Thd AI Table: Hi everyone, welcome to the AI table. Today ⁓ we have an amazing guest with us, Jessica Talisman. Jessica, you wanna do a quick intro of yourself?
Jessica Talisman: Sure. ⁓ so as Dia said, I'm Jessica Talisman. I've been working in around data, more around semantics, ⁓ ⁓ geez, 1997. That's I really started. I have at places like Amazon, Adobe, Overstock, which is now Bat Bed Bath and Beyond, and Department of Justice. right now I work consulting and I also co-run ⁓ the Knowledge Graph Academy with Tony Seal and Katarina Kari.
Thd AI Table: Yes, thanks for plugging that. I've been watching the work that you're doing with the Academy. Very exciting stuff. ⁓ we'll leave a link to that on this podcast. all so let's go ahead and dig in. My first question. something we've been seeing a bit in the news lately is how AI systems ⁓ have been to exhibit bias sometimes, ⁓ and important decision making, ⁓ things like whether a job candidate passes an AI screen. and goes on to have their resume viewed by a human. Can you talk in layman's terms for us about why these biases happen and if you have any recommendations for businesses to get around them?
Jessica Talisman: Sure. I mean one of the problems especially with AI is that it's a black box. So it comes already with training data. And if you consider the world wide web or the internet, that's bulk of the training data. So there's inherited bias. ⁓ there's a great Yeah, yeah. ⁓
Thd AI Table: Can we talk a little bit sorry to interrupt you, can we can we back up to to defining a black box for s for some folks who might not know what that is?
Jessica Talisman: Sure. Yeah, so so there's two basic components, and I'm I'm really distilling this, but basic components of of an AI model, ⁓ large language model. ⁓ one is the parametric layer, and that's the layer that none of us see. That truly is the black box. It's where the training data sits. And so we don't have any insight or idea of how these models have been trained. Now, what happens is that training data is broken up into these little chunks that become tokens, and then it's a statistical prediction model. So it's going to statistically predict ⁓ the next word is in a sentence, for example. So if a lot of the training data contains ⁓ information or or knowledge that is biased, large language model is going to inherit those biases, right? Because that's how its brain You can think of the parametric layer as the brain. That's how its brain has been trained to think in those lines. And then there's the retrieval layer, which is what we all can see or at least have some measure of control around. Not not total control, but that's where we enter our own context or what's being called context. And that's where we can give our own training data or try to steer AI towards ⁓ answers and opinions and and I guess that. if if you think about it this way, it's very hard for any of us to be unbiased. So ⁓ even organizations can insert their own bias if their own data and information within their own organization is biased towards certain opinions.
Thd AI Table: Yes, so I think ⁓ you're you're illuminating an important point that to a large extent AI is a reflection of ourselves. ⁓ in sense as humans, you know, we're not perfect, so AI is is is not gonna be perfect either. ⁓ but there are strategies that we can use to try to get it less biased.
Jessica Talisman: Total reflection. Less biased. And I mean this has been a struggle for a long time. I I, you know, I have a master's in information in library science and I am a librarian. And a part of the job of a librarian is to try to remove as much bias as possible. you're trying to stay objective. ⁓ I do so in my own writing, is to balance the points, so a point and a counterpoint. ⁓ somewhere that we've seen bias show up ⁓ was Amazon. ⁓ it's resume review agent. I guess you could call it an agent, but it was an AI implementation to help to screen resumes. And so what happened with that algorithm, and it was actually discovered and it's been written about and published, there was a documentary about it, was that it was specifically not ⁓ letting resumes pass through it had a bias against women. ⁓ anyone who went to a women's college when that algorithm was being used up to a point, were not passed through the filter. So that's an example of where the bias shows up.
Thd AI Table: Yeah, so so yeah, these can have some serious implications. walk through from your perspective strategies for businesses to that, that includes adding that context. What does that process look like?
Jessica Talisman: Yeah, so I I actually have been thinking about this a lot because something I run into quite a bit is businesses will try to scope and scale the context that they give only relative to their own world and their own language. That's a big problem because inevitably, especially if a business is interacting with the actual world. then ⁓ can easily be introduced because you create a very narrow sliver into that sort of idea of the and ⁓ that that your ⁓ data information holds, right? So an example is ⁓ you know, when I worked at Adobe, for example, a feature almost every single product was called journeys. And so that internal definition of what a journey was didn't really match the external definition of what that was. So that's one place that you see people narrowing their sort of scope and field and only ⁓ I should say, assuming that AI understands your internal world. So I think of it like touching grass. There's sort of a balance in which you have to relate how internal definitions align with external definitions, because guess what? your your customers are coming from the external world from the actual world they're not coming from internal so that's one thing the other thing is including context more than what your exact business holds you want to ground your business concepts for example in context in the real world, which maybe mean that you have to expand your small world, your myopic world to include more than what you cover and more than what what's meaningful to you in your business. And and that seems really confusing, but I've I've run into this quite a bit with businesses where they're like, we just wanna wanna wanna narrow down and edit down to only the things that matter inside of our business. ⁓ But you don't put those terms that that meaning In context within the outside world, larger than what your business is, then you're going to lose a huge lens into opportunities, also contextual understanding of your business. Because you have to understand how your business is placed within the larger realm. ⁓
Thd AI Table: Yeah, that that is a a huge point. my next question, a big part of of what we refer to in our discipline as responsible AI, ⁓ which the practice of ⁓ designing deploying AI systems that are safe and fair, ⁓ is making sure AI systems are transparent and easily traceable with respect to the decisions that They're making can you talk about how sciences can help with this?
