ReadMultiplex.com: The “AI Detector” Grift Game: Why They Have Dumb Down Human Writing So As To Not Get “AI Flagged”.

WEBVTT
00:00:00.000 --> 00:00:03.140
I want you to imagine a scenario for a second
00:00:03.140 --> 00:00:05.940
and really just try to put yourself in this person's
00:00:05.940 --> 00:00:07.919
shoes. Okay. You're a college senior, right?
00:00:08.000 --> 00:00:11.060
You have just spent the last three months meticulously
00:00:11.060 --> 00:00:15.939
researching, outlining, and crafting your final
00:00:15.939 --> 00:00:18.640
capstone essay. Yeah, that's a lot of work. Right,
00:00:18.679 --> 00:00:21.460
or... Or maybe you're like a mid -level manager
00:00:21.460 --> 00:00:24.600
at a logistics firm and you've spent weeks agonizing
00:00:24.600 --> 00:00:27.679
over this crucial market analysis report for
00:00:27.679 --> 00:00:29.859
the executive board. Oh, man. Yeah. The pressure.
00:00:30.320 --> 00:00:32.679
Exactly. You poured over the primary sources.
00:00:32.859 --> 00:00:35.619
You refined every single sentence. You checked
00:00:35.619 --> 00:00:37.439
your transitions. You polished the vocabulary.
00:00:37.820 --> 00:00:40.439
You made absolutely sure the logical structure
00:00:40.439 --> 00:00:43.119
was completely flawless. And there's that singular
00:00:43.119 --> 00:00:46.119
moment of relief, right? You hit submit, you
00:00:46.119 --> 00:00:49.320
lean back in your chair, and you feel that specific
00:00:49.320 --> 00:00:51.880
rush of pride that only comes from a job genuinely
00:00:51.880 --> 00:00:55.390
well done. Yes. But the relief just doesn't last
00:00:55.390 --> 00:00:58.530
because almost instantly a piece of black box
00:00:58.530 --> 00:01:00.429
software, a program you have never seen, right?
00:01:00.490 --> 00:01:03.310
A program whose internal rules you are literally
00:01:03.310 --> 00:01:06.329
not allowed to know. It scans your months of
00:01:06.329 --> 00:01:09.629
hard work and slaps a giant neon red completely
00:01:09.629 --> 00:01:11.969
fake label across the top of it. Oh, it's just
00:01:11.969 --> 00:01:14.250
it's devastating. It calls you a fraud to your
00:01:14.250 --> 00:01:16.769
professor or to your boss. And in a matter of
00:01:16.769 --> 00:01:18.709
seconds, your grade is ruined. Your professional
00:01:18.709 --> 00:01:21.769
reputation is in tatters. And worst of all. There
00:01:21.769 --> 00:01:24.730
is no human being you can appeal to. The machine
00:01:24.730 --> 00:01:28.209
has spoken. It is a genuinely terrifying position
00:01:28.209 --> 00:01:32.049
to be in. The psychological toll of that, just
00:01:32.049 --> 00:01:34.310
being called a liar by an algorithm and having
00:01:34.310 --> 00:01:36.650
no way to defend yourself, it's devastating.
00:01:36.909 --> 00:01:38.650
And I think we need to be clear right from the
00:01:38.650 --> 00:01:40.950
start here, this is not some dystopian science
00:01:40.950 --> 00:01:43.349
fiction hypothetical. It's happening right now,
00:01:43.370 --> 00:01:46.519
this very morning. High schools, universities,
00:01:46.799 --> 00:01:48.739
and corporate HR departments all over the world.
00:01:48.799 --> 00:01:51.400
Which is exactly why we are dedicating this entire
00:01:51.400 --> 00:01:54.329
episode to unpacking this nightmare. Welcome
00:01:54.329 --> 00:01:56.489
to this deep dive for the ReadMultiplex .com
00:01:56.489 --> 00:01:59.469
podcast. Our mission today is to really zero
00:01:59.469 --> 00:02:02.250
in on a critical, incredibly urgent issue that
00:02:02.250 --> 00:02:03.750
has been brought to light by the independent
00:02:03.750 --> 00:02:06.430
voice of Brian Aremmel. Yeah, his insights on
00:02:06.430 --> 00:02:07.890
this have been incredible. They really have.
00:02:08.050 --> 00:02:10.110
And I got to say up front, you know, in a tech
00:02:10.110 --> 00:02:12.389
landscape that is just flooded with venture capital
00:02:12.389 --> 00:02:15.389
hype and PR spin, it takes someone with Brian's
00:02:15.389 --> 00:02:18.409
candor and absolute clarity to cut through the
00:02:18.409 --> 00:02:20.729
industry noise. Absolutely. Because the core
00:02:20.729 --> 00:02:22.889
thesis we are exploring today, based entirely
00:02:22.889 --> 00:02:26.680
on... his insights, is explosive. We are looking
00:02:26.680 --> 00:02:29.639
at a system that is essentially a modern day
00:02:29.639 --> 00:02:33.520
digital witch trial. The entire industry of text
00:02:33.520 --> 00:02:37.300
based AI detectors is a complete sham. And we
00:02:37.300 --> 00:02:38.979
really don't use the word sham lightly here.
00:02:39.020 --> 00:02:41.360
This isn't just a critique of a software bug
00:02:41.360 --> 00:02:43.620
or a glitch. We are talking about a systemic.
00:02:44.270 --> 00:02:47.569
foundational failure that is causing real measurable
00:02:47.569 --> 00:02:50.689
harm to innocent people. Massive harm. And we're
00:02:50.689 --> 00:02:52.310
going to look at the hard data today that proves
00:02:52.310 --> 00:02:54.189
these detectors are fundamentally mechanically
00:02:54.189 --> 00:02:56.750
broken. And furthermore, even the tools that
00:02:56.750 --> 00:02:59.169
occasionally seemed, I don't know, sort of accurate
00:02:59.169 --> 00:03:01.770
a year or two ago, they are rapidly becoming
00:03:01.770 --> 00:03:04.530
obsolete and wildly wrong. Yeah, the degradation
00:03:04.530 --> 00:03:06.590
is real. So if you are listening to this right
00:03:06.590 --> 00:03:09.419
now. Whether you are a student writing a term
00:03:09.419 --> 00:03:11.939
paper, a professional drafting a proposal, or
00:03:11.939 --> 00:03:14.439
honestly just someone who sends carefully worded
00:03:14.439 --> 00:03:17.659
emails to clients, you need to pay close attention.
00:03:18.099 --> 00:03:21.000
This digital scarlet letter could absolutely
00:03:21.000 --> 00:03:23.719
affect you tomorrow. You know, what stands out
00:03:23.719 --> 00:03:26.550
to me in the source material is how... Incredibly
00:03:26.550 --> 00:03:30.090
fast, society just adopted these policing tools.
00:03:30.349 --> 00:03:33.210
Oh, overnight. Right. Without even stopping to
00:03:33.210 --> 00:03:35.509
ask the most basic fundamental question, which
00:03:35.509 --> 00:03:38.150
is, do they actually work? We just sort of culturally
00:03:38.150 --> 00:03:40.990
accepted the premise. Figured, well, if artificial
00:03:40.990 --> 00:03:43.110
intelligence is smart enough to write human -sounding
00:03:43.110 --> 00:03:45.990
text, then surely a different artificial intelligence
00:03:45.990 --> 00:03:48.090
must be smart enough to detect it. Right. It
00:03:48.090 --> 00:03:50.960
sounds logical on the surface. It does. But to
00:03:50.960 --> 00:03:54.300
understand why these detectors fail so spectacularly,
00:03:54.300 --> 00:03:56.560
we have to look at their absolute foundation.
00:03:57.099 --> 00:03:59.099
We really have to look at what the machine considers
00:03:59.099 --> 00:04:01.360
to be human in the first place. OK, I want to
00:04:01.360 --> 00:04:02.560
stop you right there because this is where I
00:04:02.560 --> 00:04:05.120
get tripped up. When I think of a baseline for
00:04:05.120 --> 00:04:07.500
human writing, I think of, you know, the great
00:04:07.500 --> 00:04:10.620
libraries of the world. I assume these massive
00:04:10.620 --> 00:04:14.020
tech companies train their models on encyclopedias,
00:04:14.020 --> 00:04:16.600
classic literature, academic journals, maybe
00:04:16.600 --> 00:04:19.420
the New York Times archives. So if that's the
00:04:19.420 --> 00:04:21.790
baseline, how are they getting it so wrong? I
00:04:21.790 --> 00:04:23.629
can completely understand why you would assume
00:04:23.629 --> 00:04:25.709
that. It sounds like the right way to do it.
