I like it.
00:05:46.053 --> 00:05:46.634
I like it
00:05:47.033 --> 00:05:52.983
co-hosted by Mike Masnick, founder and editor of Techdirt, and me, Ben Whitelaw, the founder and editor of Everything in Moderation.
00:05:53.242 --> 00:05:55.423
Mike, you already know the big news of this week.
00:05:55.670 --> 00:05:56.925
the Bluesky bell is back.
00:05:57.752 --> 00:05:58.403
I heard it.
00:05:58.603 --> 00:05:59.083
Ding, ding
00:05:59.906 --> 00:06:03.702
Yeah, perhaps used slightly incorrectly, a bit rusty.
00:06:04.151 --> 00:06:09.521
We, we haven't used it in a while, so, so w- if we have new listeners, you should explain the Blue Sky Bell
00:06:09.666 --> 00:06:20.769
The Bluesky bell is, well, where it's been is, is in a box in a, a room of my house while we have been renovating some, parts of our London, house.
00:06:20.790 --> 00:06:23.290
And so it has been fairly absent, I would say.
00:06:23.290 --> 00:06:35.226
But it's designed, when present, to highlight the fact that Mike is a board member of Bluesky and the, you know, potential, let's say kind of implications of you talking about a story related to Bluesky.
00:06:35.297 --> 00:06:37.720
Not that you're biased, but that you have priors.
00:06:38.499 --> 00:06:43.228
Do not trust anything, I say, because I am, I am fully biased.
00:06:43.771 --> 00:06:49.942
And, and the bell is, is the first of, I think, a, a series of different, um yeah, I just like the sound.
00:06:50.201 --> 00:06:54.742
Um, a f- a number of different instruments, noises that we're gonna kind of hopefully introduce.
00:06:54.781 --> 00:06:56.922
We've been talking about it for a while, but I'm quite committed to it
00:06:57.007 --> 00:07:00.187
I am still waiting for the trade-offs tuba, so
00:07:01.091 --> 00:07:04.461
I've got a, I've got another proposal actually for, for later on in the podcast.
00:07:04.461 --> 00:07:06.942
So yeah, we're starting, to bring that back.
00:07:07.541 --> 00:07:35.869
before we jump into, new production noises for the podcast and stories that we think you should all be listening to, people might be tired of me saying, but we are very, very grateful to our very dedicated group of Patreon supporters who pay either£10 a month to get, excellent extended episodes and early access to them as well, as well as our 20 a month founder and insider supporters who get both of those things and the opportunity to suggest stories for us to cover.
00:07:36.266 --> 00:07:41.555
a number of those supporters actually got in touch with us in recent weeks to do that, and we covered those stories.
00:07:41.886 --> 00:07:45.185
So it's a very, very, valuable thing, and we couldn't do the podcast without it.
00:07:45.523 --> 00:07:56.696
we monetize both through the Patreon supporters and also sponsorships, and it really helps us cover, the many, many articles we read and the production of it and our brilliant editor, Lee.
00:07:56.696 --> 00:08:00.615
So it's all going to a, a very, very good cause.
00:08:00.942 --> 00:08:07.048
if you are interested in sponsoring the podcast, you can find out more, If you go to our website, ctrlaltspeech.com, there is more information there.
00:08:07.889 --> 00:08:16.012
And if you are tempted by the Patreon, we have a very, very good offer right now, an almost free first month, with the discount code TRUSTCON.
00:08:16.271 --> 00:08:29.915
Use all caps, T-R-U-S-T C-O-N, and that gets you, this week a very, very interesting kind of picking over this big mass migration event in the Spanish city of Ceuta and how platforms are responding to it.
00:08:30.305 --> 00:08:31.925
So have a look out for that.
00:08:32.331 --> 00:08:35.072
Mike, where do we, where do we begin this week?
00:08:35.375 --> 00:08:37.595
I think we need to watermark this conversation
00:08:38.440 --> 00:08:41.610
In, a few years' time it probably will be, and we, we, we...
00:08:41.759 --> 00:08:50.043
People will know that there was no AI involved in our, inadequate words and attempted analysis.
00:08:50.413 --> 00:08:55.842
but, you know, there is this kind of growing question of, of what can you trust online, what is real?
00:08:56.273 --> 00:09:02.850
And we're gonna talk a bit about fact-checking later, which I think was kind of the original attempt to help validate what people saw on the web.
00:09:03.102 --> 00:09:10.355
but we have been waiting for the AI labs to really come up with their own play on this, and Anthropic this week announced something that I think fits neatly into that bucket.
00:09:10.807 --> 00:09:11.107
Yeah.
00:09:11.107 --> 00:09:22.591
So, you know, this sort of all traces back to the EU's AI Act, which, you know, I think has been seen as sort of, a really early jump into the space of regulating AI in, in all sorts of ways.
00:09:22.902 --> 00:09:29.312
And one element of that is a requirement for AI produced content to be watermarked in some way.
00:09:29.692 --> 00:09:43.761
you can understand the sort of thinking behind that in that there was a lot of concern, you know, a lot of the initial concerns about generative AI tools was that it was generating content that would fool people into believing that it was real.
00:09:43.782 --> 00:09:45.331
I mean, the, the deepfakes, right?
00:09:45.331 --> 00:09:56.361
We spent years talking about the threat and risk of deepfakes and how they were fooling people, except that there was maybe not that much evidence of people actually being fooled by deepfakes.
00:09:56.466 --> 00:10:05.481
in fact, it seemed like the bigger problem was people doing what's been known as the liar's dividend, where they would claim when they were really caught doing something that it was a deepfake.
00:10:05.782 --> 00:10:17.451
but so built into the EU Act is this, requirement that AI companies somehow watermark or indicate content that is created by AI.
00:10:17.861 --> 00:10:19.971
Now, there have been efforts underway.
00:10:20.254 --> 00:10:40.357
there's a, a standard which is now pretty widely being adopted for image and audio and video AI generated content, that, from what I've heard is, getting pretty good and is fairly hard to, break, though I imagine people will come up with all sorts of ways to break it, over time.
00:10:40.607 --> 00:10:52.888
But what was trickier was the idea of, what AI is most used for, which is generating text, and how do you indicate and mark that text is, originally generated by AI.
00:10:53.524 --> 00:11:09.195
So Anthropic announced this week that they are now, very soon starting to watermark all AI text that they generate and/or process, which is a, an interesting distinction that is worth digging a little deeper into.
00:11:09.565 --> 00:11:29.298
But basically what they're doing is they have put some sort of, at this point still hidden kind of bias into their text generation that will be recognizable to a machine-readable system to indicate that the text has been processed by Anthropic's tools.
00:11:29.597 --> 00:11:45.662
And so at some point they will make an API available that will allow you to take any text and then ship it back to Anthropic, and Anthropic will give you some sort of score as to whether or not that text was processed by Claude in, in some form or another.
00:11:46.022 --> 00:11:50.885
Now, this, there's the, the regulatory aspect of it, which is that they're sort of required to do this.
00:11:50.895 --> 00:11:58.798
It-- as people were talking about this, it came out that apparently Gemini, Google's, AI model, has been doing this for some time already.
00:11:58.928 --> 00:12:03.038
at some point, I assume that OpenAI will release something similar.
00:12:03.437 --> 00:12:05.272
the reactions to this have been all over the map.
00:12:05.481 --> 00:12:08.581
a lot of people are really, really mad about it.
00:12:08.855 --> 00:12:15.282
some for perhaps reasonable reasons, some I think for less than reasonable reasons.
00:12:15.591 --> 00:12:17.361
They are very much...
00:12:17.751 --> 00:12:25.339
You know, there are people who have-- This may surprise you, Ben,'cause I know you, you think well of people in the world and think people are generally good.
00:12:25.695 --> 00:12:36.511
there are some people who use AI to cheat in some form or another, and it is possible that a tool like this might make it harder for them to get away with the, the cheating side of what they do.
00:12:36.912 --> 00:12:40.211
and so some people I think are just mad about that aspect of it.
00:12:40.437 --> 00:12:45.589
just on that, I have a very interesting s- kind of very brief tangent about exactly that point.
00:12:46.004 --> 00:12:52.721
Cause I think a lot of the reaction has been from, like, people who are using AI models for the LinkedIn thought leadership, right?
00:12:52.971 --> 00:12:59.292
You know, there's, there's-- We talked about the AI slop button that's been created on LinkedIn.
00:12:59.591 --> 00:13:00.731
LinkedIn is the worst for it.
00:13:01.121 --> 00:13:15.779
This week, I f- saw two posts that were very, very similar in nature, bar a few kind of structural differences, and called out the fact that these posts written six or seven hours apart were very, very similar.
00:13:16.179 --> 00:13:34.234
And it turned out, long story short, that one of the two people had fed all of the, posts about actually this topic that we're talking about now, anthropic watermarking, had fed all of those into Claude and asked for a, had asked for a take that he could post onto LinkedIn.
00:13:34.716 --> 00:13:35.476
Wow
00:13:35.798 --> 00:13:49.082
I, to be honest, I don't believe that his page isn't running, isn't being run by Claude or OpenClaude or something, because his response to my questions was very, very, sounded very much like a, an AI, chatbot.
00:13:49.679 --> 00:14:00.731
The DM that I got from him sounded very much like an AI chatbot, and frankly, I would love some sort of watermarking in that moment because these two posts were incredibly similar.
00:14:01.168 --> 00:14:07.105
he made up some sort of, reasoning for why, the post is incredibly similar, but he had essentially kind of plagiarized it using AI.
00:14:07.698 --> 00:14:13.568
did you click the, the new LinkedIn button that allows you to, to claim that certain content is AI slop?
00:14:13.775 --> 00:14:15.846
didn't even think to do that, which is
00:14:16.102 --> 00:14:17.822
You well, you should, right?
00:14:17.832 --> 00:14:23.625
That is the new feature on LinkedIn is that you can, you can say,"I think this is AI slop," and, LinkedIn may do
00:14:23.671 --> 00:14:23.980
Yeah.
00:14:24.181 --> 00:14:26.980
But the point is, you know, as you say, it is very, very prevalent
00:14:27.419 --> 00:14:27.730
Yeah.
00:14:28.149 --> 00:14:32.912
there are certain cases and, and, you know, we're seeing more and more examples of it.
00:14:33.192 --> 00:14:55.859
there was a discussion in a, private group chat, so I'm gonna kinda be a little careful about what I say here, about a certain paper that recently came out that is relevant to a lot of our interests, where there was very clear indications that, at least the summary of the paper was written entirely by AI.
00:14:56.298 --> 00:15:00.859
It had a lot of hallmarks of AI drafted nature.
00:15:00.869 --> 00:15:03.158
And so, like, now people are sort of looking at this thing.
00:15:03.158 --> 00:15:18.664
Now, the flip side of that is that, that leads people to, like, completely freaking out about these things and maybe overreacting and assuming that things are AI generated or AI slop, as people would like to say, when it's not AI slop.
00:15:18.735 --> 00:15:24.996
And so you run the r- the real risk of, I mean, you have sort of, type one and type two errors, right?