Jessica Talisman: Sure. I just wrote a piece called an Ontology that evaluated the data landscape right now, and I I brought an information science lens to it. ⁓ of ⁓ problems is because we have that black box effect where we can't see into how certain decisions are made, ⁓ leaves ⁓ a lot of organizations and people kind of handicapped in not being able to really transparently into these models. ⁓ So the thing that's starting to emerge is the idea of data sovereignty. And sovereignty is the idea that you can own your own data, regardless of a vendor platform. Every organization, every ⁓ institution should be able to own its own data very transparently and ex and and that in and of itself is one of the keys to being able to transparently explain how LLM arrived at certain answers. We're this come up ⁓ the EU AI Act, once it goes into actual force, or the data, the data act, the US government's data act. ⁓ there are requirements for data sovereignty. And those requirements are going to help Hopefully, for to actually implement explainable AI. And explainable AI is that you can actually point to and detail how an answer was generated and what data was used to generate that answer. So I'll give you an idea. If you have ⁓ ⁓ based say that you have a database and the database that vendor offers some sort of agentic AI within the product. of the issues is who owns responsibility for explaining how those answers ⁓ were Right. In a lot of those cases, the the user or the customer doesn't even have visibility into that data. So one of the base requirements is understanding and having ownership, really well-organized, curated ownership of your data, and being able to Cite sources. So it's just like if if you're writing something or you're quoting something, you know, best practices is to cite sources. It's not, you know, otherwise it's plagiarism. ⁓ then we get into a world of ⁓ ⁓ starting to focus on and prioritize things like provenance, which is a very important concept within both history and library science and information science. So is when you can actually attach attribution to your data object. objects and your information and that things are clearly organized in a way that you can easily through an evaluation process point back to the source identify where an answer came from.
Thd AI Table: ⁓ within orgs from a leadership perspective, what roles should have purview over this, do you think?
Jessica Talisman: That's interesting. That's a really interesting question because you have sort of a musical chair situation with CDOs. ⁓ They don't really have a long tenure. So the natural answer would be governance, you know, whoever's head of governance and CDOs, at least for that aspect. But time has passed, what I'm realizing more and more, especially through my own work, is it really is a collaborative cross team effort.
Thd AI Table: Mm-hmm.
Jessica Talisman: It it is something that needs to be owned by everyone. It needs to be a muscle that's built within an organization. And there needs to be sort of organizational policies in place so that there's repeatable patterns. of the problems is most organizations, not all, have ⁓ sort of vectors of control in terms of each person's ⁓ implementations of AI. So implementations and deployments are wildly different and everyone's using different data. In fact, it it is tending to create deeper silos rather than breaking down silos, which points to the fact that, like, okay, if if we don't all get on the same page have ⁓ ⁓ strategy around this, ⁓ then then there can be no leadership for lack of strategy.
Thd AI Table: I agree. Strategy first, everyone. Strategy first. Especially with these lawsuits coming in, strategy first.
Jessica Talisman: Strategy first. It's getting so lost. Strategy first. And like the stakes are high. Yeah.
Thd AI Table: my next question, people interact with ⁓ LLMs Claude, Gemini, ChatGPT, we're all familiar with hallucinations or when an AI system provides you some output that isn't true or doesn't ⁓ make any can you talk to us about what makes this happen? And is there anything that folks can do to reduce ⁓ hallucinations?
Jessica Talisman: Yeah, and so for hallucinations, you know, I'm a big proponent and supporter of know, knowledge graphs of symbolic AI. So a lot of people don't realize that yes, we are all engaging with what we know of as AI, but the AI that we're playing with is generative AI. It's large language models, which is a st it's a statistical prediction model. So again, it's it's just statistically predicting. Another to that ⁓ is there's other types of AI. And what's emerging from the research is symbolic AI, it counterbalances hallucinations. what is symbolic AI? It's things like taxonomies and ontologies and knowledge graphs. So those are becoming increasingly ⁓ to organizations. So means curating and organizing your data. It means describing your data in a machine-readable format. They become symbols so that what happens is the statistical prediction model ⁓ is to have Reference points to actual facts within a system, because you can say something like, you know, has this podcast that's like an actual fact, and I can verify it because it's represented, for example, my taxonomy or my knowledge graph or you know, within my ontology. So one vector of control for hallucinations is to actually build fact-based symbolic models, you know, things that are easily consumable by machines. They're machine readable. ⁓ And they lead to easier interpretation ⁓ responses. Now, again, we have control over how the models arrive on our doorsteps ⁓ or our you know computers. There, there's that black box effect. So ⁓ of issues is when an answer cannot be easily explained, that's that points needing to build out that sort of fact-based system, those reference points, and it becomes a point of stability. Now the one thing is it's not it's not a one-off, right? Building these things are not one-off. It's not like spin up and generate an ontology for for each person's little sliver of AI implementation. There needs to be sort of consistency. But the other thing for hallucinations is that you know symbolic AI systems, so that means you know your taxonomy, your ontology, that it becomes your organization's memory as well. And that includes provenance. So you're basically building your organizational's brain. so to minimize hallucinations, you need those reference points. To the earlier question, you need to have explainable AI, you need evaluations. and start ⁓ to a record or a log. of these sort of, you know, if we're talking about performance metrics, we need to start measuring the performance of provenance, truthfulness, reliability, those sorts of things. And a lot of organizations are not doing that. They're just accepting the hallucination as fact.