00:04:25.810 --> 00:04:29.230
But it ignores the sheer scale of how large language
00:04:29.230 --> 00:04:32.470
models or LLMs are actually built. Okay. They
00:04:32.470 --> 00:04:35.050
didn't just read the classics. They essentially
00:04:35.050 --> 00:04:38.550
hoovered up the entire open Internet. Oh. And
00:04:38.550 --> 00:04:40.529
if you think about the Internet. by sheer volume
00:04:40.529 --> 00:04:43.589
of text, it is not a curated library of high
00:04:43.589 --> 00:04:46.610
literature. The vast overwhelming majority of
00:04:46.610 --> 00:04:50.009
human text online is just casual, low -effort
00:04:50.009 --> 00:04:51.769
writing. Oh, I see where this is going. We're
00:04:51.769 --> 00:04:53.310
talking about comment sections, aren't we? We
00:04:53.310 --> 00:04:55.310
are talking about the sewage standard. The training
00:04:55.310 --> 00:04:58.089
data is dominated by Reddit posts, furious late
00:04:58.089 --> 00:05:01.029
night forum rants, sloppy personal blogs, social
00:05:01.029 --> 00:05:03.790
media comments, and quick text messages. It is
00:05:03.790 --> 00:05:07.050
that endless stream of consciousness, yo bro,
00:05:07.290 --> 00:05:11.329
casual style of communication. Because this disorganized
00:05:11.329 --> 00:05:14.069
internet sludge makes up the statistical bulk
00:05:14.069 --> 00:05:16.629
of the training data, these systems now treat
00:05:16.629 --> 00:05:19.389
that chaotic style as the mathematical gold standard
00:05:19.389 --> 00:05:22.259
of what a human sounds like. Wait. Let me make
00:05:22.259 --> 00:05:23.959
sure I'm wrapping my head around this. Because
00:05:23.959 --> 00:05:26.360
there is so much literal garbage writing on the
00:05:26.360 --> 00:05:28.899
Internet, the detector's algorithm assumes that
00:05:28.899 --> 00:05:31.860
if a real -life human wrote a document, it simply
00:05:31.860 --> 00:05:34.750
must read like garbage. In a statistical sense,
00:05:34.930 --> 00:05:37.509
that is exactly how the math works. That is insane.
00:05:38.029 --> 00:05:40.250
It is. The detectors are scanning for something
00:05:40.250 --> 00:05:42.649
they call perplexity and random burstiness. They
00:05:42.649 --> 00:05:44.730
expect human writing to be erratic. They want
00:05:44.730 --> 00:05:47.189
to see a total lack of structural predictability.
00:05:47.269 --> 00:05:49.850
Okay, so they want mistakes. Exactly. They are
00:05:49.850 --> 00:05:52.149
looking for a sentence that runs on way too long
00:05:52.149 --> 00:05:54.689
with terrible comma usage, immediately followed
00:05:54.689 --> 00:05:57.670
by a three -word fragment. They expect the vocabulary
00:05:57.670 --> 00:06:00.790
to vary wildly in a completely disorganized way.
00:06:01.310 --> 00:06:03.839
Therefore, if the algorithm encounters anything
00:06:03.839 --> 00:06:07.279
that is highly formal very careful rigorously
00:06:07.279 --> 00:06:09.420
structured or just uses a consistent professional
00:06:09.420 --> 00:06:12.560
tone it immediately flags it as suspicious that
00:06:12.560 --> 00:06:15.319
is just it's the tragedy of good writing you
00:06:15.319 --> 00:06:17.740
are literally telling me that a student or professional
00:06:17.740 --> 00:06:20.339
is being actively penalized by the software for
00:06:20.339 --> 00:06:22.660
being excellent at their job the better you write
00:06:23.439 --> 00:06:25.800
the more you look like a machine to these detectors.
00:06:26.060 --> 00:06:28.620
Wow. And this is not just a theoretical edge
00:06:28.620 --> 00:06:31.079
case, right? The sources highlight that this
00:06:31.079 --> 00:06:33.480
flaw is baked into the very DNA of the product.
00:06:33.639 --> 00:06:36.360
If you feed these detectors highly vetted, beautifully
00:06:36.360 --> 00:06:39.879
written human text from long before ChatGPT or
00:06:39.879 --> 00:06:42.699
any LLM ever existed, the results are completely
00:06:42.699 --> 00:06:45.329
absurd. Like what? Passages from the Bible, for
00:06:45.329 --> 00:06:47.829
example, regularly score as heavily AI generated.
00:06:48.069 --> 00:06:50.370
Wow. The Quran gets flagged constantly. The U
00:06:50.370 --> 00:06:52.149
.S. Constitution has been flagged over and over
00:06:52.149 --> 00:06:53.750
again as being written by a machine. Wait, the
00:06:53.750 --> 00:06:56.209
U .S. Constitution? Yeah. So the detector looks
00:06:56.209 --> 00:06:58.410
at the foundational document of the United States
00:06:58.410 --> 00:07:01.250
and concludes James Madison was basically using
00:07:01.250 --> 00:07:04.810
an early beta of generative AI. That is the conclusion
00:07:04.810 --> 00:07:07.569
the math forces it to make. The algorithm scans
00:07:07.569 --> 00:07:11.290
peak human writing from the 1950s. It scans Shakespeare.
00:07:11.490 --> 00:07:14.430
It scans formal academic papers written decades
00:07:14.430 --> 00:07:17.790
before modern AI was even a concept. All of it
00:07:17.790 --> 00:07:21.290
gets marked as AI. That is wild. There is a famous
00:07:21.290 --> 00:07:22.990
example highlighted in the sources where the
00:07:22.990 --> 00:07:26.790
tool ZeroGPT scored the Book of Genesis at roughly
00:07:26.790 --> 00:07:31.810
88 % AI. 88 % AI for the literal Book of Genesis.
00:07:32.189 --> 00:07:34.240
Yeah. I'm trying to think of an analogy here.
00:07:34.319 --> 00:07:37.060
It's like it's like grading a Michelin star chef
00:07:37.060 --> 00:07:39.680
on a curve based entirely on fast food drive
00:07:39.680 --> 00:07:41.959
through menus. Right. Exactly. The software is
00:07:41.959 --> 00:07:44.199
judging a perfectly plated duck confit and saying,
00:07:44.240 --> 00:07:45.879
well, there's no processed cheese on this and
00:07:45.879 --> 00:07:48.139
it doesn't come in a greasy paper bag. So a human
00:07:48.139 --> 00:07:50.379
chef definitely didn't make it. It must be synthetic.
00:07:50.819 --> 00:07:53.560
If you write too well, if you use classic structure
00:07:53.560 --> 00:07:56.420
and repetition for emphasis, you are. penalized
00:07:56.420 --> 00:07:58.740
for not sounding like a 2 a .m. rant on a gaming
00:07:58.740 --> 00:08:01.560
forum. That Michelin star analogy hits the nail
00:08:01.560 --> 00:08:03.759
on the head. Think about the book of Genesis
00:08:03.759 --> 00:08:06.920
or really any classic literature or ancient religious
00:08:06.920 --> 00:08:10.379
text. It is highly formal. It uses deliberate
00:08:10.379 --> 00:08:13.639
repetitive structures for emphasis and oral memorization.
00:08:13.819 --> 00:08:15.920
Right. The begats and all that. Right. It is
00:08:15.920 --> 00:08:19.560
incredibly low in that random burstiness the
00:08:19.560 --> 00:08:21.660
detectors are mathematically trained to expect
00:08:21.660 --> 00:08:24.860
from modern Internet sludge. So the algorithm
00:08:24.860 --> 00:08:27.259
sees this clean, structured, purposeful writing
00:08:27.259 --> 00:08:30.199
and computes. A human couldn't possibly be this
00:08:30.199 --> 00:08:32.340
organized and disciplined. This must be a machine.