00:15:25.017 --> 00:15:31.480
The false negatives and false positives where, things that aren't AI generated are going to be accused of being AI generated.
00:15:31.807 --> 00:15:35.615
things that are AI generated are not going to be, caught up in this.
00:15:35.846 --> 00:15:36.666
And it just become...
00:15:36.676 --> 00:15:39.186
You have this sort of witch hunt aspect of it.
00:15:39.628 --> 00:16:03.961
And what makes this worse, and this is part of the way that Anthropic has rolled this out, which I understand technically they probably feel, not doing it this way would be even harder, is that even though the, AI Act is fairly clear that the watermarking requirements or the transparency requirements don't need to apply to merely processing content.
00:16:03.981 --> 00:16:12.371
So if you're doing, like, a spell check or a grammar check, which is, you know, one of the most popular uses of these tools, it doesn't need to watermark it.
00:16:12.422 --> 00:16:15.292
But Anthropic is saying,"We're watermarking those also."
00:16:15.653 --> 00:16:16.153
Right
00:16:16.542 --> 00:16:18.701
And, like, this is how I use AI.
00:16:18.850 --> 00:16:33.530
I have a whole tool that I've built personally that I write my articles, and then I have a whole very, very deliberately trained on my writing, specific to what I'm looking for, specific to what I'm doing.
00:16:33.750 --> 00:16:35.421
I have it do a, a review.
00:16:35.681 --> 00:16:39.390
Mostly it gives me feedback, and then I go in and make the edits.
00:16:39.561 --> 00:16:49.022
But it also will, you know, especially where I do, like, run-on sentences or, typos, it will call those out, and it will make suggestions for, like, how to, edit them.
00:16:49.350 --> 00:16:54.301
And so now those, in theory, may also include a kind of AI watermark.
00:16:54.311 --> 00:17:13.057
And so you raise this question that comes up in, in all sorts of situations where it's like, if I write the whole thing, have an AI review it, and it catches a typo and says,"You should rewrite this sentence," and I rewrite it the way it suggests, and then somebody runs that through an AI checker, do then I get accused of, it being AI?
00:17:13.462 --> 00:17:22.242
Or you can go even further where there are all sorts of stories of people who have legitimate reasons for using these AI tools to help them write.
00:17:22.242 --> 00:17:29.353
You have people who are not fluent in a language who might use it to im- improve their fluency to make themselves more understood.
00:17:29.653 --> 00:17:36.251
I've heard of, stories of people with disabilities who use the technology to help them in, in other ways.
00:17:36.662 --> 00:17:48.215
And now if we're suddenly, putting this sort of scarlet letter on anything that was processed in any way by that can cause real damage as well.
00:17:48.256 --> 00:17:56.179
And so I think there's a lot of different tricks and, questions raised by like how well this will work and what it's doing.
00:17:56.647 --> 00:18:20.633
I think it's funny because like part of the, EU Commission's guidelines, I was looking this over in preparing for this, is they were saying that, people should know when they are interacting with AI or exposed to AI-generated content and it says this will help them make informed decisions, calibrate their trust and reliance on AI, and avoid misinformation or deception, which is a worthy goal.
00:18:21.143 --> 00:18:25.840
But I'm wondering, does this actually do that or does this make this worse?
00:18:25.840 --> 00:18:40.221
Because now suddenly you're going to have all these people who are freaking out about these things, accusing people of,"Oh, you just used AI to generate this," and actually sowing more mistrust and distrust of content that maybe doesn't deserve that.
00:18:40.352 --> 00:18:52.558
And so there's a lot of different factors that go into this and, and, Anthropic has to do this and, and Google and, OpenAI and others in some sense will have to do this to comply with the EU's new regulations.
00:18:52.909 --> 00:18:58.321
But I, I'm wondering if the long-term impact of this is actually, more trust or less
00:18:58.653 --> 00:18:58.983
Yeah.
00:18:58.983 --> 00:19:01.943
I mean, it gets to the heart of kind of what authorship is, doesn't it?
00:19:01.943 --> 00:19:03.923
And, and who owns a piece of work.
00:19:03.923 --> 00:19:16.884
And I think that's why it kinda gets under so many people's skins because e- everyone, you know, feels, I think quite strongly about, if they're creating original content and using a tool to augment it in some way, it's still theirs.
00:19:17.433 --> 00:19:20.327
my question would be like, Is this gonna matter in a couple of years' time?
00:19:21.191 --> 00:19:32.810
you know, i- i- are we not gonna be in a world where everyone is using these tools so ubiquitously that actually whether it's kind of been edited or authored by Claude it's kind of neither here nor there?
00:19:33.040 --> 00:19:34.500
And I, I feel like we're almost...
00:19:34.811 --> 00:19:47.394
there are elements of that already, I feel, where we don't know if somebody has kind of authored a, a LinkedIn post or a blog post, or if they've come up with a, an idea for a podcast, let's say.
00:19:47.752 --> 00:19:58.699
and yes, this, the way that Anthropic are approaching this isn't going to help clarify who's originated the content and who's augmented it, but everyone's going to be doing it anyway, so does it really matter?
00:19:59.146 --> 00:20:06.053
Well, you know, and actually there is an element of this which people are saying is that, you know, people are going to figure out how to strip out the watermarks.
00:20:06.053 --> 00:20:17.970
I mean, the text watermark w- with video and pictures and audio, I think it's easier to, put in a watermark that is harder to remove, though I, I imagine people will quickly develop tools that will remove those watermarks.
00:20:18.210 --> 00:20:29.955
With the text ones, it's very, very, like, it's going to be trickier, but it, it's trickier to build in the watermark in the first place, and so therefore it will be easier to figure out ways to remove it.
00:20:29.955 --> 00:20:46.146
And Anthropic has even said editing a piece of text that's generated by Claude may or may not remove the watermark, which is, is sort of an indication of how it works, is going to be certain orders of words or, you know, punctuation or something that's, that gives some sort of indication.
00:20:46.146 --> 00:20:52.480
So the editing will depend on whether or not you disturb the elements of the watermark enough, to make it invisible.
00:20:52.480 --> 00:20:59.356
But that also means you could accidentally, in some cases, create text that would set off the watermark when it doesn't.
00:20:59.688 --> 00:21:31.230
But also, I think what will happen is that those in the know, those who you know, have more, resources and ability and knowledge are probably going to figure out ways to avoid the watermark or avoid being accused of using this stuff, whereas people, with less resources are more likely to be called out for using this stuff, even if the more, powerful or wealthy or whoever are using the tools in the same way.
00:21:31.421 --> 00:21:47.382
And so I think you, you actually create a, bigger divide where, yes, lots of people, almost everyone is going to be using these tools in some form or another, but, those who are well-resourced may figure out ways to not get accused of it, while those who are not will get accused of it.
00:21:47.392 --> 00:21:56.462
And so you sort of create this, further version of a digital divide by, you know, having this kind of witch hunt sort of situation going on.
00:21:56.732 --> 00:21:58.853
So I, I worry about that as well.
00:21:58.867 --> 00:21:59.178
Yeah.
00:21:59.498 --> 00:22:05.133
And, and I mean, I, feel like those, that divide is already in place actually.
00:22:05.133 --> 00:22:13.182
You know, I think, I think there's-- this obviously has the potential to exacerbate it, but I think the, there is a kind of group of haves and have-nots in, in the AI race.
00:22:13.232 --> 00:22:25.529
And, you know, this discussion of, who will become an underclass if they can't use AI tools to the same extent as, Sam Altman and his merry band of CEOs is, has been a long kind of standing discussion.
00:22:25.925 --> 00:22:31.443
the question is really whether this is just a kind of compliance move or whether there's actually kind of belief that this will help.
00:22:31.789 --> 00:22:45.602
That's what it kind of comes down to really, doesn't it, is does Anthropic feel like this is going to alleviate some of the, the questions around, around whether people are producing content of their own or, or just cribbing, as the person who I spotted on LinkedIn did?
00:22:46.031 --> 00:22:48.132
Or whether it's just a kind of thing to keep EU happy.
00:22:48.607 --> 00:22:49.397
It feels like the latter.
00:22:49.691 --> 00:22:55.270
yeah, I think this is, this is purely driven by the EU regulations and the compliance aspect of it.
00:22:55.570 --> 00:23:02.509
You know, there are services out there that right now that claim to, they will recognize AI-generated content.
00:23:02.799 --> 00:23:06.467
I've pointed out, mistakes that I think that technology has made.
00:23:06.707 --> 00:23:20.005
It's sort of funny because those technologies also rely on AI, and it's, somewhat amusing at times to see people who absolutely hate AI pointing to those services as proof that, you know, this or that is AI slop.
00:23:20.267 --> 00:23:24.355
and it, it's all just like, it's all AI up and down the stack.
00:23:24.657 --> 00:23:31.480
and, I think just the attitude of, thinking that we have to blame people for using these tools.
00:23:31.641 --> 00:23:38.528
Like, again, I get that, and I have called out, you know, when I've seen AI-generated slop, right?
00:23:38.538 --> 00:23:42.107
Like, there are times where you look at it and you're like,"Oh my God," you know,"have a little self-respect.
00:23:42.107 --> 00:23:52.391
Why are you doing this?" but there are also times where I think, there are good reasons for people to use the technology and, to help them do certain things that they might not otherwise be able to do.
00:23:52.760 --> 00:24:15.391
And so, to me, like so many things, like this is the same argument I've made in the past in regarding other technologies, is that this is a new technology that changes the way people around the globe interact and do things, and we as a society need to learn how to deal with that properly and to understand it and to come to terms with what it means and what we think of it.
00:24:15.771 --> 00:24:29.059
and, you know, I think rushing to the space where it's just like,"Oh, we need to, call out people for using the technology and potentially bash them for it," is not gonna be particularly helpful in, in the long
00:24:29.281 --> 00:24:30.912
Yeah, I think it's-- that's true.
00:24:30.912 --> 00:24:41.094
I mean, the AI Act obviously was fairly early on in terms of trying to kind of put some guardrails and, and some kind of structure around how AI, the AI industry would evolve.
00:24:41.693 --> 00:24:49.363
at the same time, you know, there are already news outlets that have reporters who are solely AI personas, right?
00:24:49.363 --> 00:24:50.042
And who, who,
00:24:50.489 --> 00:24:51.979
Yeah, it's just garbage though.
00:24:52.962 --> 00:24:55.292
I mean, some of it is and some of it isn't.
00:24:55.363 --> 00:25:03.499
I think there, there's a kind of variety of, of, of quality in there, and some people are leaning into it more, and some of it is, is obviously kind of fairly weak.
00:25:03.855 --> 00:25:11.430
But it, again, this idea that the kind of genie's out of the bottle, i- in many ways, and this is like a kind of, feels like a band-aid on, on, on that-
00:25:11.731 --> 00:25:12.182
yeah.
00:25:12.221 --> 00:25:26.902
And r- right, you know, to, to be fair, right, there are certain examples where you could see where the AI can be useful, and there are stories where it's like you have, like, local newspapers that don't have enough resources to cover things like local high school sports or something like that.
00:25:27.122 --> 00:25:36.042
And if, if you can get stories out of it that wouldn't get covered otherwise, like, maybe that is a better solution than just not covering the, the space at all.