Thd AI Table: Yes, that's what I'm seeing as well. And you talk about this work, is this something that you see falling under CDOs as well? Or is more flexible? Are there other places it could it make sense for it to be within an organization? I know not all organizations have a CDO.
Jessica Talisman: Yeah, I mean if there's no CDO, especially, I mean, ideally a CDO is leading, but again, so many organizations do not have CDOs. ⁓ so I I again think it needs to be something that's like a policy governance. You know, I was thinking about it you you s organizations, especially because of things like GDPR, which was, you know, the legislation to to ⁓ you know regarding privacy that came out of the EU. And one of the important things is like how did organizations adapt and adopt that sort of policy of protecting people's private information by law? And so that had to be instituted as part of governance and classification, right? So having a system in place that could reliably, you know, and consistently protect throughout an organization. And that same thing has to happen. So ⁓ really is foundational. ⁓ I I call it knowledge infrastructure. It's something we don't have. Most organizations do not have today. ⁓ organizations do not have people trained up ⁓ ⁓ in information science. in semantics ⁓ in knowledge, management. ⁓ and ⁓ my is. Is that organizations are going to start, you know, we hear of the great upskilling or, you know, those skill sets need to be in place, such as knowledge management and semantics and information, all of those things. ⁓ right now organizations limping along and people are kind of trying to do things on the fly ⁓ without having ⁓ that themselves of how to do that.
Thd AI Table: Yeah, to me I I see that as well. I think it's a matter of data there are a few ⁓ shining stars are doing this work well and I I think that we're gonna see them break out from pack. because I think there is serious ROI that's connected to being ⁓ a mature, you know, some folks calling an intelligent enterprise. So I think ⁓ When organizations are moving towards that territory where they want to become an in an intelligent enterprise, where they understand AI not simply as a tool, but as something that has some ⁓ ramifications for how business operates and how it structures itself to operate. ⁓ when you when you dig into changing that this is a part of the structural change that will need to occur for for businesses who want to, you know, be at the top of the pack.
Jessica Talisman: And if you look at the training data that's most consumed and used by AI, because there was a great Washington Post article last year that evaluated, for example, Llama, which is Facebook's open model, and 15 million training sources and was able to evaluate because that's an open model. So they were able to evaluate what what constituted the training data. ⁓ And top 10 sources came from pretty well structured, good knowledge sources like New York Times and Wikidata. Wikidata ⁓ is And Wikipedia, it's a knowledge graph. DBpedia was another big source. It's a knowledge graph. So that's the other part of it is like that knowledge management piece that steps beyond data and goes towards, you know, we we see this happening with content and documents and and you know, people's share points that they didn't really care to to organize before. Those sorts of things start to become really important.
Thd AI Table: Yeah. we're seeing more organizations ⁓ that are training their own models. Can you about why it's dangerous to let an AI model train or ingest freely internal corporate or personal communications and ⁓ what boundary might look like when it comes to that?
Jessica Talisman: Yeah, so ⁓ I mean I think that's gonna be the new wave ⁓ w once everything ⁓ it's gonna have to settle down in some however things shake out, but these the models are becoming the open models are becoming our frontier models are becoming increasingly expensive. So naturally people are turning into, you know, training their own small to medium-sized models. ⁓ and just the same problems that occur with the black box effect of these frontier models, ⁓ you you want to be careful and curious. First of all, what you're giving the model, but also you have customer data, you have financial data. There are things that are proprietary that don't belong in training data that could be very, very dangerous. so security is starting to become an increasing priority across it. You can look at things like open claw, we're seeing that mirror effect of what's happening with models that are open to the public, you know, or available to the public. Yeah. The same rules apply with small and medium-sized models as well. So people's own models that they're training internally. So there need to be guardrails. that's ⁓ really important. I mean, that takes a lot of curation, a lot of organization. security to be looped in from the very beginning because there are vulnerabilities in these models. You open up doors that otherwise wouldn't be available. there are things like does your model reach reach out to the internet? is there any sort of interoperability or or dependency on on the internet for answers or for, you know, as soon as you open up that door, ⁓ expose a vulnerability again. ⁓ who has access, who gets to include information or knowledge? Is there are there point people who are determining, you know, and curating and organizing the information available to the small to medium size models? But the security ⁓ threats that are emerging are are pretty serious. And so there there have already been hacks that that that have been pretty scary. So I think organizations are starting to feel the effects.
Thd AI Table: Yeah, and I think it highlights the point you made earlier about the importance of collaboration when it comes to this work, making sure you have a security person that's in governance, data person participating in governance, someone from legal ⁓
Jessica Talisman: You know what's interesting, like w Having been at some of these organizations, one thing that went out the door that used to be a priority prior to 2021 was vendor evaluations. Really rigorous vendor evaluations of software tools platforms. If I work for an organization and I want to onboard, for example, or I want to purchase or acquire, you know, in any of those effects, another tool, right, for my organization. I'm I I need a knowledge graph tool for my organization. I go through rigorous vendor analysis that's set out by the organization. You can't pass go, I have to present at a council to show them the tool I want. Security goes through a big evaluation. is that not happening with AI models?