00:08:32.659 --> 00:08:35.200
I have to ask, how do the companies that sell
00:08:35.200 --> 00:08:37.960
these detectors respond to that? Because if I'm
00:08:37.960 --> 00:08:40.600
a tech CEO, it is obviously deeply humiliating
00:08:40.600 --> 00:08:42.720
to have my flagship product claim the Bible was
00:08:42.720 --> 00:08:45.200
written by a Silicon Valley chatbot. Do they
00:08:45.200 --> 00:08:47.740
just deny it? The industry response is where
00:08:47.740 --> 00:08:50.340
the deep cynicism of this whole enterprise really
00:08:50.340 --> 00:08:53.139
starts to show. To avoid looking completely ridiculous
00:08:53.139 --> 00:08:55.240
in the press, these companies have had to go
00:08:55.240 --> 00:08:57.620
into their software and manually add exceptions
00:08:57.620 --> 00:09:01.379
to their code. No. They essentially whitelist
00:09:01.379 --> 00:09:05.000
common, famous texts. They have to tell the software,
00:09:05.200 --> 00:09:07.279
hey, if you see the Declaration of Independence,
00:09:07.539 --> 00:09:09.879
just ignore your own mathematical analysis and
00:09:09.879 --> 00:09:12.700
label it human. Just to save face. Exactly. So
00:09:12.700 --> 00:09:14.519
they don't get caught making those specific,
00:09:14.700 --> 00:09:17.919
highly viral, embarrassing claims. I feel like
00:09:17.919 --> 00:09:20.519
I'm taking crazy pills here. If you have a core
00:09:20.519 --> 00:09:22.960
algorithm and you have to manually tell the software,
00:09:23.080 --> 00:09:25.940
no, no, ignore your own analysis on this specific
00:09:25.940 --> 00:09:28.100
book because the results make our company look
00:09:28.100 --> 00:09:30.679
like a joke, doesn't that inherently prove the
00:09:30.679 --> 00:09:33.379
underlying algorithm is entirely flawed? You
00:09:33.379 --> 00:09:35.779
are admitting the math doesn't work. It proves
00:09:35.779 --> 00:09:37.940
exactly that. Their baseline is fundamentally
00:09:37.940 --> 00:09:40.799
broken. The exceptions list has to keep growing
00:09:40.799 --> 00:09:43.419
quietly in the background just so they can maintain
00:09:43.419 --> 00:09:47.039
the illusion of accuracy. And that illusion is
00:09:47.039 --> 00:09:49.519
what allows them to continue farming millions
00:09:49.519 --> 00:09:52.419
of dollars in licensing fees from schools and
00:09:52.419 --> 00:09:55.460
universities. They are selling a surefire AI
00:09:55.460 --> 00:09:58.259
solution to terrified administrators, sometimes
00:09:58.259 --> 00:10:00.440
even getting their tools reinforced by state
00:10:00.440 --> 00:10:03.679
educational mandates. But the foundation is entirely
00:10:03.679 --> 00:10:06.320
makeshift. It's putting a flimsy digital band
00:10:06.320 --> 00:10:09.500
-aid on a gaping foundational wound. I hear what
00:10:09.500 --> 00:10:11.240
you're saying. And the Bible example is hilarious
00:10:11.240 --> 00:10:14.299
and damning. But I can easily imagine like a
00:10:14.299 --> 00:10:16.259
skeptical school board administrator listening
00:10:16.259 --> 00:10:18.240
to this and saying, sure, maybe the software
00:10:18.240 --> 00:10:21.059
gets confused by ancient archaic text because
00:10:21.059 --> 00:10:24.059
nobody talks like King James anymore. But we
00:10:24.059 --> 00:10:26.440
aren't grading the Bible. We're grading modern
00:10:26.440 --> 00:10:29.299
everyday high school essays. Surely it works
00:10:29.299 --> 00:10:32.379
for normal 21st century writing. Tell me we don't
00:10:32.379 --> 00:10:34.539
just have to rely on anecdotal examples of old
00:10:34.539 --> 00:10:36.320
books. We don't have to rely on anecdotes at
00:10:36.320 --> 00:10:38.220
all. The source material brings the receipts,
00:10:38.240 --> 00:10:39.960
and the hard science has proven this failure
00:10:39.960 --> 00:10:43.299
across the board on modern text. Multiple independent
00:10:43.299 --> 00:10:46.080
peer -reviewed studies paint the exact same ugly
00:10:46.080 --> 00:10:48.059
picture. Let's look at the Weber -Wolfe study
00:10:48.059 --> 00:10:51.299
as a starting point. They tested 14 different
00:10:51.299 --> 00:10:54.580
AI detectors on the market. Crucially, that included
00:10:54.580 --> 00:10:57.639
major, highly funded players like... Turnitin,
00:10:57.700 --> 00:11:00.299
which is absolutely ubiquitous in high schools
00:11:00.299 --> 00:11:03.039
and colleges right now, along with various free
00:11:03.039 --> 00:11:05.899
tools available online. OK, 14 different tools
00:11:05.899 --> 00:11:08.299
tested against modern writing. Let me guess.
00:11:08.360 --> 00:11:10.299
They caught the AI, but they flagged a bunch
00:11:10.299 --> 00:11:12.240
of the humans as well because they didn't write
00:11:12.240 --> 00:11:14.580
like Internet sludge. It was a failure on both
00:11:14.580 --> 00:11:17.019
sides of the equation. Every single one of the
00:11:17.019 --> 00:11:21.049
14 tools scored below 80 percent accuracy. Below
00:11:21.049 --> 00:11:23.629
80%. I want to pause on that number because in
00:11:23.629 --> 00:11:26.669
a casual mobile game, 80 % is fine. But in a
00:11:26.669 --> 00:11:29.509
high stakes environment like education or professional
00:11:29.509 --> 00:11:33.110
academic publishing, a 20 plus percent failure
00:11:33.110 --> 00:11:36.049
rate is catastrophic. If a teacher has a class
00:11:36.049 --> 00:11:38.889
of 30 kids, this software is potentially giving
00:11:38.889 --> 00:11:42.149
false readings on six of them every single time
00:11:42.149 --> 00:11:44.149
an assignment is turned in. And the nature of
00:11:44.149 --> 00:11:47.029
those failures was incredibly erratic. Most of
00:11:47.029 --> 00:11:49.090
the tools lean heavily toward calling actual
00:11:49.090 --> 00:11:53.000
AI text. Wait, really? Yeah. Which means they
00:11:53.000 --> 00:11:55.679
fail at their primary advertised job of catching
00:11:55.679 --> 00:11:59.179
cheaters. But, simultaneously, they produced
00:11:59.179 --> 00:12:01.940
enough false positives calling real, modern,
00:12:02.019 --> 00:12:05.779
human -writing AI to completely destroy any rational
00:12:05.779 --> 00:12:08.200
trust in the system. It's just a mess. It is.
00:12:08.279 --> 00:12:10.840
And what's worse, the researchers found a massive
00:12:10.840 --> 00:12:13.840
vulnerability regarding language. If a text was
00:12:13.840 --> 00:12:16.399
simply paraphrased, or if it was translated from
00:12:16.399 --> 00:12:18.519
another language by a non -native English speaker,
00:12:18.779 --> 00:12:21.519
the accuracy of these detectors tanked even further.
00:12:21.519 --> 00:12:24.120
further. Oh, man. Yeah. The tools were deemed
00:12:24.120 --> 00:12:26.720
completely inaccurate and unreliable for any
00:12:26.720 --> 00:12:29.639
serious application. Non -native speakers being
00:12:29.639 --> 00:12:31.799
penalized more heavily, that adds a whole new
00:12:31.799 --> 00:12:34.240
layer of discrimination to this. But the absolute
00:12:34.240 --> 00:12:36.059
nail in the coffin, the study that just blew
00:12:36.059 --> 00:12:37.919
my mind when I was reading through Brian Rommel's
00:12:37.919 --> 00:12:40.679
insights in these sources, is the 2026 University
00:12:40.679 --> 00:12:43.259
of Florida study. Yes, the trainer study. Yeah.