00:25:36.289 --> 00:25:51.682
but yeah, it's just a question of like, what is society's attitude towards all these things as well, and does something like this, increase thoughtful approaches to using the technology, or does it just create this kind of like, we're gonna burn these people at the stake because they use this
00:25:52.347 --> 00:25:52.718
Yeah.
00:25:52.907 --> 00:25:53.228
Yeah.
00:25:53.228 --> 00:25:56.768
kind of a further way that we can be polarized as a society.
00:25:57.131 --> 00:25:58.111
Yeah, yeah
00:25:58.548 --> 00:26:01.988
talking of which, a gentle segue into our eye-catching section.
00:26:02.397 --> 00:26:04.579
Um, I wondered, you know, what,
00:26:04.739 --> 00:26:06.479
Speaking of people we can burn at the stake.
00:26:08.279 --> 00:26:20.473
I wasn't quite as explicit as that, but, um, I, I, I wonder what would come up if you put Mark Zuckerberg's 6,000-word just kind of novella on AI through a watermarking tool.
00:26:20.723 --> 00:26:28.786
I wonder just how much of it might be created by whatever, model he uses, probably Muse Spark, but we don't know.
00:26:28.894 --> 00:26:31.565
because everyone has been talking about this.
00:26:31.836 --> 00:26:33.355
Some people are saying it's an essay.
00:26:33.695 --> 00:26:36.207
I think it's a kind of badly edited monologue.
00:26:36.605 --> 00:26:45.377
his essay entitled"The Future Is for Everyone," and if watermarking kind of divided opinion, particularly among the LinkedIn crowd, then this has as well, hasn't it?
00:26:45.807 --> 00:26:46.307
Yeah.
00:26:46.613 --> 00:26:55.289
you know, I, I refer to it as a manifesto, um, because, yeah, it has that sort of feel where, you know, the future is for everyone.
00:26:55.637 --> 00:27:02.509
and, you know, I think a lot of people have been, completely bashing it, mostly for good reason.
00:27:02.814 --> 00:27:10.026
you know, my take on it, he, he talks about a whole bunch of different things, but basically he's just like a very strong, he supports AI.
00:27:10.036 --> 00:27:14.346
He thinks AI should be accessible and available for everyone.
00:27:14.696 --> 00:27:24.807
you know, he always refers to super intelligence, which is this horrible, stupid, annoying, misleading term that, that he has really embraced when it comes to AI technologies.
00:27:25.172 --> 00:27:36.492
but he thinks that more people should have access to it and that they should have access to local, open weight, distributed models, which I agree with.
00:27:36.542 --> 00:27:40.969
And so there's like a lot in there that I'm reading and I was like, yeah, he's right.
00:27:40.979 --> 00:27:58.352
He's making these arguments that I think are actually correct, surrounded by an absolute ton of bullshit and nonsense about how amazing data centers are and how they're you know, revitalizing communities and all this nonsense.
00:27:58.696 --> 00:28:11.448
and then, the larger point though is, and I wrote something about this this week as well, is that he talks about the importance of decentralization and how the power shouldn't be in the control of just three giant centralized companies.
00:28:11.938 --> 00:28:26.438
And I'm just like,"Who-- Y-you're Mark Zuckerberg." Like, the whole reason we're concerned about the giant power of three or four centralized companies is partly because of your doing and things that you have done for the past 20 years.
00:28:26.478 --> 00:28:46.433
Who the fuck are you to, like, come in there and say, like,"Oh," like,"the world, it doesn't end well when new technologies are just in the hands of a few companies." It's like for 20 years, you've taken steps to make sure that Meta is the centralized giant, and you've, actively prevented decentralization within the the social space.
00:28:46.784 --> 00:28:54.714
So I wrote something basically saying, like, there are things in here that he's right about, but you should not trust him to be the messenger on it.
00:28:54.865 --> 00:28:58.694
And the only reason he's saying these things is because they're losing the race.
00:28:58.694 --> 00:29:09.752
I mean, we've talked about this before, that you know, Meta is way behind the other models, and therefore he has to put forth this idea that like, yes, the open decentralized models matter.
00:29:10.012 --> 00:29:18.838
But the second that Meta were ever close to the top of the race again, he would shut things down so fast because that's what he's done.
00:29:18.878 --> 00:29:29.234
And, and I pointed to the example of Threads and ActivityPub, where when they launched Threads, they were like, we believe in openness, we believe in the open social network where it should be on protocols.
00:29:29.234 --> 00:29:47.463
We shouldn't have control." And then as soon as Threads suddenly took off and became much bigger because they shoved everybody in Instagram into it, then suddenly they're like,"Oh yeah, ActivityPub, yeah, we'll, we'll maybe support it." And like everybody who was working on that basically left Threads and they don't really care about it anymore.
00:29:47.813 --> 00:29:49.402
this is all just marketing.
00:29:49.651 --> 00:29:56.682
And so I think some people are dismissing everything that's in there I do think there are some interesting things that are in there that are worth paying attention to.
00:29:56.853 --> 00:30:04.508
I just, you know, Mark is the wrong messenger for it because you shouldn't trust him'cause he's, he's not, he's not believable on this.
00:30:04.508 --> 00:30:05.938
He's not, he has no credibility.
00:30:06.262 --> 00:30:09.492
And he thinks he'll have amnesia from the last two decades.
00:30:09.682 --> 00:30:25.163
I think, uh, we've got a couple other stories we wanna touch on before we- Mm we do this, but, uh, the perfect visual metaphor for this essay, Mike, is, a video that Zuckerberg put out on his socials this week of him sparring with a UFC fighter on a barge.
00:30:25.213 --> 00:30:25.933
Have you seen this?
00:30:26.317 --> 00:30:27.166
No, I have not.
00:30:27.292 --> 00:30:34.847
It's, you know, you, you talked about there kind of being like a couple of ideas in, amongst a sea of terrible, terrible takes in this essay.
00:30:35.137 --> 00:30:36.185
Well, there's a, a video.
00:30:36.185 --> 00:30:38.165
You know he's a kinda big MMA guy now.
00:30:38.865 --> 00:30:49.130
There's a video of him taking on the number one UFC fighter on a barge in the middle of the ocean, and It's a kind of perfect visual representation of this.
00:30:49.140 --> 00:31:02.236
Like, he is, awash with, bad ideas and bad takes, and he is standing in the middle of the ocean swiping away, thrashing away, hoping for something to kind of, come good, and that for me sums up where Meta is in the AI race right now.
00:31:02.276 --> 00:31:04.155
I'll, I'll, I'll post it in the show notes.
00:31:04.705 --> 00:31:18.521
talking of kind of, AI companies trying to show themselves in a good light, an eye-catching story that I saw this week, Mike, and I wanted to talk through, was OpenAI's head of ethics leaving after one year in post.
00:31:18.882 --> 00:31:18.892
Mm.
00:31:18.942 --> 00:31:40.369
Now, this is a woman called Chloe Bakalar, who people who are very, very in the weeds on the industry will know was the head ethicist or chief ethicist of Meta for I think five or six years, and joined OpenAI last year as part of an effort to, her job was to kind of create, processes for, for model development, in ethical ways.
00:31:40.380 --> 00:31:43.789
She was thinking a lot about how humans interacted with AIs.
00:31:44.109 --> 00:31:52.541
You know, all the kind of, like, slightly philosophical questions that I guess don't get answered by teams who are working on research or, product or the technology itself.
00:31:53.031 --> 00:31:59.549
She, is leaving, and the company have announced that she's not going to be replaced, which I think is interesting.
00:31:59.978 --> 00:32:01.147
you know, I have a couple of thoughts on this.
00:32:01.147 --> 00:32:21.818
I mean, not necessarily very surprising in the run-up to an IPO that you might shed one of the people who wants to do very thoughtful deliberation o- on, on models and maybe, like, have a couple more meetings and ask some kind of pointed questions of, developers about how they're, structuring the data and developing these models.
00:32:21.997 --> 00:32:30.949
kinda makes sense that you, you might shepherd that person out the door, and, the justification that they've given is that ethics is now kind of baked into a bunch of teams within OpenAI.
00:32:30.949 --> 00:32:35.479
It's, it's core, yeah, core to a lot of what, the research they do.
00:32:35.669 --> 00:32:36.009
And, you
00:32:36.113 --> 00:32:40.563
Who needs ethics when you have, uh, you know, ChatGPT handling ethics for
00:32:40.800 --> 00:32:42.181
Yeah, for sure.
00:32:42.455 --> 00:32:50.958
and you might be g- be able to get on board the idea that there's a product team, who is, taking all the work that Chloe's done over the last year and kind of implementing it in new ways.
00:32:51.217 --> 00:33:05.619
the other kind of question is, like, whether this is a further distinction between OpenAI and its big rivals, Anthropic, which obviously has done lots of work around, the kind of philosophy, of models and how models behave and the morals that a model has.
00:33:06.006 --> 00:33:14.288
they have a philosopher who's been responsible for creating Claude's soul, if a model can have a soul, and we've talked about that in the past.
00:33:14.288 --> 00:33:28.823
And they've done lots of work, with the, kind of Pope and the Vatican City, and kind of really just, I guess, trying to kind of instill the idea that they are a company that has, models that kind of think differently or act differently in certain scenarios.
00:33:29.182 --> 00:33:36.587
And so, you th-this feels like another fork in the road between what OpenAI are doing and what Anthropic have tried to do or to kind of show.
00:33:37.078 --> 00:33:42.290
Now, I don't for one think that Anthropic is entirely whiter than white or the good guys here.
00:33:42.661 --> 00:33:54.097
Um, but I think it's important to note, and, uh, again, we have in technology companies for the past 15 or 20 years had ethicists, people who have been thinking about kind of big ethical questions.
00:33:54.404 --> 00:33:58.726
in the early days, I don't think they had philosophy degrees, or had, were actually trained in ethics.
00:33:58.736 --> 00:34:03.103
I think the early, early ones were, like, comms officers and people who were kind of shoved into that role.
00:34:03.123 --> 00:34:08.744
So maybe we're further on than, we were, but I think this is a kind of noted development as, as, those IPOs come along.
00:34:08.925 --> 00:34:09.186
Yeah.
00:34:09.186 --> 00:34:34.583
I mean, the thing that I would point out is like for most tech companies, and I don't know the details here, I don't know that any of this is public, but you know, for most tech companies, the way it works when you join is you get, an equity grant that vests over time, and most of it is you don't get anything until you've been there for one full year, and then you get, there's a cliff, and you get your first, you know, quarter of the equity grant at one year.
00:34:34.733 --> 00:34:46.092
So somebody leaving before an entire year is up is sort of eyebrow-raising, especially right before an IPO that in theory would make one who holds OpenAI equity very, very wealthy.
00:34:46.394 --> 00:34:48.760
so there's, definitely an eyebrow raise there.
00:34:48.789 --> 00:34:56.503
Leaving before one year is noteworthy in that it suggests something, something went wrong.
00:34:56.954 --> 00:35:04.155
Uh, whether it's with OpenAI, whether it was with this individual, who knows or some combination of both perhaps.