Thd AI Table: I think it's because everyone's in a rush. ⁓ very interesting to watch you're right, the priority that's been placed ⁓ on evaluating solutions, it has significantly over the years. And I would argue, in this moment it it makes a more sense for it to be increasing given the ramifications of of adopting something that is faulty. We still haven't figured that out in the legal system. I mean there are lots of lawsuits are going through the court right now but we haven't determined who's liable things go wrong. Is it is it the tool provider, is it the company that contracted with the tool who who's ⁓ responsible here. ⁓ ⁓
Jessica Talisman: Right. So
Thd AI Table: While that's up in the air, it makes sense for everybody to cover themselves. so
Jessica Talisman: There's a concept in in in libraries that's in archives, which is really interesting, which is ⁓ in the American College Research Libraries Association, they they they created sort of tenets of AI. But but one thing that's pretty preval prevalent is no human, no AI. And I really, really like that because if we don't have like humans in the loop and evaluating ⁓ before we get to sort of that emergency or or legal then course ⁓ when happens, it's too late. and ⁓ you so you don't have those things in front of, but ⁓ there are things in terms of copyright, right? I mean, there's there's other issues that extend further. and those just policies haven't been updated. We had the d the digital. Millennial Copyright Act. And that ⁓ was created when internet came around because it a format or a medium that hadn't existed before. So it was to protect, you know, people's work, but also extend a little bit further because they needed, you know, was this new medium. We haven't addressed this new medium, this new format. No.
Thd AI Table: No. We're we're still having conversations. We're still figuring it out. it's it's up in the air. ⁓ sometimes I think companies forget that part that it's up in the air. Like ⁓ doing all this stuff, but like the legal framework behind it is not solid. It it just because we have some people can do something one day doesn't mean that we can do it the next. so it'll be interesting to see how everything shakes out. We're we're kinda all along for the ride.
Jessica Talisman: So up here. It's not there.
Thd AI Table: But I do think that it's important that people like yourself in and and share insights their professional perspective ⁓ of we need to consider as we think about these new policies that'll need to come into place.
Jessica Talisman: I mean I I I do have like on respect, there are some ⁓ organizations, one that I'm working with, that are very highly regulated. And they are, you know, some of these organizations are being very thoughtful, careful, and mitigated. And and I think those are the ones that are gonna emerge. And they may be working more slowly, you know, good things take time.
Thd AI Table: Yeah, they're more thoughtful. That's my that's my experience as well. And I just want to give a shout out to the banks because ⁓ a lot of the banks a lot of the bank ⁓ executives are really thinking deeply about this. And they're not just out and adopting tools and using a checklist and say, hey, AI, we're doing it, we're doing it all. they're really thinking about how to it in ways to best serve their customers.
Jessica Talisman: Yes. Yeah.
Thd AI Table: ⁓ Okay, so my next question, how does a company maintain control over truth language and product names and industry regulations are changing frequently?
Jessica Talisman: gosh, that goes back to the knowledge management muscle. It's like if you don't have people that are managing, it's like a chain. I was thinking about this in terms of change management for ontology, taxonomy, knowledge management work. So working on a project right now at an unnamed company. ⁓ And is the most interesting thing is how it's education, education first and foremost. Most people do not understand. There's there's think that it's a one and done. I build this knowledge structure, this taxonomy, this ontology, this thing that's gonna help to reduce or eliminate hallucinations ⁓ and it's done. And just not the case. It's like, would you do that with your database? As fast as the data in your database changes is as fast and as as careful as you have to be about managing the knowledge that you give to your mom.
Thd AI Table: Mm-hmm. ⁓
Jessica Talisman: It is lockstep. so I'm, ⁓ know, I still see people saying, okay, well, can I just have AI generate this for me, the knowledge structures, you know, the taxonomies, the ontologies for me, and call it done. Not really. Like, no, you can't because if your data is continuing to change, then you're basically, of course it's gonna go stale. That's the secondary complaint. It's very interesting because I'm not sure if people register it. The first complaint is takes a lot of time to organize and to build and to curate. It's like okay. Yeah. ⁓ then they say, okay, ⁓ can AI do it for me? Is usually the second question. It's like if you want AI to organize your business, sure.
Thd AI Table: Yeah, that's a hard.
Jessica Talisman: But you're gonna get that quality of of understanding, which is, you know, introducing hallucinations because you haven't taken the time to curate and to determine what is meaningful for your business. then the third step is: okay, so I built it once ⁓ and it's going great. And you're like, okay, thumbs up, that's that's wonderful. It's going great. They they took the time to build it. And then you hear people complaining, it's stale. Ontologies are stale, taxonomies are stale. And it's like, yeah, because you didn't maintain it. You don't have people that are constantly curating and updating it at the rate of the business. Of course it's stale. So ⁓ something that's occurred to me is like a lot of this involves abstract logic and reasoning. A lot of this includes areas that we haven't cared about. And then there's the the adopted and adapted principle of build fast and break things, which we cannot apply to AI. is not the type of solution or implementation that we can build fast and break things. Like we're moving back to that data maturity. This is something that actually good things take time to build it needs to be a constant iterative process.
Thd AI Table: and I mean I would argue that we can build fast and break things, but we probably shouldn't. ⁓ I really want folks to like have a vision of of what that means when we say that phrase of what are the things we are breaking, ⁓ or that we can potentially break it comes to AI because those are serious things and the biggest thing that we can break is we can break trust.
Jessica Talisman: Yeah.