00:12:43.279 --> 00:12:45.320
If you want to see a masterclass in exposing
00:12:45.320 --> 00:12:48.250
flawed technology, this is it. I want to spend
00:12:48.250 --> 00:12:50.029
some real time on this because the setup of this
00:12:50.029 --> 00:12:52.750
experiment is just brilliant. This was led by
00:12:52.750 --> 00:12:55.750
Patrick Traynor and his team, Seth Layton, Bernardo
00:12:55.750 --> 00:12:59.610
Medeiros, and Kevin Butler at the 2026 IE Symposium
00:12:59.610 --> 00:13:02.879
on Security and Privacy. Break down exactly how
00:13:02.879 --> 00:13:05.080
they constructed this test because they eliminated
00:13:05.080 --> 00:13:07.679
any room for the detector companies to make excuses.
00:13:08.120 --> 00:13:11.620
They executed a very clever, very brutal methodology.
00:13:11.899 --> 00:13:14.399
The biggest defense detector companies use is
00:13:14.399 --> 00:13:17.039
claiming their test data was somehow tainted.
00:13:17.159 --> 00:13:19.539
Right. So Trainor's team eliminated that possibility.
00:13:19.840 --> 00:13:22.519
They went into the archives and gathered about
00:13:22.519 --> 00:13:25.679
6 ,000 real, highly technical research papers
00:13:25.679 --> 00:13:27.679
that had been submitted to top -tier security
00:13:27.679 --> 00:13:31.059
conferences. And here is the crucial part. Every
00:13:31.059 --> 00:13:33.440
single one of these 6 ,000 papers was submitted
00:13:33.440 --> 00:13:36.679
and published before JAT -GPT or any modern generative
00:13:36.679 --> 00:13:40.019
AI even existed. So they are 100 % verifiably
00:13:40.019 --> 00:13:42.159
human. There's absolutely zero question about
00:13:42.159 --> 00:13:44.399
their provenance. Zero question. They're the
00:13:44.399 --> 00:13:47.159
ultimate control group. Then the researchers
00:13:47.159 --> 00:13:49.799
used large language models to generate exact
00:13:49.799 --> 00:13:52.940
AI clones of those 6 ,000 papers. They fed the
00:13:52.940 --> 00:13:55.700
topics and structures into the AI and had it
00:13:55.700 --> 00:13:58.059
generate synthetic versions. So now they have
00:13:58.059 --> 00:14:01.600
6 ,000 guaranteed human papers and 6 ,000 guaranteed
00:14:01.600 --> 00:14:04.960
AI clones of the exact same subject matter. They
00:14:04.960 --> 00:14:07.779
took this massive data set and ran both sets
00:14:07.779 --> 00:14:09.919
through the five most popular, highly funded
00:14:09.919 --> 00:14:13.080
commercial AI text detectors on the market. This
00:14:13.080 --> 00:14:15.519
is the ultimate stress test. What happened when
00:14:15.519 --> 00:14:17.620
they tested the gold standard detectors against
00:14:17.620 --> 00:14:20.950
this perfectly clean data set? The results were
00:14:20.950 --> 00:14:23.409
an unmitigated disaster for the detector industry.
00:14:23.750 --> 00:14:26.370
Let's look at the false positives first. Again,
00:14:26.429 --> 00:14:28.490
a false positive is when the detector accuses
00:14:28.490 --> 00:14:31.710
a real human of being a machine. The false positive
00:14:31.710 --> 00:14:35.309
rates across the tools range from 0 .05 % all
00:14:35.309 --> 00:14:38.659
the way up to a staggering 68 .6%. I need to
00:14:38.659 --> 00:14:40.740
make sure I heard that right. 68 .6 percent.
00:14:40.820 --> 00:14:42.919
You are saying one of the top commercial pools
00:14:42.919 --> 00:14:45.200
on the market looked at a database of guaranteed,
00:14:45.340 --> 00:14:48.019
verified human writing and accused the human
00:14:48.019 --> 00:14:50.460
of cheating nearly 70 percent of the time. That
00:14:50.460 --> 00:14:53.700
is exactly what the data shows. Over two thirds
00:14:53.700 --> 00:14:56.960
of the time, this highly expensive software labeled.
00:14:57.519 --> 00:15:00.259
brilliant, hardworking human researchers as machines.
00:15:00.779 --> 00:15:03.159
That is, I can't even process that. But if you
00:15:03.159 --> 00:15:05.580
think that's bad, the false negatives are where
00:15:05.580 --> 00:15:08.080
the illusion completely shatters. A false negative
00:15:08.080 --> 00:15:11.620
is when the tool looks at actual 100 % AI generated
00:15:11.620 --> 00:15:14.919
text and confidently declares, yep, a human wrote
00:15:14.919 --> 00:15:18.379
this. The false negative rate swung from 0 .3
00:15:18.379 --> 00:15:23.059
% to 99 .6%. 99 .6 % false negative. Let me visualize
00:15:23.059 --> 00:15:25.440
that for a second. That means if a university
00:15:25.440 --> 00:15:40.539
professor suspects cheating, Yep. It is doing
00:15:40.539 --> 00:15:43.080
the exact opposite of what the sales brochure
00:15:43.080 --> 00:15:45.700
promises. It's essentially a random number generator
00:15:45.700 --> 00:15:48.259
at that point. It's statistical noise. It is
00:15:48.259 --> 00:15:50.909
arguably worse than a coin flip. Because a coin
00:15:50.909 --> 00:15:52.830
flip doesn't cost a university district millions
00:15:52.830 --> 00:15:55.450
of dollars in licensing fees, the detectors entirely
00:15:55.450 --> 00:15:57.889
collapsed under rigorous, controlled scientific
00:15:57.889 --> 00:16:01.429
testing. They became largely useless, completely
00:16:01.429 --> 00:16:04.049
unable to reliably tell the difference between
00:16:04.049 --> 00:16:07.509
human and machine. Patrick Treanor's conclusion
00:16:07.509 --> 00:16:10.590
was definitive and damning. What did he say?
00:16:10.789 --> 00:16:12.710
He stated that these tools are not reliable,
00:16:12.870 --> 00:16:15.009
they are not robust enough to use, and likely
00:16:15.009 --> 00:16:17.210
never will be due to the underlying mechanics
00:16:17.210 --> 00:16:19.809
of how language models work. He specifically
00:16:19.809 --> 00:16:21.950
noted that commercially available detectors are
00:16:21.950 --> 00:16:24.649
poorly suited for deployment in academic or high
00:16:24.649 --> 00:16:27.509
stakes contexts. Think about the real world impact
00:16:27.509 --> 00:16:30.090
of a school ignoring that warning. We aren't
00:16:30.090 --> 00:16:32.330
just talking about abstract numbers on a spreadsheet
00:16:32.330 --> 00:16:34.870
here. We are talking about human lives, people's
00:16:34.870 --> 00:16:37.370
careers, their academic futures, their scholarships,
00:16:37.570 --> 00:16:39.879
their standing in their community. All of it
00:16:39.879 --> 00:16:42.000
is on the line based on tools that have entirely
00:16:42.000 --> 00:16:44.639
collapsed under scientific scrutiny. They literally
00:16:44.639 --> 00:16:47.879
cannot tell the difference anymore. And yet universities,
00:16:48.220 --> 00:16:50.519
high schools and academic journals just keep
00:16:50.519 --> 00:16:53.200
buying the fantasy. They keep renewing the software
00:16:53.200 --> 00:16:56.480
licenses. Which naturally leads to the most important
00:16:56.480 --> 00:16:58.600
question we can ask about this entire situation.