00:35:04.155 --> 00:35:07.286
But like, to me that, is eyebrow-raising.
00:35:07.327 --> 00:35:17.809
I think, the whole thing of like companies hiring ethicists is interesting, but hiring an, you know, an ethics specialist doesn't make you an ethical company.
00:35:17.960 --> 00:35:21.630
It's often, it feels like a marketing exercise to some extent.
00:35:21.679 --> 00:35:26.570
Like actually being ethical is different than just, hiring an e-ethicist.
00:35:26.630 --> 00:35:29.166
and like, different people can have different takes.
00:35:29.166 --> 00:35:38.295
Like the whole nature of ethics is that, there is no exact right answer, and different people will disagree on what is the ethical thing to do in different scenarios.
00:35:38.710 --> 00:35:42.625
but yeah, you know, it's definitely raises some, some eyebrows about what's happening at OpenAI.
00:35:43.083 --> 00:35:43.992
Yeah, indeed.
00:35:44.407 --> 00:35:51.677
and then just before we, we wrap up, you want to mention a, an eye-catching story that you think has a lot of implications going forward.
00:35:51.686 --> 00:35:53.277
Do you wanna just give us a quick overview of that?
00:35:53.293 --> 00:35:53.862
Yeah.
00:35:54.103 --> 00:36:08.971
it is a quick one because, uh, the Ninth Circuit made a ruling, uh, Meta and, TikTok had, challenged, it's really a series of, there's like thousands of cases that the, the sort of child safety cases and everything else is, that's going on.
00:36:09.304 --> 00:36:15.125
and they had been denied a Section 230, dismissal early on, and they had appealed it.
00:36:15.454 --> 00:36:38.309
And the Ninth Circuit, which is the appeals court that covers most of the western part of the United States, is very important, ruled that Section 230 is not actually an immunity from suit, but rather merely a defense to liability, and said effectively that companies cannot appeal the loss of a 230.
00:36:38.338 --> 00:36:45.365
You know, they-- you try and get a case dismissed on 230 grounds, and historically, if you've lost that, you can immediately appeal it.
00:36:45.666 --> 00:36:50.715
And here, the Ninth Circuit is saying you can't immediately appeal it because it is not an immunity from suit.
00:36:51.081 --> 00:37:10.260
that is, it feels like a very narrow ruling, but it is an incredibly important one because the fact that Section 230 has been considered an immunity from lawsuit, which I think is the most obvious read of what the law actually says, has been the reason why Section 230 is so important.
00:37:10.690 --> 00:37:21.699
And now that has been changed, you know, the court kinda claims that this was never really, analyzed, and other courts have said differently in the past, but they sort of dismissed that as sloppy language.
00:37:21.949 --> 00:37:35.246
I saw somebody went through and did analysis saying there were 442 court rulings that have s- called Section 230 an immunity, and for the Ninth Circuit to just kind of dismiss it as sloppy.
00:37:35.476 --> 00:37:57.556
It also sort of rewrites the history of Section 230 in that, you know, they claim that, the original version of the language didn't include this language suggesting it was an immunity, and therefore, when they added that language, they said it was a minimal change, and if this really was about adding immunity, they wouldn't have called it a minimal change.
00:37:57.856 --> 00:38:00.416
That is also a complete rewriting of the history.
00:38:00.425 --> 00:38:02.186
Jeff Kosseff spoke out publicly.
00:38:02.186 --> 00:38:06.686
Jeff Kosseff, who wrote the book on the history of Section 230, said that's wrong.
00:38:06.695 --> 00:38:15.753
That's like, you know, the people who were there in the room deliberately said that they made this change on purpose to create to make Section 230 an immunity from lawsuits.
00:38:16.155 --> 00:38:18.072
so this could be a monumental change.
00:38:18.072 --> 00:38:22.063
I think it, takes away much of the value of Section 230.
00:38:22.284 --> 00:38:26.112
the question is will Meta and/or TikTok challenge this?
00:38:26.121 --> 00:38:30.237
They can ask the Ninth Circuit to rehear it, banc, which is with the larger panel.
00:38:30.467 --> 00:38:31.807
They could go to the Supreme Court.
00:38:32.025 --> 00:38:44.626
both of those are fraught with risks and challenges because you know, the Ninth Circuit at a larger scale could also agree with this, or the Supreme Court could as well, and that would be very bad for the future of Section 230.
00:38:44.757 --> 00:38:45.117
Mm.
00:38:45.358 --> 00:38:54.117
And do you, so do you think, Mike, this is the kind of recent lawsuits against these platforms being rewritten into the kind of legal system?
00:38:54.117 --> 00:39:00.666
It's, you know, it's, it's New Mexico, it's California, it's the product design, issue being reapplied, isn't it?
00:39:00.958 --> 00:39:01.217
Yeah.
00:39:01.217 --> 00:39:15.853
I mean, if you look at the last five years of court decisions related to Section 230, one by one, courts have come up with new interpretations of Section 230 that effectively chip away at each of the benefits of Section 230.
00:39:16.443 --> 00:39:32.822
So bit by bit, even as Congress is talking about repealing or reforming, the courts have done that work bit by bit, changing the way they interpret the law so that each of the, underlying benefits of 230, which is that it gets rid of bad mistargeted lawsuits quickly, are going away.
00:39:32.871 --> 00:39:44.092
And I think we're going to have very long and very damaging consequences from that, but it's the courts, you know, effectively rewriting the law on their own because they're not happy about these companies.
00:39:44.481 --> 00:39:44.882
Yeah.
00:39:45.271 --> 00:39:53.181
And, and I think this is probably a good time to plug your, Section 230 podcast, Mike, that you did last year.
00:39:53.996 --> 00:39:56.523
doing my good Otherwise objectionable.
00:39:56.815 --> 00:40:06.429
not a description of, you or I, but uh, a, a kind of narrative-driven podcast that Mike did, which is, is very good on, kind of the history of Section 230.
00:40:06.438 --> 00:40:08.268
We'll include that in the show notes.
00:40:08.402 --> 00:40:10.137
More listening for you listeners.
00:40:10.476 --> 00:40:11.635
but this podcast is not done.
00:40:11.635 --> 00:40:12.516
We're not finished here.
00:40:12.612 --> 00:40:14.135
we're wrapping up the free episode now.
00:40:14.503 --> 00:40:16.838
we've talked a bit about, some really important stories.
00:40:16.838 --> 00:40:27.025
We're also gonna talk in the Patreon episode about, a major kind of social and immigration issue that's been bubbling under the surface in, in Europe over the last month.
00:40:27.306 --> 00:40:34.831
We're gonna do a similar treatment to what we did last week with the, Indian meta story and kind of unpack what happened and what it means.
00:40:34.831 --> 00:40:38.628
There's been some interesting developments this week, and, uh, looking forward to jumping into it now
00:40:40.485 --> 00:40:43.764
That's the end of this week's free edition of Ctrl Alt Speech.
00:40:44.146 --> 00:40:51.025
To hear the extended episode with additional coverage, become a Patreon supporter at patreon.com/ctrlaltspeech.
00:40:51.286 --> 00:40:53.824
That's C-T-R-L Alt Speech.
00:40:54.405 --> 00:41:00.826
If your organization is interested in sponsoring the podcast, contact us via our website at ctrlaltspeech.com.
00:41:01.264 --> 00:41:01.985
Thanks for listening
00:00:00.047 --> 00:00:03.657
So Mike, we're gonna talk today about how you can't always trust what you can see online.
00:00:03.777 --> 00:00:07.466
That feels like a kind of running theme through Ctrl Alt Speech.
00:00:07.746 --> 00:00:09.787
Um, but
00:00:09.884 --> 00:00:13.064
But you can trust what you can trust when you listen to us.
00:00:13.105 --> 00:00:14.064
You can trust us
00:00:14.467 --> 00:00:15.836
you can, you can.
00:00:16.137 --> 00:00:17.339
That's a, high bar.
00:00:17.339 --> 00:00:19.000
It means we cannot make any mistakes today.
00:00:20.109 --> 00:00:20.579
Yes
00:00:20.579 --> 00:00:23.710
Um, not that I think anybody is fact-checking us.
00:00:23.943 --> 00:00:36.296
but you might not be able to trust everything, but there are some good things about the internet, and I wanted to ask you, what have you been watching or enjoying that makes you feel like the internet is worthwhile sticking with?
00:00:36.723 --> 00:00:37.253
Yes.
00:00:37.264 --> 00:00:49.371
Well, you know, I think one of the things that I think lots of us think a fun outgrowth of the internet is the rise of memes and meme culture and being able to share memes and things like that.
00:00:49.790 --> 00:00:56.546
So I found this thing which is brand new, it's only been out for a few days, called Meme Museum.
00:00:56.866 --> 00:01:08.754
So it's meme-museum.at, and it is built on the AT Protocol, which, you know, I am a supporter of the AT Protocol, ding, due
00:01:08.840 --> 00:01:09.620
please go bold.
00:01:09.843 --> 00:01:10.944
to, to Blue Sky.
00:01:11.243 --> 00:01:12.474
This is not Blue Sky.
00:01:12.483 --> 00:01:13.974
This is the AT Protocol.
00:01:14.343 --> 00:01:21.293
and there's actually a really interesting article from the guy who built it, this guy Joe Grmusica, I think is, is the name.
00:01:21.661 --> 00:01:27.590
and was just saying he had this idea of something he thought would be cool, which was a tool to help you manage your memes.
00:01:27.921 --> 00:01:35.861
So if you have, memes and you wanna be able to share them on social media, there are people who have, like, folders full of memes and things like that.
00:01:36.203 --> 00:01:38.784
But this is built into the AT Protocol.
00:01:39.180 --> 00:01:56.963
but he wrote a really interesting article about how it was just something that he had this thought and he could build it, and then it goes on about how nice it is to have that protocol that you can just build something and be able to, you know, build something into, a social media system, and it's a cool tool for managing all your memes.
00:01:57.043 --> 00:01:57.403
And I was
00:01:57.439 --> 00:01:58.019
I love that
00:01:58.161 --> 00:01:59.010
I like memes.
00:01:59.370 --> 00:02:05.667
I like being able to build on open protocols, and so this made me happy about the internet this week.
00:02:06.417 --> 00:02:07.037
How about you?
00:02:07.048 --> 00:02:10.354
mean, I, I've actually worked with Joe quite closely on a couple of projects.
00:02:10.784 --> 00:02:13.594
Yeah, he's done some great work, and, he's a really fun guy.
00:02:13.625 --> 00:02:17.170
So it's no surprise that he's done something cool, on the out protocol.
00:02:17.411 --> 00:02:20.131
And, and just before we move on, what, what is in your meme museum?
00:02:20.290 --> 00:02:22.871
I think people are gonna be wondering, what are you collecting?
00:02:23.067 --> 00:02:23.407
I am...
00:02:23.599 --> 00:02:27.366
I, well, I mean, I've just been playing around, so there's not that much in there yet.
00:02:27.387 --> 00:02:29.537
I need to, I need to start adding stuff to it.
00:02:29.787 --> 00:02:34.486
I have not been as good about some other people with, with my meme game.