Thd AI Table: ⁓ we can break customer trust, we can break societal trust, we can break trust between job and job offers. and these are things that once it's broken it's it's hard to clean clean up. ⁓
Jessica Talisman: Yes. Yeah. Like have you ⁓ had a chat bot experience where where you're trying to reach customer service the company has allocated the to a chatbot, right? And and you can't it's just a little window and you can type whatever you want ⁓ none of the answers or none of your your questions or your entries actually match. ⁓ accurately. I mean that's like that's like a first line of trust and and a chatbot architecture is the simplest AI architecture you can engineer. It's like a closed sort of you know situation. So if you can't build like a reliable chatbot architecture, you're you know that's that's your first broken trust bond with your consumers or your customers.
Thd AI Table: ⁓ yeah, and the flip side of that, because I I wanna be clear that, you know, I'm not anti AI, I'm just anti stupid AI. ⁓ so as an example, know, similar to your chatbot example, but where AI is used for good.
Jessica Talisman: yeah.
Thd AI Table: my personal bank uses AI when I call them. ⁓ They information about me and my different accounts and the AI answers a phone and they say, Hey, and ⁓ using information that they have on me, they take a guess on why I'm calling. Nine times out of ten, they are correct about why I'm calling. ⁓ You ⁓ hey, haven't ⁓ used credit card in ⁓ this days, and they know that I usually use it every day. Did you lose your card? ⁓ And it really takes a weight off. Now this one time
Jessica Talisman: Yeah. Right, right.
Thd AI Table: They didn't know what I was calling for and they easily let me get transferred to a human. So that maintained the trust between me and this organization because one, nine times out of ten, they made my life easier. They knew what I was calling for, and we just got to the point. And the one time when they didn't, it was clear that they didn't just have this tool simply to keep me from a person. It was put there to like help make my day better. And it did. So, you know, companies can adopt tools.
Jessica Talisman: There you go.
Thd AI Table: ⁓ in ways that are intelligent, but it does take strategy, it takes thinking. And the thing ⁓ that there are firms that are doing things in this way and you'll have to compete against them. So that's why like it's important for everybody to kinda ⁓ get on because there are shining stars already that are starting to stand out that aren't, you know, implementing in ways that are sloppy
Jessica Talisman: Yeah, yeah.
Thd AI Table: so one other question. We're seeing companies th ⁓ more and more are starting to adopt responsible AI policies. What are of the things that you think should be addressed that are often overlooked when it comes to
Jessica Talisman: Something that happened is these models became commercially available, the adoption was based off of automation first, removing augmentation step. So what augmentation means is that you're augmenting human work before you trust it to automate. There should always be human in the loop. those that critical step is something I would love to see more organizations adapt to and adopt. ⁓ definitely slows ⁓ timelines down a bit, but it also gives you a certain amount of certainty and a lens into ⁓ aligned AI is with what you expect the automation to look like. the other one is the knowledge management. most organizations do not have dedicated knowledge management or ontologists or knowledge graph departments. I want to see investment. That's that needs to happen. if you do want to invest in AI, you also have to invest in knowledge. ⁓ and so that's missing. And and that means either you upskill your employees or ⁓
Thd AI Table: Uh-huh.
Jessica Talisman: You start to hire experts in those fields, information scientists, library scientists, ontologists. ⁓ start to hire. Now, the th the most compassionate thing, because that leads to another thing, is that is that there's a huge disruption of jobs. people either losing their jobs or ⁓ I've ⁓ seen and talked to people recently where they've just accepted demotions. And it's scary. I mean, like there needs to be a human-first approach as well. And so one of the most compassionate things a leader can do, and I've seen some great leaders do this through Graph Academy. We've implemented Knowledge Graph Academy internally within organizations. And that's upskill data scientists, data peoples, governance, content strategists in place ⁓ to understand the management piece and the knowledge infrastructure piece. And it is the most passionate thing an organization leaders can do is to invest in education of their people.
Thd AI Table: I would say not just compassionate but I it's also part of a good business strategy personally. we see taking very different approaches what this ⁓ looks like ⁓ I think be interesting ⁓ say five years the road to see the differences in outcomes. I know FedEx is a an example I always ⁓ go to who who's taken the route of ⁓ upskilling their staff and using ⁓ committing to it as as something that augments human potential and thinking about how do they change existing within the organization to be ⁓ in alignment the innovation that we are seeing become available. anything else you wanna share with our audience? Pet peeve, something you wanna get off your chest and you're like, I want I don't wanna see this anymore or I wanna see more of this. Something anything else you're passionate about?
Jessica Talisman: ⁓ gosh, dialogue. I mean this is more of a general. There needs to be dialogue and reading. ⁓ I seem really, really basic, but there are such you know, we're reading less. There's less And it's through like reading and dialogue and critical thinking and just being able to sort of, you know, break down the silos within these di these conversations that are happening online. there's a certain, you know, there's a certain competitiveness and guarded secrets of of how people are doing things or implementing and ⁓ ⁓ you know that where learning happens, you know, sharing each other's learnings, it's each other's wins, but also really trying to understand that and sit with problems rather than Focusing so much on solutions. So the problem solving effect. I mean, I write about this a lot on my Substack newsletter. because I see it as as larger than than ⁓ just immediately how we're doing AI within our organizations. There's like a a deeper cultural and personal impact. ⁓
Thd AI Table: Yes.
Jessica Talisman: Yeah. And I d I just really value those those tenants. And, you know, I have one kid in college, one kid that just graduated and and something that's really impressive with them and their friends is that they're, you know, there's this movement where they're actually swearing off AI and going into careers that have nothing to do with AI, trying to avoid it. and so there's, you know, it it's important we keep the dialogue open and to also realize that like what we're building into the future, our kids involved, you know, whether we like it or not, and they're impacted. So there's just a larger, heavier weight that's gonna have huge societal impact. So that's, you know, one of the reasons it's really important to care.