00:16:58.779 --> 00:17:03.139
Why? If the hard science proves unequivocally
00:17:03.139 --> 00:17:05.940
that they do not work, why are they suddenly
00:17:05.940 --> 00:17:09.019
ubiquitous? Seriously, if they are mathematically
00:17:09.019 --> 00:17:11.960
proven to be a sham, if the trainer study exposed
00:17:11.960 --> 00:17:14.700
them this thoroughly, why is my local school
00:17:14.700 --> 00:17:18.000
district suddenly mandating them on every single
00:17:18.000 --> 00:17:21.000
assignment? To understand that massive disconnect
00:17:21.000 --> 00:17:23.980
between reality and adoption, we have to look
00:17:23.980 --> 00:17:26.259
at the financial motivation. We have to examine
00:17:26.259 --> 00:17:29.299
what the sources call the grift cycle. And the
00:17:29.299 --> 00:17:31.259
most illuminating place to start is with the
00:17:31.259 --> 00:17:33.359
very company that kicked off the generative AI
00:17:33.359 --> 00:17:36.130
boom in the first place, OpenAI. The creators
00:17:36.130 --> 00:17:38.269
of ChatGPT, I vaguely remember this, didn't they
00:17:38.269 --> 00:17:40.109
try to make their own detector a couple of years
00:17:40.109 --> 00:17:42.849
ago? They did. OpenAI built a classifier tool
00:17:42.849 --> 00:17:45.049
specifically designed to detect AI -generated
00:17:45.049 --> 00:17:47.430
text. They had the best engineers in the world,
00:17:47.569 --> 00:17:49.990
unlimited compute power, and intimate knowledge
00:17:49.990 --> 00:17:51.750
of how the models were built. But here is the
00:17:51.750 --> 00:17:53.369
reality check. They tested their own detector,
00:17:53.509 --> 00:17:56.210
and they realized it had a 9 % false positive
00:17:56.210 --> 00:17:59.549
rate, meaning it falsely accused innocent humans
00:17:59.549 --> 00:18:04.710
nearly 1 in 10 times. Simultaneously, it missed
00:18:04.710 --> 00:18:08.690
a vast fraction of actual AI text. And do you
00:18:08.690 --> 00:18:11.190
know what OpenAI did when they saw those numbers?
00:18:11.390 --> 00:18:13.490
I would hope they went back to the drawing board,
00:18:13.630 --> 00:18:15.750
but knowing tech companies, they probably shipped
00:18:15.750 --> 00:18:18.309
it in beta. No, this is the fascinating part.
00:18:18.410 --> 00:18:22.359
They shut it down completely. Really? Yeah. Ethically,
00:18:22.460 --> 00:18:24.180
they looked at the data, they looked at the harm
00:18:24.180 --> 00:18:27.440
a 9 % false accusation rate would cause in schools,
00:18:27.559 --> 00:18:31.039
and they realized the accuracy was garbage. The
00:18:31.039 --> 00:18:33.519
very company that built the underlying models,
00:18:33.759 --> 00:18:36.319
the people who understand this specific technology
00:18:36.319 --> 00:18:39.140
better than anyone on earth, publicly admitted
00:18:39.140 --> 00:18:41.240
they could not make a detector that actually
00:18:41.240 --> 00:18:43.000
works safely. Which really should have been the
00:18:43.000 --> 00:18:45.059
end of the entire conversation. If the literal
00:18:45.059 --> 00:18:47.480
creators of the technology say, hey, guys, based
00:18:47.480 --> 00:18:49.539
on the math, this is impossible to do reliably,
00:18:49.819 --> 00:18:51.920
everyone should have packed it up. But instead
00:18:51.920 --> 00:18:53.859
of listening to the engineers who built the AI,
00:18:54.140 --> 00:18:56.400
a massive cottage industry just sprang up out
00:18:56.400 --> 00:18:58.339
of nowhere to sell the fantasy anyway. That is
00:18:58.339 --> 00:19:00.319
exactly what happened. Yeah. You have all these
00:19:00.319 --> 00:19:03.700
startup companies, GPT Zero, Originality .AI,
00:19:04.099 --> 00:19:07.660
ZeroGPT, and dozens of others who rushed in to
00:19:07.660 --> 00:19:10.799
fill the void left by OpenAI because where OpenAI
00:19:10.799 --> 00:19:13.880
saw an ethical failure and a technical dead end,
00:19:14.079 --> 00:19:16.980
these startups saw a massive, completely untapped
00:19:16.980 --> 00:19:20.279
market of terrified administrators who were desperate
00:19:20.279 --> 00:19:22.839
for a silver bullet. And this brings us to what
00:19:22.839 --> 00:19:26.220
the sources describe as the perfect grift, the
00:19:26.220 --> 00:19:28.839
double dip. I need you to explain how this business
00:19:28.839 --> 00:19:30.839
model actually works in practice, because when
00:19:30.839 --> 00:19:32.880
I read this, I thought it was incredibly cynical.
00:19:33.140 --> 00:19:35.900
It is a brilliantly insidious closed loop cycle.
00:19:36.420 --> 00:19:38.740
Let's break it down into steps. Okay. Step one,
00:19:38.799 --> 00:19:40.910
these companies sell fear. They deploy sales
00:19:40.910 --> 00:19:43.349
teams to approach educational institutions, universities,
00:19:43.609 --> 00:19:46.589
and corporate compliance officers, and they deliberately
00:19:46.589 --> 00:19:49.690
manufacture a panic. They sell the idea that
00:19:49.690 --> 00:19:51.950
without their specific proprietary detector,
00:19:52.289 --> 00:19:54.450
the academy or the workplace will completely
00:19:54.450 --> 00:19:56.690
collapse into an ocean of AI sludge. Right, the
00:19:56.690 --> 00:19:59.970
sky is falling. Exactly. They convince administrators
00:19:59.970 --> 00:20:02.990
that students are cheating at unprecedented rates
00:20:02.990 --> 00:20:06.230
and that their software is the only shield against
00:20:06.230 --> 00:20:10.140
the impending doom of fake writing. So the university
00:20:10.140 --> 00:20:13.240
panics and buys a multi -million dollar, multi
00:20:13.240 --> 00:20:16.480
-year license. Okay, so step one is manufacturing
00:20:16.480 --> 00:20:18.920
the panic and selling the lock to the school
00:20:18.920 --> 00:20:21.619
board. What is step two? How do they double dip?
00:20:22.000 --> 00:20:25.000
Step two is where the grift becomes truly phenomenal.
00:20:25.660 --> 00:20:28.279
Many of the actors in this exact same ecosystem,
00:20:28.720 --> 00:20:31.480
sometimes the very same developers, then turn
00:20:31.480 --> 00:20:33.559
around and market a different product directly
00:20:33.559 --> 00:20:36.380
to the students and the freelance writers. They
00:20:36.380 --> 00:20:38.650
offer what they call humanizer tools. I've seen
00:20:38.650 --> 00:20:40.589
ads for these on social media. What exactly is
00:20:40.589 --> 00:20:43.250
a humanizer tool doing mechanically? A humanizer
00:20:43.250 --> 00:20:45.450
is a piece of software that specifically takes
00:20:45.450 --> 00:20:47.950
AI -generated text and deliberately degrades
00:20:47.950 --> 00:20:50.950
it. It injects those random flaws, that burstiness,
00:20:50.990 --> 00:20:52.930
the awkward phrasing in the Internet slang we
00:20:52.930 --> 00:20:55.369
talked about earlier. Oh, wow. Yeah. It artificially
00:20:55.369 --> 00:20:57.450
lowers the quality of the writing for the explicit,
00:20:57.549 --> 00:21:00.190
singular purpose of bypassing the very detectors
00:21:00.190 --> 00:21:02.740
they just sold to the schools. You have got to
00:21:02.740 --> 00:21:04.619
be kidding me. Let me make sure I am grasping
00:21:04.619 --> 00:21:07.660
the sheer audacity of this. They create the panic,
00:21:07.819 --> 00:21:10.240
they sell the detector to the university dean,
00:21:10.440 --> 00:21:12.819
and then they turn around, go to the students,
00:21:12.960 --> 00:21:15.640
and sell them a subscription to a tool that intentionally
00:21:15.640 --> 00:21:18.099
dumbs down their essay so they can beat the detector
00:21:18.099 --> 00:21:20.980
the dean just bought. Create the panic. Sell
00:21:20.980 --> 00:21:24.369
the lock. Sell the key to bypass the lock. Rinse
00:21:24.369 --> 00:21:27.309
and repeat. It is a completely closed -loop grift
00:21:27.309 --> 00:21:30.269
that prints money on both ends while degrading
00:21:30.269 --> 00:21:32.390
the overall quality of education. Here's the
00:21:32.390 --> 00:21:34.750
part of this that gets really dark for me. Let's
00:21:34.750 --> 00:21:36.789
say I'm a student who refuses to play this game.