00:02:34.747 --> 00:02:47.902
So maybe this will be my inspiration to up my meme game,'cause I'm-- I feel like memes are a little bit more of a millennial thing, and I am solidly in the Gen X group.
00:02:47.913 --> 00:02:54.836
But I feel sort of millennial at heart at times, and so this is my inspiration to maybe up my meme game
00:02:54.933 --> 00:02:55.283
Yeah.
00:02:55.343 --> 00:02:55.633
Nice.
00:02:55.633 --> 00:03:02.456
If you've got any, ideas, listeners, of memes that Mike should be curating in his museum, get in touch, podcast.ctrl-alt-speech.com.
00:03:02.877 --> 00:03:07.776
You have full unfettered rights to send him memes of all kinds.
00:03:09.396 --> 00:03:10.187
That's what I heard.
00:03:10.409 --> 00:03:10.800
Ben?
00:03:10.800 --> 00:03:11.599
What have you done?
00:03:13.056 --> 00:03:16.973
Um, well, I, My kind of fun thing this week, Mike, is, something for my son.
00:03:16.973 --> 00:03:28.790
Obviously, he made an appearance on the podcast a few weeks ago while he was having his bath and going to bed, and so You know, it's not the only interruption that I've had from him while sat at this very desk.
00:03:28.881 --> 00:03:41.591
He likes to, normally after a nap, come into my office, look at, you know, what I'm up to, and start bashing the keyboard, and that's, that's fairly problematic when you've got kind of documents and, emails up and things like that.
00:03:41.841 --> 00:03:44.131
You know, he manages to close a lot of windows.
00:03:44.270 --> 00:03:49.234
He manages to, you know, he managed to send an email at one point, which I was fairly horrified at.
00:03:49.614 --> 00:03:58.824
Um, so I've, I, I, I've found a tool, a website called tinyfingers.net, which is specifically created for this purpose.
00:03:59.093 --> 00:04:14.491
It is a full-screen website where babies can come along, and they can smash the keyboard to their heart's content, and it brings up little, you know, kind of emojis and emoticons and, has a nice kind of snowscape pattern as well, and I've used it a couple of times.
00:04:14.491 --> 00:04:21.721
It works incredibly well for keeping him happy and, uh, s- avoids any kind of professional, issues on my side.
00:04:21.721 --> 00:04:23.141
So great all round
00:04:23.428 --> 00:04:23.889
Yeah.
00:04:23.992 --> 00:04:30.543
I mean, I'm looking at it right now and it says 1,072 kids are smashing keyboards right now.
00:04:30.903 --> 00:04:39.528
Now I would be remiss if I didn't question the trust and safety implications of a website like this.
00:04:39.528 --> 00:04:39.538
But
00:04:39.788 --> 00:04:43.428
the kind of screen time social media ban, impetus
00:04:43.536 --> 00:04:44.257
Exactly.
00:04:44.257 --> 00:04:52.112
Like, does this comply with, uh, you know, is this site covered by the, bans on social media for kids?
00:04:52.362 --> 00:04:59.293
I mean, this is definitely for kids under the age of 16, and so I, I have to wonder about some of that.
00:05:00.536 --> 00:05:01.877
Look, look, don't get me started.
00:05:12.242 --> 00:05:20.826
Hello, and welcome to Ctrl-Alt-Speech, the podcast where we make sense of the major debates shaping online speech, platform power, content moderation, and the future of the internet.
00:05:21.185 --> 00:05:37.153
It's August the 13th, 2026, and this week we're discussing why everyone is freaking out about Anthropic's watermarking announcement, what's next for AI ethics, and for our Patreon supporters, how misinformation took hold in a tiny Spanish enclave a few weeks ago.
00:05:37.622 --> 00:05:44.252
Ctrl-Alt-Speech is the weekly moral audit of the internet, co-hosted f- fairly grand.
00:05:45.403 --> 00:05:46.043
I like it.
00:05:46.053 --> 00:05:46.634
I like it
00:05:47.033 --> 00:05:52.983
co-hosted by Mike Masnick, founder and editor of Techdirt, and me, Ben Whitelaw, the founder and editor of Everything in Moderation.
00:05:53.242 --> 00:05:55.423
Mike, you already know the big news of this week.
00:05:55.670 --> 00:05:56.925
the Bluesky bell is back.
00:05:57.752 --> 00:05:58.403
I heard it.
00:05:58.603 --> 00:05:59.083
Ding, ding
00:05:59.906 --> 00:06:03.702
Yeah, perhaps used slightly incorrectly, a bit rusty.
00:06:04.151 --> 00:06:09.521
We, we haven't used it in a while, so, so w- if we have new listeners, you should explain the Blue Sky Bell
00:06:09.666 --> 00:06:20.769
The Bluesky bell is, well, where it's been is, is in a box in a, a room of my house while we have been renovating some, parts of our London, house.
00:06:20.790 --> 00:06:23.290
And so it has been fairly absent, I would say.
00:06:23.290 --> 00:06:35.226
But it's designed, when present, to highlight the fact that Mike is a board member of Bluesky and the, you know, potential, let's say kind of implications of you talking about a story related to Bluesky.
00:06:35.297 --> 00:06:37.720
Not that you're biased, but that you have priors.
00:06:38.499 --> 00:06:43.228
Do not trust anything, I say, because I am, I am fully biased.
00:06:43.771 --> 00:06:49.942
And, and the bell is, is the first of, I think, a, a series of different, um yeah, I just like the sound.
00:06:50.201 --> 00:06:54.742
Um, a f- a number of different instruments, noises that we're gonna kind of hopefully introduce.
00:06:54.781 --> 00:06:56.922
We've been talking about it for a while, but I'm quite committed to it
00:06:57.007 --> 00:07:00.187
I am still waiting for the trade-offs tuba, so
00:07:01.091 --> 00:07:04.461
I've got a, I've got another proposal actually for, for later on in the podcast.
00:07:04.461 --> 00:07:06.942
So yeah, we're starting, to bring that back.
00:07:07.541 --> 00:07:35.869
before we jump into, new production noises for the podcast and stories that we think you should all be listening to, people might be tired of me saying, but we are very, very grateful to our very dedicated group of Patreon supporters who pay either£10 a month to get, excellent extended episodes and early access to them as well, as well as our 20 a month founder and insider supporters who get both of those things and the opportunity to suggest stories for us to cover.
00:07:36.266 --> 00:07:41.555
a number of those supporters actually got in touch with us in recent weeks to do that, and we covered those stories.
00:07:41.886 --> 00:07:45.185
So it's a very, very, valuable thing, and we couldn't do the podcast without it.
00:07:45.523 --> 00:07:56.696
we monetize both through the Patreon supporters and also sponsorships, and it really helps us cover, the many, many articles we read and the production of it and our brilliant editor, Lee.
00:07:56.696 --> 00:08:00.615
So it's all going to a, a very, very good cause.
00:08:00.942 --> 00:08:07.048
if you are interested in sponsoring the podcast, you can find out more, If you go to our website, ctrlaltspeech.com, there is more information there.
00:08:07.889 --> 00:08:16.012
And if you are tempted by the Patreon, we have a very, very good offer right now, an almost free first month, with the discount code TRUSTCON.
00:08:16.271 --> 00:08:29.915
Use all caps, T-R-U-S-T C-O-N, and that gets you, this week a very, very interesting kind of picking over this big mass migration event in the Spanish city of Ceuta and how platforms are responding to it.
00:08:30.305 --> 00:08:31.925
So have a look out for that.
00:08:32.331 --> 00:08:35.072
Mike, where do we, where do we begin this week?
00:08:35.375 --> 00:08:37.595
I think we need to watermark this conversation
00:08:38.440 --> 00:08:41.610
In, a few years' time it probably will be, and we, we, we...
00:08:41.759 --> 00:08:50.043
People will know that there was no AI involved in our, inadequate words and attempted analysis.
00:08:50.413 --> 00:08:55.842
but, you know, there is this kind of growing question of, of what can you trust online, what is real?
00:08:56.273 --> 00:09:02.850
And we're gonna talk a bit about fact-checking later, which I think was kind of the original attempt to help validate what people saw on the web.
00:09:03.102 --> 00:09:10.355
but we have been waiting for the AI labs to really come up with their own play on this, and Anthropic this week announced something that I think fits neatly into that bucket.
00:09:10.807 --> 00:09:11.107
Yeah.
00:09:11.107 --> 00:09:22.591
So, you know, this sort of all traces back to the EU's AI Act, which, you know, I think has been seen as sort of, a really early jump into the space of regulating AI in, in all sorts of ways.
00:09:22.902 --> 00:09:29.312
And one element of that is a requirement for AI produced content to be watermarked in some way.
00:09:29.692 --> 00:09:43.761
you can understand the sort of thinking behind that in that there was a lot of concern, you know, a lot of the initial concerns about generative AI tools was that it was generating content that would fool people into believing that it was real.
00:09:43.782 --> 00:09:45.331
I mean, the, the deepfakes, right?
00:09:45.331 --> 00:09:56.361
We spent years talking about the threat and risk of deepfakes and how they were fooling people, except that there was maybe not that much evidence of people actually being fooled by deepfakes.
00:09:56.466 --> 00:10:05.481
in fact, it seemed like the bigger problem was people doing what's been known as the liar's dividend, where they would claim when they were really caught doing something that it was a deepfake.
00:10:05.782 --> 00:10:17.451
but so built into the EU Act is this, requirement that AI companies somehow watermark or indicate content that is created by AI.
00:10:17.861 --> 00:10:19.971
Now, there have been efforts underway.
00:10:20.254 --> 00:10:40.357
there's a, a standard which is now pretty widely being adopted for image and audio and video AI generated content, that, from what I've heard is, getting pretty good and is fairly hard to, break, though I imagine people will come up with all sorts of ways to break it, over time.
00:10:40.607 --> 00:10:52.888
But what was trickier was the idea of, what AI is most used for, which is generating text, and how do you indicate and mark that text is, originally generated by AI.
00:10:53.524 --> 00:11:09.195
So Anthropic announced this week that they are now, very soon starting to watermark all AI text that they generate and/or process, which is a, an interesting distinction that is worth digging a little deeper into.
00:11:09.565 --> 00:11:29.298
But basically what they're doing is they have put some sort of, at this point still hidden kind of bias into their text generation that will be recognizable to a machine-readable system to indicate that the text has been processed by Anthropic's tools.
00:11:29.597 --> 00:11:45.662
And so at some point they will make an API available that will allow you to take any text and then ship it back to Anthropic, and Anthropic will give you some sort of score as to whether or not that text was processed by Claude in, in some form or another.
00:11:46.022 --> 00:11:50.885
Now, this, there's the, the regulatory aspect of it, which is that they're sort of required to do this.
00:11:50.895 --> 00:11:58.798
It-- as people were talking about this, it came out that apparently Gemini, Google's, AI model, has been doing this for some time already.
00:11:58.928 --> 00:12:03.038
at some point, I assume that OpenAI will release something similar.
00:12:03.437 --> 00:12:05.272
the reactions to this have been all over the map.
00:12:05.481 --> 00:12:08.581
a lot of people are really, really mad about it.