Thd AI Table: Thank you so much for for putting it in that broader context. ⁓ Thanks coming on.
Jessica Talisman: Thank you, dear. It was really fun.
Jessica Talisman: Sure. ⁓ so as Dia said, I'm Jessica Talisman. I've been working in around data, more around semantics, ⁓ ⁓ geez, 1997. That's I really started. I have at places like Amazon, Adobe, Overstock, which is now Bat Bed Bath and Beyond, and Department of Justice. right now I work consulting and I also co-run ⁓ the Knowledge Graph Academy with Tony Seal and Katarina Kari.
Thd AI Table: Yes, thanks for plugging that. I've been watching the work that you're doing with the Academy. Very exciting stuff. ⁓ we'll leave a link to that on this podcast. all so let's go ahead and dig in. My first question. something we've been seeing a bit in the news lately is how AI systems ⁓ have been to exhibit bias sometimes, ⁓ and important decision making, ⁓ things like whether a job candidate passes an AI screen. and goes on to have their resume viewed by a human. Can you talk in layman's terms for us about why these biases happen and if you have any recommendations for businesses to get around them?
Jessica Talisman: Sure. I mean one of the problems especially with AI is that it's a black box. So it comes already with training data. And if you consider the world wide web or the internet, that's bulk of the training data. So there's inherited bias. ⁓ there's a great Yeah, yeah. ⁓
Thd AI Table: Can we talk a little bit sorry to interrupt you, can we can we back up to to defining a black box for s for some folks who might not know what that is?
Jessica Talisman: Sure. Yeah, so so there's two basic components, and I'm I'm really distilling this, but basic components of of an AI model, ⁓ large language model. ⁓ one is the parametric layer, and that's the layer that none of us see. That truly is the black box. It's where the training data sits. And so we don't have any insight or idea of how these models have been trained. Now, what happens is that training data is broken up into these little chunks that become tokens, and then it's a statistical prediction model. So it's going to statistically predict ⁓ the next word is in a sentence, for example. So if a lot of the training data contains ⁓ information or or knowledge that is biased, large language model is going to inherit those biases, right? Because that's how its brain You can think of the parametric layer as the brain. That's how its brain has been trained to think in those lines. And then there's the retrieval layer, which is what we all can see or at least have some measure of control around. Not not total control, but that's where we enter our own context or what's being called context. And that's where we can give our own training data or try to steer AI towards ⁓ answers and opinions and and I guess that. if if you think about it this way, it's very hard for any of us to be unbiased. So ⁓ even organizations can insert their own bias if their own data and information within their own organization is biased towards certain opinions.
Thd AI Table: Yes, so I think ⁓ you're you're illuminating an important point that to a large extent AI is a reflection of ourselves. ⁓ in sense as humans, you know, we're not perfect, so AI is is is not gonna be perfect either. ⁓ but there are strategies that we can use to try to get it less biased.
Jessica Talisman: Total reflection. Less biased. And I mean this has been a struggle for a long time. I I, you know, I have a master's in information in library science and I am a librarian. And a part of the job of a librarian is to try to remove as much bias as possible. you're trying to stay objective. ⁓ I do so in my own writing, is to balance the points, so a point and a counterpoint. ⁓ somewhere that we've seen bias show up ⁓ was Amazon. ⁓ it's resume review agent. I guess you could call it an agent, but it was an AI implementation to help to screen resumes. And so what happened with that algorithm, and it was actually discovered and it's been written about and published, there was a documentary about it, was that it was specifically not ⁓ letting resumes pass through it had a bias against women. ⁓ anyone who went to a women's college when that algorithm was being used up to a point, were not passed through the filter. So that's an example of where the bias shows up.
Thd AI Table: Yeah, so so yeah, these can have some serious implications. walk through from your perspective strategies for businesses to that, that includes adding that context. What does that process look like?
Jessica Talisman: Yeah, so I I actually have been thinking about this a lot because something I run into quite a bit is businesses will try to scope and scale the context that they give only relative to their own world and their own language. That's a big problem because inevitably, especially if a business is interacting with the actual world. then ⁓ can easily be introduced because you create a very narrow sliver into that sort of idea of the and ⁓ that that your ⁓ data information holds, right? So an example is ⁓ you know, when I worked at Adobe, for example, a feature almost every single product was called journeys. And so that internal definition of what a journey was didn't really match the external definition of what that was. So that's one place that you see people narrowing their sort of scope and field and only ⁓ I should say, assuming that AI understands your internal world. So I think of it like touching grass. There's sort of a balance in which you have to relate how internal definitions align with external definitions, because guess what? your your customers are coming from the external world from the actual world they're not coming from internal so that's one thing the other thing is including context more than what your exact business holds you want to ground your business concepts for example in context in the real world, which maybe mean that you have to expand your small world, your myopic world to include more than what you cover and more than what what's meaningful to you in your business. And and that seems really confusing, but I've I've run into this quite a bit with businesses where they're like, we just wanna wanna wanna narrow down and edit down to only the things that matter inside of our business. ⁓ But you don't put those terms that that meaning In context within the outside world, larger than what your business is, then you're going to lose a huge lens into opportunities, also contextual understanding of your business. Because you have to understand how your business is placed within the larger realm. ⁓
Thd AI Table: Yeah, that that is a a huge point. my next question, a big part of of what we refer to in our discipline as responsible AI, ⁓ which the practice of ⁓ designing deploying AI systems that are safe and fair, ⁓ is making sure AI systems are transparent and easily traceable with respect to the decisions that They're making can you talk about how sciences can help with this?