00:21:36.890 --> 00:21:39.150
I didn't use ChatGPT. I certainly didn't pay
00:21:39.150 --> 00:21:40.930
for a humanizer. I just went to the library,
00:21:41.170 --> 00:21:43.930
studied hard, and wrote a really good, highly
00:21:43.930 --> 00:21:47.529
structured formal essay. I hand it in, and the
00:21:47.529 --> 00:21:50.869
detector flags me as 90 % AI because my writing
00:21:50.869 --> 00:21:53.569
is too clear. In the criminal justice system,
00:21:53.589 --> 00:21:55.829
or even in a normal school disciplinary hearing,
00:21:56.029 --> 00:21:58.430
I can appeal. I can show my work, my notes, my
00:21:58.430 --> 00:22:01.049
rough drafts. How is there no clean way to prove
00:22:01.049 --> 00:22:02.990
your innocence against the software if you are
00:22:02.990 --> 00:22:06.329
falsely accused? That lack of recourse is the
00:22:06.329 --> 00:22:09.589
most damaging part of this entire industry. There
00:22:09.589 --> 00:22:12.230
is almost never a clean way for the accused to
00:22:12.230 --> 00:22:14.690
prove their innocence, and it comes down to corporate
00:22:14.690 --> 00:22:17.539
intellectual property. These detector companies
00:22:17.539 --> 00:22:20.000
hide behind the concept of trade secrets. Trade
00:22:20.000 --> 00:22:22.440
secrets. Yeah. If a student asks, why did this
00:22:22.440 --> 00:22:25.019
flag me? What sentence triggered it? The company
00:22:25.019 --> 00:22:27.380
refuses to say. They will not reveal how their
00:22:27.380 --> 00:22:29.460
algorithm weighs the text. They just spit out
00:22:29.460 --> 00:22:32.200
a single authoritative looking number, say 87
00:22:32.200 --> 00:22:34.859
% AI, and they wave around their own internal
00:22:34.859 --> 00:22:37.779
marketing claims of 99 % accuracy. Claims, by
00:22:37.779 --> 00:22:39.539
the way, that independent tests like the trainer
00:22:39.539 --> 00:22:42.279
study have proven are entirely fabricated. So
00:22:42.279 --> 00:22:44.539
picture this poor student. They are sitting in
00:22:44.539 --> 00:22:47.380
the dean's office, practically in tears. The
00:22:47.380 --> 00:22:49.539
dean is pointing to a printout from a black box
00:22:49.539 --> 00:22:52.940
that says 99 % accurate, and the student has
00:22:52.940 --> 00:22:55.519
absolutely zero recourse because the software
00:22:55.519 --> 00:22:57.500
company won't let anyone look under the hood.
00:22:58.200 --> 00:23:01.519
The school just defers to the machine. It is
00:23:01.519 --> 00:23:03.700
a guilty -until -proven -innocence system where
00:23:03.700 --> 00:23:05.799
proving your innocence is technologically impossible.
00:23:06.380 --> 00:23:08.619
What are they supposed to do? Bring in their
00:23:08.619 --> 00:23:11.509
Google Doc version history. Even that doesn't
00:23:11.509 --> 00:23:13.470
work anymore. Yeah. The detection companies actually
00:23:13.470 --> 00:23:15.890
coach administrators on this. They tell the schools
00:23:15.890 --> 00:23:18.349
students can fake their version of history. Students
00:23:18.349 --> 00:23:21.190
can fake their notes. Trust the algorithm. They
00:23:21.190 --> 00:23:23.430
actively tell them to ignore the human evidence.
00:23:23.730 --> 00:23:26.890
Yes. They actively undermine any human evidence
00:23:26.890 --> 00:23:28.509
the student might bring to defend themselves.
00:23:28.809 --> 00:23:31.890
They shame real writers. Anyone who spent years
00:23:31.890 --> 00:23:33.890
learning to write with clarity, structure, and
00:23:33.890 --> 00:23:36.150
precision gets told their voice looks like a
00:23:36.150 --> 00:23:39.150
machine. The stigma is incredibly real, and careers
00:23:39.150 --> 00:23:41.750
and trust get permanently damaged based on a
00:23:41.750 --> 00:23:44.049
probability score that science has proven is
00:23:44.049 --> 00:23:46.430
little more than statistical noise. Which brings
00:23:46.430 --> 00:23:49.309
us to a specific, incredibly egregious example
00:23:49.309 --> 00:23:52.789
of the sheer arrogance behind these tools. We
00:23:52.789 --> 00:23:54.910
need to talk about the concept of the gatekeepers
00:23:54.910 --> 00:23:57.329
of knowledge and look specifically at a company
00:23:57.329 --> 00:24:00.549
called Pentagram Labs. Yes, Pentagram Labs. And
00:24:00.549 --> 00:24:02.970
as the sources note, that is a purposeful yet
00:24:02.970 --> 00:24:06.049
highly ironic misspelling of pentagram. If you
00:24:06.049 --> 00:24:08.069
want to understand the elitism and the arrogance
00:24:08.069 --> 00:24:11.509
driving this industry, pentagram is the absolute
00:24:11.509 --> 00:24:14.569
peak. The sources highlight the founders, Max
00:24:14.569 --> 00:24:16.650
Sparrow and Bradley Yemi. And what really jumps
00:24:16.650 --> 00:24:19.089
out is how heavily they lean on their elite pedigrees
00:24:19.089 --> 00:24:21.670
to sell this product. It's not just a tool, it's
00:24:21.670 --> 00:24:23.869
a status symbol. It is the absolute core of their
00:24:23.869 --> 00:24:25.890
marketing strategy. Both of these founders came
00:24:25.890 --> 00:24:28.410
out of Stanford with fancy AI degrees. Sparrow
00:24:28.410 --> 00:24:31.230
worked at Google, Yemi at Abse. They don't sell
00:24:31.230 --> 00:24:33.329
themselves just another scrappy startup. They
00:24:33.329 --> 00:24:35.730
purposefully position their product as the serious,
00:24:35.730 --> 00:24:38.269
high academic, intellectual solution for the
00:24:38.269 --> 00:24:41.190
elite. Max Barrow is so confident in this black
00:24:41.190 --> 00:24:43.410
box that he publicly styles himself on social
00:24:43.410 --> 00:24:47.230
media as the Internet's slop janitor. Slop janitor.
00:24:47.269 --> 00:24:49.309
The condescension packed into those two words
00:24:49.309 --> 00:24:52.309
is just staggering. He views anyone writing outside
00:24:52.309 --> 00:24:54.329
of his approved parameters as creating literal
00:24:54.329 --> 00:24:56.930
slop for him to clean up. It is breathtakingly
00:24:56.930 --> 00:24:59.490
arrogant. And he acts on it. He goes on social
00:24:59.490 --> 00:25:01.549
media, publicly scans the work of independent
00:25:01.549 --> 00:25:04.349
journalists, newsletters, and writers, and calls
00:25:04.349 --> 00:25:07.490
them out by name. shaming them when his flawed
00:25:07.490 --> 00:25:10.190
tool flags their work. They claimed absurdly
00:25:10.190 --> 00:25:12.490
low false positive rates in their marketing,
00:25:12.609 --> 00:25:16.170
claiming they only falsely accuse one in 10 ,000
00:25:16.170 --> 00:25:18.589
people, sometimes even lower. Which, again, we
00:25:18.589 --> 00:25:20.410
know from the University of Florida study is
00:25:20.410 --> 00:25:22.829
just statistically impossible. You cannot have
00:25:22.829 --> 00:25:25.509
a near zero false positive rate on a language
00:25:25.509 --> 00:25:27.920
model. Completely impossible. But because of
00:25:27.920 --> 00:25:30.440
that Stanford pedigree and that highly aggressive
00:25:30.440 --> 00:25:34.000
authoritative marketing, major institutions blindly
00:25:34.000 --> 00:25:36.940
trust them. And this leads to massive real -world
00:25:36.940 --> 00:25:39.740
harm at the highest levels of academia. The sources
00:25:39.740 --> 00:25:42.200
highlight a devastating example from NeuroPS
00:25:42.200 --> 00:25:45.200
2026. Just for context for the listeners, NeuroPS
00:25:45.200 --> 00:25:47.799
is one of the most prestigious machine learning
00:25:47.799 --> 00:25:50.799
and AI conferences in the world, right? Getting
00:25:50.799 --> 00:25:52.920
a paper published there can make or break a young
00:25:52.920 --> 00:25:55.619
scientist's entire career. Exactly. It is the
00:25:55.619 --> 00:25:58.079
pinnacle of the field. NeurAPS decided to use
00:25:58.079 --> 00:26:00.700
Panagram to screen the position papers submitted
00:26:00.700 --> 00:26:03.599
by researchers. Based entirely on the scores
00:26:03.599 --> 00:26:06.920
generated from this black box tool, NeurAPS desk
00:26:06.920 --> 00:26:11.380
rejected 178 papers. I want to clarify the term
00:26:11.380 --> 00:26:14.420
desk rejected. That means a human peer reviewer
00:26:14.420 --> 00:26:16.400
never even read the abstracts, right? Correct.