00:12:08.855 --> 00:12:15.282
some for perhaps reasonable reasons, some I think for less than reasonable reasons.
00:12:15.591 --> 00:12:17.361
They are very much...
00:12:17.751 --> 00:12:25.339
You know, there are people who have-- This may surprise you, Ben,'cause I know you, you think well of people in the world and think people are generally good.
00:12:25.695 --> 00:12:36.511
there are some people who use AI to cheat in some form or another, and it is possible that a tool like this might make it harder for them to get away with the, the cheating side of what they do.
00:12:36.912 --> 00:12:40.211
and so some people I think are just mad about that aspect of it.
00:12:40.437 --> 00:12:45.589
just on that, I have a very interesting s- kind of very brief tangent about exactly that point.
00:12:46.004 --> 00:12:52.721
Cause I think a lot of the reaction has been from, like, people who are using AI models for the LinkedIn thought leadership, right?
00:12:52.971 --> 00:12:59.292
You know, there's, there's-- We talked about the AI slop button that's been created on LinkedIn.
00:12:59.591 --> 00:13:00.731
LinkedIn is the worst for it.
00:13:01.121 --> 00:13:15.779
This week, I f- saw two posts that were very, very similar in nature, bar a few kind of structural differences, and called out the fact that these posts written six or seven hours apart were very, very similar.
00:13:16.179 --> 00:13:34.234
And it turned out, long story short, that one of the two people had fed all of the, posts about actually this topic that we're talking about now, anthropic watermarking, had fed all of those into Claude and asked for a, had asked for a take that he could post onto LinkedIn.
00:13:34.716 --> 00:13:35.476
Wow
00:13:35.798 --> 00:13:49.082
I, to be honest, I don't believe that his page isn't running, isn't being run by Claude or OpenClaude or something, because his response to my questions was very, very, sounded very much like a, an AI, chatbot.
00:13:49.679 --> 00:14:00.731
The DM that I got from him sounded very much like an AI chatbot, and frankly, I would love some sort of watermarking in that moment because these two posts were incredibly similar.
00:14:01.168 --> 00:14:07.105
he made up some sort of, reasoning for why, the post is incredibly similar, but he had essentially kind of plagiarized it using AI.
00:14:07.698 --> 00:14:13.568
did you click the, the new LinkedIn button that allows you to, to claim that certain content is AI slop?
00:14:13.775 --> 00:14:15.846
didn't even think to do that, which is
00:14:16.102 --> 00:14:17.822
You well, you should, right?
00:14:17.832 --> 00:14:23.625
That is the new feature on LinkedIn is that you can, you can say,"I think this is AI slop," and, LinkedIn may do
00:14:23.671 --> 00:14:23.980
Yeah.
00:14:24.181 --> 00:14:26.980
But the point is, you know, as you say, it is very, very prevalent
00:14:27.419 --> 00:14:27.730
Yeah.
00:14:28.149 --> 00:14:32.912
there are certain cases and, and, you know, we're seeing more and more examples of it.
00:14:33.192 --> 00:14:55.859
there was a discussion in a, private group chat, so I'm gonna kinda be a little careful about what I say here, about a certain paper that recently came out that is relevant to a lot of our interests, where there was very clear indications that, at least the summary of the paper was written entirely by AI.
00:14:56.298 --> 00:15:00.859
It had a lot of hallmarks of AI drafted nature.
00:15:00.869 --> 00:15:03.158
And so, like, now people are sort of looking at this thing.
00:15:03.158 --> 00:15:18.664
Now, the flip side of that is that, that leads people to, like, completely freaking out about these things and maybe overreacting and assuming that things are AI generated or AI slop, as people would like to say, when it's not AI slop.
00:15:18.735 --> 00:15:24.996
And so you run the r- the real risk of, I mean, you have sort of, type one and type two errors, right?
00:15:25.017 --> 00:15:31.480
The false negatives and false positives where, things that aren't AI generated are going to be accused of being AI generated.
00:15:31.807 --> 00:15:35.615
things that are AI generated are not going to be, caught up in this.
00:15:35.846 --> 00:15:36.666
And it just become...
00:15:36.676 --> 00:15:39.186
You have this sort of witch hunt aspect of it.
00:15:39.628 --> 00:16:03.961
And what makes this worse, and this is part of the way that Anthropic has rolled this out, which I understand technically they probably feel, not doing it this way would be even harder, is that even though the, AI Act is fairly clear that the watermarking requirements or the transparency requirements don't need to apply to merely processing content.
00:16:03.981 --> 00:16:12.371
So if you're doing, like, a spell check or a grammar check, which is, you know, one of the most popular uses of these tools, it doesn't need to watermark it.
00:16:12.422 --> 00:16:15.292
But Anthropic is saying,"We're watermarking those also."
00:16:15.653 --> 00:16:16.153
Right
00:16:16.542 --> 00:16:18.701
And, like, this is how I use AI.
00:16:18.850 --> 00:16:33.530
I have a whole tool that I've built personally that I write my articles, and then I have a whole very, very deliberately trained on my writing, specific to what I'm looking for, specific to what I'm doing.
00:16:33.750 --> 00:16:35.421
I have it do a, a review.
00:16:35.681 --> 00:16:39.390
Mostly it gives me feedback, and then I go in and make the edits.
00:16:39.561 --> 00:16:49.022
But it also will, you know, especially where I do, like, run-on sentences or, typos, it will call those out, and it will make suggestions for, like, how to, edit them.
00:16:49.350 --> 00:16:54.301
And so now those, in theory, may also include a kind of AI watermark.
00:16:54.311 --> 00:17:13.057
And so you raise this question that comes up in, in all sorts of situations where it's like, if I write the whole thing, have an AI review it, and it catches a typo and says,"You should rewrite this sentence," and I rewrite it the way it suggests, and then somebody runs that through an AI checker, do then I get accused of, it being AI?
00:17:13.462 --> 00:17:22.242
Or you can go even further where there are all sorts of stories of people who have legitimate reasons for using these AI tools to help them write.
00:17:22.242 --> 00:17:29.353
You have people who are not fluent in a language who might use it to im- improve their fluency to make themselves more understood.
00:17:29.653 --> 00:17:36.251
I've heard of, stories of people with disabilities who use the technology to help them in, in other ways.
00:17:36.662 --> 00:17:48.215
And now if we're suddenly, putting this sort of scarlet letter on anything that was processed in any way by that can cause real damage as well.
00:17:48.256 --> 00:17:56.179
And so I think there's a lot of different tricks and, questions raised by like how well this will work and what it's doing.
00:17:56.647 --> 00:18:20.633
I think it's funny because like part of the, EU Commission's guidelines, I was looking this over in preparing for this, is they were saying that, people should know when they are interacting with AI or exposed to AI-generated content and it says this will help them make informed decisions, calibrate their trust and reliance on AI, and avoid misinformation or deception, which is a worthy goal.
00:18:21.143 --> 00:18:25.840
But I'm wondering, does this actually do that or does this make this worse?
00:18:25.840 --> 00:18:40.221
Because now suddenly you're going to have all these people who are freaking out about these things, accusing people of,"Oh, you just used AI to generate this," and actually sowing more mistrust and distrust of content that maybe doesn't deserve that.
00:18:40.352 --> 00:18:52.558
And so there's a lot of different factors that go into this and, and, Anthropic has to do this and, and Google and, OpenAI and others in some sense will have to do this to comply with the EU's new regulations.
00:18:52.909 --> 00:18:58.321
But I, I'm wondering if the long-term impact of this is actually, more trust or less
00:18:58.653 --> 00:18:58.983
Yeah.
00:18:58.983 --> 00:19:01.943
I mean, it gets to the heart of kind of what authorship is, doesn't it?
00:19:01.943 --> 00:19:03.923
And, and who owns a piece of work.
00:19:03.923 --> 00:19:16.884
And I think that's why it kinda gets under so many people's skins because e- everyone, you know, feels, I think quite strongly about, if they're creating original content and using a tool to augment it in some way, it's still theirs.
00:19:17.433 --> 00:19:20.327
my question would be like, Is this gonna matter in a couple of years' time?
00:19:21.191 --> 00:19:32.810
you know, i- i- are we not gonna be in a world where everyone is using these tools so ubiquitously that actually whether it's kind of been edited or authored by Claude it's kind of neither here nor there?
00:19:33.040 --> 00:19:34.500
And I, I feel like we're almost...
00:19:34.811 --> 00:19:47.394
there are elements of that already, I feel, where we don't know if somebody has kind of authored a, a LinkedIn post or a blog post, or if they've come up with a, an idea for a podcast, let's say.
00:19:47.752 --> 00:19:58.699
and yes, this, the way that Anthropic are approaching this isn't going to help clarify who's originated the content and who's augmented it, but everyone's going to be doing it anyway, so does it really matter?
00:19:59.146 --> 00:20:06.053
Well, you know, and actually there is an element of this which people are saying is that, you know, people are going to figure out how to strip out the watermarks.
00:20:06.053 --> 00:20:17.970
I mean, the text watermark w- with video and pictures and audio, I think it's easier to, put in a watermark that is harder to remove, though I, I imagine people will quickly develop tools that will remove those watermarks.
00:20:18.210 --> 00:20:29.955
With the text ones, it's very, very, like, it's going to be trickier, but it, it's trickier to build in the watermark in the first place, and so therefore it will be easier to figure out ways to remove it.
00:20:29.955 --> 00:20:46.146
And Anthropic has even said editing a piece of text that's generated by Claude may or may not remove the watermark, which is, is sort of an indication of how it works, is going to be certain orders of words or, you know, punctuation or something that's, that gives some sort of indication.
00:20:46.146 --> 00:20:52.480
So the editing will depend on whether or not you disturb the elements of the watermark enough, to make it invisible.
00:20:52.480 --> 00:20:59.356
But that also means you could accidentally, in some cases, create text that would set off the watermark when it doesn't.
00:20:59.688 --> 00:21:31.230
But also, I think what will happen is that those in the know, those who you know, have more, resources and ability and knowledge are probably going to figure out ways to avoid the watermark or avoid being accused of using this stuff, whereas people, with less resources are more likely to be called out for using this stuff, even if the more, powerful or wealthy or whoever are using the tools in the same way.
00:21:31.421 --> 00:21:47.382
And so I think you, you actually create a, bigger divide where, yes, lots of people, almost everyone is going to be using these tools in some form or another, but, those who are well-resourced may figure out ways to not get accused of it, while those who are not will get accused of it.
00:21:47.392 --> 00:21:56.462
And so you sort of create this, further version of a digital divide by, you know, having this kind of witch hunt sort of situation going on.
00:21:56.732 --> 00:21:58.853
So I, I worry about that as well.
00:21:58.867 --> 00:21:59.178
Yeah.
00:21:59.498 --> 00:22:05.133
And, and I mean, I, feel like those, that divide is already in place actually.
00:22:05.133 --> 00:22:13.182
You know, I think, I think there's-- this obviously has the potential to exacerbate it, but I think the, there is a kind of group of haves and have-nots in, in the AI race.