Jessica Talisman: Sure. I just wrote a piece called an Ontology that evaluated the data landscape right now, and I I brought an information science lens to it. ⁓ of ⁓ problems is because we have that black box effect where we can't see into how certain decisions are made, ⁓ leaves ⁓ a lot of organizations and people kind of handicapped in not being able to really transparently into these models. ⁓ So the thing that's starting to emerge is the idea of data sovereignty. And sovereignty is the idea that you can own your own data, regardless of a vendor platform. Every organization, every ⁓ institution should be able to own its own data very transparently and ex and and that in and of itself is one of the keys to being able to transparently explain how LLM arrived at certain answers. We're this come up ⁓ the EU AI Act, once it goes into actual force, or the data, the data act, the US government's data act. ⁓ there are requirements for data sovereignty. And those requirements are going to help Hopefully, for to actually implement explainable AI. And explainable AI is that you can actually point to and detail how an answer was generated and what data was used to generate that answer. So I'll give you an idea. If you have ⁓ ⁓ based say that you have a database and the database that vendor offers some sort of agentic AI within the product. of the issues is who owns responsibility for explaining how those answers ⁓ were Right. In a lot of those cases, the the user or the customer doesn't even have visibility into that data. So one of the base requirements is understanding and having ownership, really well-organized, curated ownership of your data, and being able to Cite sources. So it's just like if if you're writing something or you're quoting something, you know, best practices is to cite sources. It's not, you know, otherwise it's plagiarism. ⁓ then we get into a world of ⁓ ⁓ starting to focus on and prioritize things like provenance, which is a very important concept within both history and library science and information science. So is when you can actually attach attribution to your data object. objects and your information and that things are clearly organized in a way that you can easily through an evaluation process point back to the source identify where an answer came from.
Thd AI Table: ⁓ within orgs from a leadership perspective, what roles should have purview over this, do you think?
Jessica Talisman: That's interesting. That's a really interesting question because you have sort of a musical chair situation with CDOs. ⁓ They don't really have a long tenure. So the natural answer would be governance, you know, whoever's head of governance and CDOs, at least for that aspect. But time has passed, what I'm realizing more and more, especially through my own work, is it really is a collaborative cross team effort.
Thd AI Table: Mm-hmm.
Jessica Talisman: It it is something that needs to be owned by everyone. It needs to be a muscle that's built within an organization. And there needs to be sort of organizational policies in place so that there's repeatable patterns. of the problems is most organizations, not all, have ⁓ sort of vectors of control in terms of each person's ⁓ implementations of AI. So implementations and deployments are wildly different and everyone's using different data. In fact, it it is tending to create deeper silos rather than breaking down silos, which points to the fact that, like, okay, if if we don't all get on the same page have ⁓ ⁓ strategy around this, ⁓ then then there can be no leadership for lack of strategy.
Thd AI Table: I agree. Strategy first, everyone. Strategy first. Especially with these lawsuits coming in, strategy first.
Jessica Talisman: Strategy first. It's getting so lost. Strategy first. And like the stakes are high. Yeah.
Thd AI Table: my next question, people interact with ⁓ LLMs Claude, Gemini, ChatGPT, we're all familiar with hallucinations or when an AI system provides you some output that isn't true or doesn't ⁓ make any can you talk to us about what makes this happen? And is there anything that folks can do to reduce ⁓ hallucinations?
Jessica Talisman: Yeah, and so for hallucinations, you know, I'm a big proponent and supporter of know, knowledge graphs of symbolic AI. So a lot of people don't realize that yes, we are all engaging with what we know of as AI, but the AI that we're playing with is generative AI. It's large language models, which is a st it's a statistical prediction model. So again, it's it's just statistically predicting. Another to that ⁓ is there's other types of AI. And what's emerging from the research is symbolic AI, it counterbalances hallucinations. what is symbolic AI? It's things like taxonomies and ontologies and knowledge graphs. So those are becoming increasingly ⁓ to organizations. So means curating and organizing your data. It means describing your data in a machine-readable format. They become symbols so that what happens is the statistical prediction model ⁓ is to have Reference points to actual facts within a system, because you can say something like, you know, has this podcast that's like an actual fact, and I can verify it because it's represented, for example, my taxonomy or my knowledge graph or you know, within my ontology. So one vector of control for hallucinations is to actually build fact-based symbolic models, you know, things that are easily consumable by machines. They're machine readable. ⁓ And they lead to easier interpretation ⁓ responses. Now, again, we have control over how the models arrive on our doorsteps ⁓ or our you know computers. There, there's that black box effect. So ⁓ of issues is when an answer cannot be easily explained, that's that points needing to build out that sort of fact-based system, those reference points, and it becomes a point of stability. Now the one thing is it's not it's not a one-off, right? Building these things are not one-off. It's not like spin up and generate an ontology for for each person's little sliver of AI implementation. There needs to be sort of consistency. But the other thing for hallucinations is that you know symbolic AI systems, so that means you know your taxonomy, your ontology, that it becomes your organization's memory as well. And that includes provenance. So you're basically building your organizational's brain. so to minimize hallucinations, you need those reference points. To the earlier question, you need to have explainable AI, you need evaluations. and start ⁓ to a record or a log. of these sort of, you know, if we're talking about performance metrics, we need to start measuring the performance of provenance, truthfulness, reliability, those sorts of things. And a lot of organizations are not doing that. They're just accepting the hallucination as fact.