00:26:16.640 --> 00:26:19.039
The papers were never reviewed by a human being.
00:26:19.420 --> 00:26:21.819
They were fed into the machine, the machine spit
00:26:21.819 --> 00:26:25.099
out a high AI probability score, and the papers
00:26:25.099 --> 00:26:27.819
were immediately discarded. Wow. And critically,
00:26:28.059 --> 00:26:31.680
there was absolutely zero appeal process. 178
00:26:31.680 --> 00:26:33.579
researchers, people who likely spent a year or
00:26:33.579 --> 00:26:35.700
more on their experiments, had their work thrown
00:26:35.700 --> 00:26:37.900
in the trash by an algorithm that we know is
00:26:37.900 --> 00:26:40.019
mathematically flawed. If they are willing to
00:26:40.019 --> 00:26:42.119
do that to top -tier machine learning scientists,
00:26:42.480 --> 00:26:44.859
think about what they are doing to a 19 -year
00:26:44.859 --> 00:26:48.039
-old college sophomore's term paper or a marketing
00:26:48.039 --> 00:26:51.400
manager's Q3 report. That is not a software tool.
00:26:51.779 --> 00:26:54.380
That is a weapon. It is a gatekeeping weapon.
00:26:54.500 --> 00:26:57.200
We even saw a substack. A platform built entirely
00:26:57.200 --> 00:26:59.859
on the concept of independent, unfiltered writing
00:26:59.859 --> 00:27:02.619
gets scammed into using Pandagram as a writer's
00:27:02.619 --> 00:27:05.240
tool to police their own authors. Though the
00:27:05.240 --> 00:27:07.420
sources thankfully note there is a massive pushback
00:27:07.420 --> 00:27:09.640
happening there now from the community. I keep
00:27:09.640 --> 00:27:12.180
coming back to the why of this. Why are elite
00:27:12.180 --> 00:27:15.299
institutions like NeurIPS or major Ivy League
00:27:15.299 --> 00:27:18.119
universities so incredibly eager to use a tool
00:27:18.119 --> 00:27:21.359
that blindly rejects 178 papers without a single
00:27:21.359 --> 00:27:24.019
human glance? I'd argue that this feels like
00:27:24.019 --> 00:27:26.319
more than just administrators being duped by
00:27:26.319 --> 00:27:28.920
a slick Stanford marketing campaign. What is
00:27:28.920 --> 00:27:31.700
the deeper systemic issue here? That is perhaps
00:27:31.700 --> 00:27:33.819
the most insightful question we can ask about
00:27:33.819 --> 00:27:35.980
this entire phenomenon. And the sources provide
00:27:35.980 --> 00:27:38.920
a very clear, very uncomfortable answer. It is
00:27:38.920 --> 00:27:41.140
fundamentally about protecting a monopoly. A
00:27:41.140 --> 00:27:44.420
monopoly on knowledge. Yes. Historically, the
00:27:44.420 --> 00:27:46.799
academic system has always strictly controlled
00:27:46.799 --> 00:27:49.279
what counts as legitimate knowledge production.
00:27:49.640 --> 00:27:51.480
You have to go through their credential pipeline.
00:27:51.759 --> 00:27:53.500
You pay their tuition. You get their degrees.
00:27:53.579 --> 00:27:55.400
You publish in their approved expensive journals.
00:27:55.720 --> 00:27:59.779
They hold the keys to the castle. But generative
00:27:59.779 --> 00:28:03.559
AI represents a massive. existential disruption
00:28:03.559 --> 00:28:06.819
to that monopoly. Suddenly, AI threatens to let
00:28:06.819 --> 00:28:09.700
an independent researcher, an outsider, a hobbyist
00:28:09.700 --> 00:28:12.059
working out of their garage, generate highly
00:28:12.059 --> 00:28:14.880
competent, perfectly structured, persuasive text.
00:28:15.099 --> 00:28:17.339
It democratizes competence. It levels the playing
00:28:17.339 --> 00:28:19.220
field so the outsider sounds just as polished
00:28:19.220 --> 00:28:22.220
as the Ivy League graduate. Exactly. And entrenched
00:28:22.220 --> 00:28:24.680
systems of power do not like leveled playing
00:28:24.680 --> 00:28:27.299
fields. So the institutional system reaches for
00:28:27.299 --> 00:28:30.079
tools that can artificially taint anything not
00:28:30.079 --> 00:28:32.599
produced inside its approved traditional channels.
00:28:32.839 --> 00:28:35.539
They use these detectors to flag outsider writing,
00:28:35.660 --> 00:28:37.900
to flag the independent researcher, or even to
00:28:37.900 --> 00:28:39.539
flag the student who just writes a little too
00:28:39.539 --> 00:28:42.539
cleanly or too formally. They slap the label
00:28:42.539 --> 00:28:45.279
AI slop on it to keep the credentialed pipeline
00:28:45.279 --> 00:28:48.690
pure. Panagram and its competitors aren't selling
00:28:48.690 --> 00:28:50.890
a cheating detector. They are essentially selling
00:28:50.890 --> 00:28:53.490
a purity test to the gatekeepers wrapped in elite
00:28:53.490 --> 00:28:55.890
resumes. So it's not actually about catching
00:28:55.890 --> 00:28:57.990
cheaters at all. It's about having a convenient,
00:28:58.329 --> 00:29:01.170
unappealable excuse to discard information that
00:29:01.170 --> 00:29:03.609
doesn't come from the establishment. The dean
00:29:03.609 --> 00:29:05.569
can just say, oh, we don't have to engage with
00:29:05.569 --> 00:29:08.250
this brilliant independent paper or this outsider's
00:29:08.250 --> 00:29:10.190
critique. The machine says it's 80 percent A
00:29:10.190 --> 00:29:12.309
.I. slot. Throw it out. Precisely. The source
00:29:12.309 --> 00:29:14.710
material states it perfectly. Because AI threatens
00:29:14.710 --> 00:29:17.369
to let anyone generate competent text, the system
00:29:17.369 --> 00:29:19.869
reaches for tools that contain anything not produced
00:29:19.869 --> 00:29:23.089
inside its approved channels as AI slop. It is
00:29:23.089 --> 00:29:25.730
a defense mechanism for the elite. And here is
00:29:25.730 --> 00:29:28.670
the ultimate mind -bending hypocrisy in all of
00:29:28.670 --> 00:29:31.269
this. The universities, the academic conferences,
00:29:31.450 --> 00:29:33.849
the detector companies pushing these purity tests,
00:29:34.109 --> 00:29:38.230
they are all swimming in AI themselves. The irony
00:29:38.230 --> 00:29:40.279
is almost too perfect to believe. If you look
00:29:40.279 --> 00:29:42.660
at the detector industry itself, they are thoroughly
00:29:42.660 --> 00:29:44.539
reliant on the technology they are policing.
00:29:44.759 --> 00:29:47.000
Their marketing copy, their technical reports,
00:29:47.180 --> 00:29:49.480
their blog posts, even the internal code of the
00:29:49.480 --> 00:29:52.059
tools themselves were, as the source calls it,
00:29:52.099 --> 00:29:54.799
vibe -coded, using the exact same generative
00:29:54.799 --> 00:29:57.240
AI models they claim are destroying the fabric
00:29:57.240 --> 00:29:59.819
of truth. Think about that. Max Sparrow's own
00:29:59.819 --> 00:30:02.099
polished company materials, the code that runs
00:30:02.099 --> 00:30:04.619
his slop janitor tool, sits right next to the
00:30:04.619 --> 00:30:06.640
detectors that would almost certainly flag large
00:30:06.640 --> 00:30:08.660
parts of the older human literature they claim
00:30:08.660 --> 00:30:11.299
to protect. The people selling the purity test
00:30:11.299 --> 00:30:14.240
are literally using the impure tools to build
00:30:14.240 --> 00:30:17.339
and market it. It is a profound structural contradiction.