00:22:13.232 --> 00:22:25.529
And, you know, this discussion of, who will become an underclass if they can't use AI tools to the same extent as, Sam Altman and his merry band of CEOs is, has been a long kind of standing discussion.
00:22:25.925 --> 00:22:31.443
the question is really whether this is just a kind of compliance move or whether there's actually kind of belief that this will help.
00:22:31.789 --> 00:22:45.602
That's what it kind of comes down to really, doesn't it, is does Anthropic feel like this is going to alleviate some of the, the questions around, around whether people are producing content of their own or, or just cribbing, as the person who I spotted on LinkedIn did?
00:22:46.031 --> 00:22:48.132
Or whether it's just a kind of thing to keep EU happy.
00:22:48.607 --> 00:22:49.397
It feels like the latter.
00:22:49.691 --> 00:22:55.270
yeah, I think this is, this is purely driven by the EU regulations and the compliance aspect of it.
00:22:55.570 --> 00:23:02.509
You know, there are services out there that right now that claim to, they will recognize AI-generated content.
00:23:02.799 --> 00:23:06.467
I've pointed out, mistakes that I think that technology has made.
00:23:06.707 --> 00:23:20.005
It's sort of funny because those technologies also rely on AI, and it's, somewhat amusing at times to see people who absolutely hate AI pointing to those services as proof that, you know, this or that is AI slop.
00:23:20.267 --> 00:23:24.355
and it, it's all just like, it's all AI up and down the stack.
00:23:24.657 --> 00:23:31.480
and, I think just the attitude of, thinking that we have to blame people for using these tools.
00:23:31.641 --> 00:23:38.528
Like, again, I get that, and I have called out, you know, when I've seen AI-generated slop, right?
00:23:38.538 --> 00:23:42.107
Like, there are times where you look at it and you're like,"Oh my God," you know,"have a little self-respect.
00:23:42.107 --> 00:23:52.391
Why are you doing this?" but there are also times where I think, there are good reasons for people to use the technology and, to help them do certain things that they might not otherwise be able to do.
00:23:52.760 --> 00:24:15.391
And so, to me, like so many things, like this is the same argument I've made in the past in regarding other technologies, is that this is a new technology that changes the way people around the globe interact and do things, and we as a society need to learn how to deal with that properly and to understand it and to come to terms with what it means and what we think of it.
00:24:15.771 --> 00:24:29.059
and, you know, I think rushing to the space where it's just like,"Oh, we need to, call out people for using the technology and potentially bash them for it," is not gonna be particularly helpful in, in the long
00:24:29.281 --> 00:24:30.912
Yeah, I think it's-- that's true.
00:24:30.912 --> 00:24:41.094
I mean, the AI Act obviously was fairly early on in terms of trying to kind of put some guardrails and, and some kind of structure around how AI, the AI industry would evolve.
00:24:41.693 --> 00:24:49.363
at the same time, you know, there are already news outlets that have reporters who are solely AI personas, right?
00:24:49.363 --> 00:24:50.042
And who, who,
00:24:50.489 --> 00:24:51.979
Yeah, it's just garbage though.
00:24:52.962 --> 00:24:55.292
I mean, some of it is and some of it isn't.
00:24:55.363 --> 00:25:03.499
I think there, there's a kind of variety of, of, of quality in there, and some people are leaning into it more, and some of it is, is obviously kind of fairly weak.
00:25:03.855 --> 00:25:11.430
But it, again, this idea that the kind of genie's out of the bottle, i- in many ways, and this is like a kind of, feels like a band-aid on, on, on that-
00:25:11.731 --> 00:25:12.182
yeah.
00:25:12.221 --> 00:25:26.902
And r- right, you know, to, to be fair, right, there are certain examples where you could see where the AI can be useful, and there are stories where it's like you have, like, local newspapers that don't have enough resources to cover things like local high school sports or something like that.
00:25:27.122 --> 00:25:36.042
And if, if you can get stories out of it that wouldn't get covered otherwise, like, maybe that is a better solution than just not covering the, the space at all.
00:25:36.289 --> 00:25:51.682
but yeah, it's just a question of like, what is society's attitude towards all these things as well, and does something like this, increase thoughtful approaches to using the technology, or does it just create this kind of like, we're gonna burn these people at the stake because they use this
00:25:52.347 --> 00:25:52.718
Yeah.
00:25:52.907 --> 00:25:53.228
Yeah.
00:25:53.228 --> 00:25:56.768
kind of a further way that we can be polarized as a society.
00:25:57.131 --> 00:25:58.111
Yeah, yeah
00:25:58.548 --> 00:26:01.988
talking of which, a gentle segue into our eye-catching section.
00:26:02.397 --> 00:26:04.579
Um, I wondered, you know, what,
00:26:04.739 --> 00:26:06.479
Speaking of people we can burn at the stake.
00:26:08.279 --> 00:26:20.473
I wasn't quite as explicit as that, but, um, I, I, I wonder what would come up if you put Mark Zuckerberg's 6,000-word just kind of novella on AI through a watermarking tool.
00:26:20.723 --> 00:26:28.786
I wonder just how much of it might be created by whatever, model he uses, probably Muse Spark, but we don't know.
00:26:28.894 --> 00:26:31.565
because everyone has been talking about this.
00:26:31.836 --> 00:26:33.355
Some people are saying it's an essay.
00:26:33.695 --> 00:26:36.207
I think it's a kind of badly edited monologue.
00:26:36.605 --> 00:26:45.377
his essay entitled"The Future Is for Everyone," and if watermarking kind of divided opinion, particularly among the LinkedIn crowd, then this has as well, hasn't it?
00:26:45.807 --> 00:26:46.307
Yeah.
00:26:46.613 --> 00:26:55.289
you know, I, I refer to it as a manifesto, um, because, yeah, it has that sort of feel where, you know, the future is for everyone.
00:26:55.637 --> 00:27:02.509
and, you know, I think a lot of people have been, completely bashing it, mostly for good reason.
00:27:02.814 --> 00:27:10.026
you know, my take on it, he, he talks about a whole bunch of different things, but basically he's just like a very strong, he supports AI.
00:27:10.036 --> 00:27:14.346
He thinks AI should be accessible and available for everyone.
00:27:14.696 --> 00:27:24.807
you know, he always refers to super intelligence, which is this horrible, stupid, annoying, misleading term that, that he has really embraced when it comes to AI technologies.
00:27:25.172 --> 00:27:36.492
but he thinks that more people should have access to it and that they should have access to local, open weight, distributed models, which I agree with.
00:27:36.542 --> 00:27:40.969
And so there's like a lot in there that I'm reading and I was like, yeah, he's right.
00:27:40.979 --> 00:27:58.352
He's making these arguments that I think are actually correct, surrounded by an absolute ton of bullshit and nonsense about how amazing data centers are and how they're you know, revitalizing communities and all this nonsense.
00:27:58.696 --> 00:28:11.448
and then, the larger point though is, and I wrote something about this this week as well, is that he talks about the importance of decentralization and how the power shouldn't be in the control of just three giant centralized companies.
00:28:11.938 --> 00:28:26.438
And I'm just like,"Who-- Y-you're Mark Zuckerberg." Like, the whole reason we're concerned about the giant power of three or four centralized companies is partly because of your doing and things that you have done for the past 20 years.
00:28:26.478 --> 00:28:46.433
Who the fuck are you to, like, come in there and say, like,"Oh," like,"the world, it doesn't end well when new technologies are just in the hands of a few companies." It's like for 20 years, you've taken steps to make sure that Meta is the centralized giant, and you've, actively prevented decentralization within the the social space.
00:28:46.784 --> 00:28:54.714
So I wrote something basically saying, like, there are things in here that he's right about, but you should not trust him to be the messenger on it.
00:28:54.865 --> 00:28:58.694
And the only reason he's saying these things is because they're losing the race.
00:28:58.694 --> 00:29:09.752
I mean, we've talked about this before, that you know, Meta is way behind the other models, and therefore he has to put forth this idea that like, yes, the open decentralized models matter.
00:29:10.012 --> 00:29:18.838
But the second that Meta were ever close to the top of the race again, he would shut things down so fast because that's what he's done.
00:29:18.878 --> 00:29:29.234
And, and I pointed to the example of Threads and ActivityPub, where when they launched Threads, they were like, we believe in openness, we believe in the open social network where it should be on protocols.
00:29:29.234 --> 00:29:47.463
We shouldn't have control." And then as soon as Threads suddenly took off and became much bigger because they shoved everybody in Instagram into it, then suddenly they're like,"Oh yeah, ActivityPub, yeah, we'll, we'll maybe support it." And like everybody who was working on that basically left Threads and they don't really care about it anymore.
00:29:47.813 --> 00:29:49.402
this is all just marketing.
00:29:49.651 --> 00:29:56.682
And so I think some people are dismissing everything that's in there I do think there are some interesting things that are in there that are worth paying attention to.
00:29:56.853 --> 00:30:04.508
I just, you know, Mark is the wrong messenger for it because you shouldn't trust him'cause he's, he's not, he's not believable on this.
00:30:04.508 --> 00:30:05.938
He's not, he has no credibility.
00:30:06.262 --> 00:30:09.492
And he thinks he'll have amnesia from the last two decades.
00:30:09.682 --> 00:30:25.163
I think, uh, we've got a couple other stories we wanna touch on before we- Mm we do this, but, uh, the perfect visual metaphor for this essay, Mike, is, a video that Zuckerberg put out on his socials this week of him sparring with a UFC fighter on a barge.
00:30:25.213 --> 00:30:25.933
Have you seen this?
00:30:26.317 --> 00:30:27.166
No, I have not.
00:30:27.292 --> 00:30:34.847
It's, you know, you, you talked about there kind of being like a couple of ideas in, amongst a sea of terrible, terrible takes in this essay.
00:30:35.137 --> 00:30:36.185
Well, there's a, a video.
00:30:36.185 --> 00:30:38.165
You know he's a kinda big MMA guy now.
00:30:38.865 --> 00:30:49.130
There's a video of him taking on the number one UFC fighter on a barge in the middle of the ocean, and It's a kind of perfect visual representation of this.
00:30:49.140 --> 00:31:02.236
Like, he is, awash with, bad ideas and bad takes, and he is standing in the middle of the ocean swiping away, thrashing away, hoping for something to kind of, come good, and that for me sums up where Meta is in the AI race right now.
00:31:02.276 --> 00:31:04.155
I'll, I'll, I'll post it in the show notes.
00:31:04.705 --> 00:31:18.521
talking of kind of, AI companies trying to show themselves in a good light, an eye-catching story that I saw this week, Mike, and I wanted to talk through, was OpenAI's head of ethics leaving after one year in post.
00:31:18.882 --> 00:31:18.892
Mm.
00:31:18.942 --> 00:31:40.369
Now, this is a woman called Chloe Bakalar, who people who are very, very in the weeds on the industry will know was the head ethicist or chief ethicist of Meta for I think five or six years, and joined OpenAI last year as part of an effort to, her job was to kind of create, processes for, for model development, in ethical ways.
00:31:40.380 --> 00:31:43.789
She was thinking a lot about how humans interacted with AIs.
00:31:44.109 --> 00:31:52.541
You know, all the kind of, like, slightly philosophical questions that I guess don't get answered by teams who are working on research or, product or the technology itself.
00:31:53.031 --> 00:31:59.549
She, is leaving, and the company have announced that she's not going to be replaced, which I think is interesting.
00:31:59.978 --> 00:32:01.147
you know, I have a couple of thoughts on this.
00:32:01.147 --> 00:32:21.818
I mean, not necessarily very surprising in the run-up to an IPO that you might shed one of the people who wants to do very thoughtful deliberation o- on, on models and maybe, like, have a couple more meetings and ask some kind of pointed questions of, developers about how they're, structuring the data and developing these models.
00:32:21.997 --> 00:32:30.949
kinda makes sense that you, you might shepherd that person out the door, and, the justification that they've given is that ethics is now kind of baked into a bunch of teams within OpenAI.
00:32:30.949 --> 00:32:35.479
It's, it's core, yeah, core to a lot of what, the research they do.
00:32:35.669 --> 00:32:36.009
And, you
00:32:36.113 --> 00:32:40.563
Who needs ethics when you have, uh, you know, ChatGPT handling ethics for
00:32:40.800 --> 00:32:42.181
Yeah, for sure.
00:32:42.455 --> 00:32:50.958
and you might be g- be able to get on board the idea that there's a product team, who is, taking all the work that Chloe's done over the last year and kind of implementing it in new ways.
00:32:51.217 --> 00:33:05.619
the other kind of question is, like, whether this is a further distinction between OpenAI and its big rivals, Anthropic, which obviously has done lots of work around, the kind of philosophy, of models and how models behave and the morals that a model has.
00:33:06.006 --> 00:33:14.288
they have a philosopher who's been responsible for creating Claude's soul, if a model can have a soul, and we've talked about that in the past.
00:33:14.288 --> 00:33:28.823
And they've done lots of work, with the, kind of Pope and the Vatican City, and kind of really just, I guess, trying to kind of instill the idea that they are a company that has, models that kind of think differently or act differently in certain scenarios.
00:33:29.182 --> 00:33:36.587
And so, you th-this feels like another fork in the road between what OpenAI are doing and what Anthropic have tried to do or to kind of show.
00:33:37.078 --> 00:33:42.290
Now, I don't for one think that Anthropic is entirely whiter than white or the good guys here.
00:33:42.661 --> 00:33:54.097
Um, but I think it's important to note, and, uh, again, we have in technology companies for the past 15 or 20 years had ethicists, people who have been thinking about kind of big ethical questions.
00:33:54.404 --> 00:33:58.726
in the early days, I don't think they had philosophy degrees, or had, were actually trained in ethics.
00:33:58.736 --> 00:34:03.103
I think the early, early ones were, like, comms officers and people who were kind of shoved into that role.
00:34:03.123 --> 00:34:08.744
So maybe we're further on than, we were, but I think this is a kind of noted development as, as, those IPOs come along.
00:34:08.925 --> 00:34:09.186
Yeah.
00:34:09.186 --> 00:34:34.583
I mean, the thing that I would point out is like for most tech companies, and I don't know the details here, I don't know that any of this is public, but you know, for most tech companies, the way it works when you join is you get, an equity grant that vests over time, and most of it is you don't get anything until you've been there for one full year, and then you get, there's a cliff, and you get your first, you know, quarter of the equity grant at one year.
00:34:34.733 --> 00:34:46.092
So somebody leaving before an entire year is up is sort of eyebrow-raising, especially right before an IPO that in theory would make one who holds OpenAI equity very, very wealthy.
00:34:46.394 --> 00:34:48.760
so there's, definitely an eyebrow raise there.
00:34:48.789 --> 00:34:56.503
Leaving before one year is noteworthy in that it suggests something, something went wrong.
00:34:56.954 --> 00:35:04.155
Uh, whether it's with OpenAI, whether it was with this individual, who knows or some combination of both perhaps.
00:35:04.155 --> 00:35:07.286
But like, to me that, is eyebrow-raising.
00:35:07.327 --> 00:35:17.809
I think, the whole thing of like companies hiring ethicists is interesting, but hiring an, you know, an ethics specialist doesn't make you an ethical company.
00:35:17.960 --> 00:35:21.630
It's often, it feels like a marketing exercise to some extent.
00:35:21.679 --> 00:35:26.570
Like actually being ethical is different than just, hiring an e-ethicist.
00:35:26.630 --> 00:35:29.166
and like, different people can have different takes.
00:35:29.166 --> 00:35:38.295
Like the whole nature of ethics is that, there is no exact right answer, and different people will disagree on what is the ethical thing to do in different scenarios.
00:35:38.710 --> 00:35:42.625
but yeah, you know, it's definitely raises some, some eyebrows about what's happening at OpenAI.
00:35:43.083 --> 00:35:43.992
Yeah, indeed.
00:35:44.407 --> 00:35:51.677
and then just before we, we wrap up, you want to mention a, an eye-catching story that you think has a lot of implications going forward.
00:35:51.686 --> 00:35:53.277
Do you wanna just give us a quick overview of that?
00:35:53.293 --> 00:35:53.862
Yeah.
00:35:54.103 --> 00:36:08.971
it is a quick one because, uh, the Ninth Circuit made a ruling, uh, Meta and, TikTok had, challenged, it's really a series of, there's like thousands of cases that the, the sort of child safety cases and everything else is, that's going on.
00:36:09.304 --> 00:36:15.125
and they had been denied a Section 230, dismissal early on, and they had appealed it.
00:36:15.454 --> 00:36:38.309
And the Ninth Circuit, which is the appeals court that covers most of the western part of the United States, is very important, ruled that Section 230 is not actually an immunity from suit, but rather merely a defense to liability, and said effectively that companies cannot appeal the loss of a 230.
00:36:38.338 --> 00:36:45.365
You know, they-- you try and get a case dismissed on 230 grounds, and historically, if you've lost that, you can immediately appeal it.
00:36:45.666 --> 00:36:50.715
And here, the Ninth Circuit is saying you can't immediately appeal it because it is not an immunity from suit.
00:36:51.081 --> 00:37:10.260
that is, it feels like a very narrow ruling, but it is an incredibly important one because the fact that Section 230 has been considered an immunity from lawsuit, which I think is the most obvious read of what the law actually says, has been the reason why Section 230 is so important.
00:37:10.690 --> 00:37:21.699
And now that has been changed, you know, the court kinda claims that this was never really, analyzed, and other courts have said differently in the past, but they sort of dismissed that as sloppy language.
00:37:21.949 --> 00:37:35.246
I saw somebody went through and did analysis saying there were 442 court rulings that have s- called Section 230 an immunity, and for the Ninth Circuit to just kind of dismiss it as sloppy.
00:37:35.476 --> 00:37:57.556
It also sort of rewrites the history of Section 230 in that, you know, they claim that, the original version of the language didn't include this language suggesting it was an immunity, and therefore, when they added that language, they said it was a minimal change, and if this really was about adding immunity, they wouldn't have called it a minimal change.
00:37:57.856 --> 00:38:00.416
That is also a complete rewriting of the history.
00:38:00.425 --> 00:38:02.186
Jeff Kosseff spoke out publicly.
00:38:02.186 --> 00:38:06.686
Jeff Kosseff, who wrote the book on the history of Section 230, said that's wrong.
00:38:06.695 --> 00:38:15.753
That's like, you know, the people who were there in the room deliberately said that they made this change on purpose to create to make Section 230 an immunity from lawsuits.
00:38:16.155 --> 00:38:18.072
so this could be a monumental change.
00:38:18.072 --> 00:38:22.063
I think it, takes away much of the value of Section 230.
00:38:22.284 --> 00:38:26.112
the question is will Meta and/or TikTok challenge this?
00:38:26.121 --> 00:38:30.237
They can ask the Ninth Circuit to rehear it, banc, which is with the larger panel.
00:38:30.467 --> 00:38:31.807
They could go to the Supreme Court.
00:38:32.025 --> 00:38:44.626
both of those are fraught with risks and challenges because you know, the Ninth Circuit at a larger scale could also agree with this, or the Supreme Court could as well, and that would be very bad for the future of Section 230.
00:38:44.757 --> 00:38:45.117
Mm.
00:38:45.358 --> 00:38:54.117
And do you, so do you think, Mike, this is the kind of recent lawsuits against these platforms being rewritten into the kind of legal system?
00:38:54.117 --> 00:39:00.666
It's, you know, it's, it's New Mexico, it's California, it's the product design, issue being reapplied, isn't it?
00:39:00.958 --> 00:39:01.217
Yeah.
00:39:01.217 --> 00:39:15.853
I mean, if you look at the last five years of court decisions related to Section 230, one by one, courts have come up with new interpretations of Section 230 that effectively chip away at each of the benefits of Section 230.
00:39:16.443 --> 00:39:32.822
So bit by bit, even as Congress is talking about repealing or reforming, the courts have done that work bit by bit, changing the way they interpret the law so that each of the, underlying benefits of 230, which is that it gets rid of bad mistargeted lawsuits quickly, are going away.
00:39:32.871 --> 00:39:44.092
And I think we're going to have very long and very damaging consequences from that, but it's the courts, you know, effectively rewriting the law on their own because they're not happy about these companies.
00:39:44.481 --> 00:39:44.882
Yeah.
00:39:45.271 --> 00:39:53.181
And, and I think this is probably a good time to plug your, Section 230 podcast, Mike, that you did last year.
00:39:53.996 --> 00:39:56.523
doing my good Otherwise objectionable.
00:39:56.815 --> 00:40:06.429
not a description of, you or I, but uh, a, a kind of narrative-driven podcast that Mike did, which is, is very good on, kind of the history of Section 230.
00:40:06.438 --> 00:40:08.268
We'll include that in the show notes.
00:40:08.402 --> 00:40:10.137
More listening for you listeners.
00:40:10.476 --> 00:40:11.635
but this podcast is not done.
00:40:11.635 --> 00:40:12.516
We're not finished here.
00:40:12.612 --> 00:40:14.135
we're wrapping up the free episode now.
00:40:14.503 --> 00:40:16.838
we've talked a bit about, some really important stories.
00:40:16.838 --> 00:40:27.025
We're also gonna talk in the Patreon episode about, a major kind of social and immigration issue that's been bubbling under the surface in, in Europe over the last month.
00:40:27.306 --> 00:40:34.831
We're gonna do a similar treatment to what we did last week with the, Indian meta story and kind of unpack what happened and what it means.
00:40:34.831 --> 00:40:38.628
There's been some interesting developments this week, and, uh, looking forward to jumping into it now
00:40:40.485 --> 00:40:43.764
That's the end of this week's free edition of Ctrl Alt Speech.
00:40:44.146 --> 00:40:51.025
To hear the extended episode with additional coverage, become a Patreon supporter at patreon.com/ctrlaltspeech.
00:40:51.286 --> 00:40:53.824
That's C-T-R-L Alt Speech.
00:40:54.405 --> 00:41:00.826
If your organization is interested in sponsoring the podcast, contact us via our website at ctrlaltspeech.com.
00:41:01.264 --> 00:41:01.985
Thanks for listening