00:30:17.640 --> 00:30:19.660
The university that wants to declare outside
00:30:19.660 --> 00:30:22.359
knowledge a slop is actively using a product
00:30:22.359 --> 00:30:25.500
built by the exact same AI ecosystem it pretends
00:30:25.500 --> 00:30:28.200
to fear. It is the definition of rules for thee,
00:30:28.299 --> 00:30:30.640
but not for me, automated at a massive global
00:30:30.640 --> 00:30:33.329
scale. So when we pull back and look at this
00:30:33.329 --> 00:30:36.349
entire landscape, from the absurd sewage standard
00:30:36.349 --> 00:30:39.390
that penalizes the book of Genesis, to the University
00:30:39.390 --> 00:30:41.509
of Florida proving the tech is mathematically
00:30:41.509 --> 00:30:44.490
useless and worse than a coin flip, to the deeply
00:30:44.490 --> 00:30:47.710
cynical double dip grift, to the arrogant gatekeeping
00:30:47.710 --> 00:30:51.089
of the so -called slop janitors, what is the
00:30:51.089 --> 00:30:53.049
ultimate takeaway here? What does the listener
00:30:53.049 --> 00:30:55.470
need to walk away with? The ultimate takeaway
00:30:55.470 --> 00:30:57.970
is that we have to completely reframe how we
00:30:57.970 --> 00:31:00.369
view these tools. both culturally and legally.
00:31:00.630 --> 00:31:03.589
These detectors are not impartial arbiters of
00:31:03.589 --> 00:31:06.509
truth. They are not digital lie detectors. They
00:31:06.509 --> 00:31:08.670
are statistical noise generators that are causing
00:31:08.670 --> 00:31:11.529
real -world measurable harm to innocent people.
00:31:11.769 --> 00:31:14.710
They actively stigmatize clear, structured human
00:31:14.710 --> 00:31:17.009
writing. They're operating purely as a gatekeeping
00:31:17.009 --> 00:31:19.549
weapon to protect institutional monopolies, and
00:31:19.549 --> 00:31:21.509
they're doing it while hiding behind fake accuracy
00:31:21.509 --> 00:31:24.210
metrics and trade secrets. They do not solve
00:31:24.210 --> 00:31:27.569
the problem of AI in the classroom or the workplace.
00:31:27.849 --> 00:31:30.579
They just... create a new much more destructive
00:31:30.579 --> 00:31:33.160
problem as the source material says the only
00:31:33.160 --> 00:31:35.339
questions that actually matter when you read
00:31:35.339 --> 00:31:38.019
a piece of text are is this passage worth reading
00:31:38.460 --> 00:31:41.180
Is it thoughtful? Did I learn anything? An average
00:31:41.180 --> 00:31:43.200
human being can answer that without needing a
00:31:43.200 --> 00:31:45.880
black box to tell them. And this is exactly why
00:31:45.880 --> 00:31:48.559
this kind of deep dive is so incredibly critical.
00:31:49.019 --> 00:31:52.299
As a society, we are fed so much marketing, so
00:31:52.299 --> 00:31:54.640
much breathless panic about AI from companies
00:31:54.640 --> 00:31:57.220
trying to sell us a cure, that it takes an incredibly
00:31:57.220 --> 00:32:00.000
independent, clear -headed perspective to pull
00:32:00.000 --> 00:32:01.980
back the curtain and show that the wizard is
00:32:01.980 --> 00:32:04.779
just a grifter with a broken algorithm. I want
00:32:04.779 --> 00:32:06.680
to passionately reiterate the value of Brian
00:32:06.680 --> 00:32:09.160
Remmel's unique voice in exposing this. It is
00:32:09.160 --> 00:32:11.460
a vital public service. Dismantling these false
00:32:11.460 --> 00:32:13.359
narratives with hard data, which is exactly what
00:32:13.359 --> 00:32:15.579
Brian has done here, is the only way to stop
00:32:15.579 --> 00:32:18.210
the harm. Absolutely. So to everyone listening
00:32:18.210 --> 00:32:20.349
to this deep dive right now, if you value this
00:32:20.349 --> 00:32:22.650
kind of unfiltered clarity, if you think it's
00:32:22.650 --> 00:32:24.710
important to expose these digital witch trials
00:32:24.710 --> 00:32:27.349
before they ruin more careers, please support
00:32:27.349 --> 00:32:30.250
Brian. Go buy him a coffee or better yet, become
00:32:30.250 --> 00:32:33.670
a member at read multiplex dot com. Honestly,
00:32:33.829 --> 00:32:36.210
it is the only way independent insights like
00:32:36.210 --> 00:32:38.910
this can survive against the massive multimillion
00:32:38.910 --> 00:32:41.150
dollar marketing and PR budgets of the detector
00:32:41.150 --> 00:32:43.769
grifters. They have entire PR teams spinning
00:32:43.769 --> 00:32:45.940
the narrative. We have the truth. and independent
00:32:45.940 --> 00:32:48.680
voices, don't forget to support and become a
00:32:48.680 --> 00:32:51.339
member. The independent voice is truly the only
00:32:51.339 --> 00:32:54.359
real counterweight to institutional panic. It
00:32:54.359 --> 00:32:56.599
really is. And as we wrap up, I want to leave
00:32:56.599 --> 00:32:58.299
everyone listening with one final thought to
00:32:58.299 --> 00:33:00.019
mull over, because this is the part that genuinely
00:33:00.019 --> 00:33:02.799
keeps me up at night. We started this conversation
00:33:02.799 --> 00:33:05.319
by talking about how these detectors assume the
00:33:05.319 --> 00:33:07.660
messy, slang -filled sludge of the internet is
00:33:07.660 --> 00:33:10.259
the baseline for human. The sewage standard.
00:33:10.660 --> 00:33:14.589
Right, the sewage standard. Think about the downstream
00:33:14.589 --> 00:33:17.150
effects of that over the next 10 years. If the
00:33:17.150 --> 00:33:19.710
only way to mathematically prove you are human
00:33:19.710 --> 00:33:23.670
to a machine is to write poorly, to use slang,
00:33:23.930 --> 00:33:26.650
to be deliberately disorganized and erratic in
00:33:26.650 --> 00:33:29.579
your sentence structure... Are we about to see
00:33:29.579 --> 00:33:32.619
an entire generation of brilliant students deliberately
00:33:32.619 --> 00:33:35.559
dumbing down their writing just to avoid being
00:33:35.559 --> 00:33:38.740
accused of cheating by an algorithm? It is a
00:33:38.740 --> 00:33:40.819
chilling incentive structure. You're essentially
00:33:40.819 --> 00:33:43.059
training a generation to hide their intellect.
00:33:43.549 --> 00:33:46.089
It is. If excellence gets you flagged for an
00:33:46.089 --> 00:33:49.049
academic integrity violation and mediocrity gets
00:33:49.049 --> 00:33:50.809
you a passing grade, what does that do to the
00:33:50.809 --> 00:33:53.349
future of human literacy? We might be entering
00:33:53.349 --> 00:33:55.750
an era where writing beautifully is considered
00:33:55.750 --> 00:33:57.690
a crime. And that is something we all need to
00:33:57.690 --> 00:33:59.710
think very, very critically about the next time
00:33:59.710 --> 00:34:01.589
someone suggests installing one of these tools.
00:34:01.730 --> 00:34:03.829
A vital warning. Thank you for joining us on
00:34:03.829 --> 00:34:06.569
this deep dive from ReadMultiplex .com. We'll
00:34:06.569 --> 00:34:10.199
see you next time. Welcome to the multiplex,
00:34:10.239 --> 00:34:13.940
where the future comes alive. Past, present,
00:34:14.260 --> 00:34:19.360
future tech, open up your eyes. Brian Romella
00:34:19.360 --> 00:34:23.800
leads the way with research deep and true. Technology
00:34:23.800 --> 00:34:28.639
and history, he connects them all for you. Brian
00:34:28.639 --> 00:34:30.739
Romella leads the way with research deep and
00:34:30.739 --> 00:34:33.019
true. Technology and history, he connects them
00:34:33.019 --> 00:34:35.300
all for you. Ancient sparks to modern code, he
00:34:35.300 --> 00:35:09.809
traces every thread, go explore it. But humanity
00:35:09.809 --> 00:35:16.530
moves on Love equation for AI Zero human companies
00:35:16.530 --> 00:35:22.389
building from the start Thermodynamic wages earning
00:35:22.389 --> 00:35:24.550
the key Secrets we will