Thd AI Table: Yes, that's what I'm seeing as well. And you talk about this work, is this something that you see falling under CDOs as well? Or is more flexible? Are there other places it could it make sense for it to be within an organization? I know not all organizations have a CDO.
Jessica Talisman: Yeah, I mean if there's no CDO, especially, I mean, ideally a CDO is leading, but again, so many organizations do not have CDOs. ⁓ so I I again think it needs to be something that's like a policy governance. You know, I was thinking about it you you s organizations, especially because of things like GDPR, which was, you know, the legislation to to ⁓ you know regarding privacy that came out of the EU. And one of the important things is like how did organizations adapt and adopt that sort of policy of protecting people's private information by law? And so that had to be instituted as part of governance and classification, right? So having a system in place that could reliably, you know, and consistently protect throughout an organization. And that same thing has to happen. So ⁓ really is foundational. ⁓ I I call it knowledge infrastructure. It's something we don't have. Most organizations do not have today. ⁓ organizations do not have people trained up ⁓ ⁓ in information science. in semantics ⁓ in knowledge, management. ⁓ and ⁓ my is. Is that organizations are going to start, you know, we hear of the great upskilling or, you know, those skill sets need to be in place, such as knowledge management and semantics and information, all of those things. ⁓ right now organizations limping along and people are kind of trying to do things on the fly ⁓ without having ⁓ that themselves of how to do that.
Thd AI Table: Yeah, to me I I see that as well. I think it's a matter of data there are a few ⁓ shining stars are doing this work well and I I think that we're gonna see them break out from pack. because I think there is serious ROI that's connected to being ⁓ a mature, you know, some folks calling an intelligent enterprise. So I think ⁓ When organizations are moving towards that territory where they want to become an in an intelligent enterprise, where they understand AI not simply as a tool, but as something that has some ⁓ ramifications for how business operates and how it structures itself to operate. ⁓ when you when you dig into changing that this is a part of the structural change that will need to occur for for businesses who want to, you know, be at the top of the pack.
Jessica Talisman: And if you look at the training data that's most consumed and used by AI, because there was a great Washington Post article last year that evaluated, for example, Llama, which is Facebook's open model, and 15 million training sources and was able to evaluate because that's an open model. So they were able to evaluate what what constituted the training data. ⁓ And top 10 sources came from pretty well structured, good knowledge sources like New York Times and Wikidata. Wikidata ⁓ is And Wikipedia, it's a knowledge graph. DBpedia was another big source. It's a knowledge graph. So that's the other part of it is like that knowledge management piece that steps beyond data and goes towards, you know, we we see this happening with content and documents and and you know, people's share points that they didn't really care to to organize before. Those sorts of things start to become really important.
Thd AI Table: Yeah. we're seeing more organizations ⁓ that are training their own models. Can you about why it's dangerous to let an AI model train or ingest freely internal corporate or personal communications and ⁓ what boundary might look like when it comes to that?
Jessica Talisman: Yeah, so ⁓ I mean I think that's gonna be the new wave ⁓ w once everything ⁓ it's gonna have to settle down in some however things shake out, but these the models are becoming the open models are becoming our frontier models are becoming increasingly expensive. So naturally people are turning into, you know, training their own small to medium-sized models. ⁓ and just the same problems that occur with the black box effect of these frontier models, ⁓ you you want to be careful and curious. First of all, what you're giving the model, but also you have customer data, you have financial data. There are things that are proprietary that don't belong in training data that could be very, very dangerous. so security is starting to become an increasing priority across it. You can look at things like open claw, we're seeing that mirror effect of what's happening with models that are open to the public, you know, or available to the public. Yeah. The same rules apply with small and medium-sized models as well. So people's own models that they're training internally. So there need to be guardrails. that's ⁓ really important. I mean, that takes a lot of curation, a lot of organization. security to be looped in from the very beginning because there are vulnerabilities in these models. You open up doors that otherwise wouldn't be available. there are things like does your model reach reach out to the internet? is there any sort of interoperability or or dependency on on the internet for answers or for, you know, as soon as you open up that door, ⁓ expose a vulnerability again. ⁓ who has access, who gets to include information or knowledge? Is there are there point people who are determining, you know, and curating and organizing the information available to the small to medium size models? But the security ⁓ threats that are emerging are are pretty serious. And so there there have already been hacks that that that have been pretty scary. So I think organizations are starting to feel the effects.
Thd AI Table: Yeah, and I think it highlights the point you made earlier about the importance of collaboration when it comes to this work, making sure you have a security person that's in governance, data person participating in governance, someone from legal ⁓
Jessica Talisman: You know what's interesting, like w Having been at some of these organizations, one thing that went out the door that used to be a priority prior to 2021 was vendor evaluations. Really rigorous vendor evaluations of software tools platforms. If I work for an organization and I want to onboard, for example, or I want to purchase or acquire, you know, in any of those effects, another tool, right, for my organization. I'm I I need a knowledge graph tool for my organization. I go through rigorous vendor analysis that's set out by the organization. You can't pass go, I have to present at a council to show them the tool I want. Security goes through a big evaluation. is that not happening with AI models?
Thd AI Table: I think it's because everyone's in a rush. ⁓ very interesting to watch you're right, the priority that's been placed ⁓ on evaluating solutions, it has significantly over the years. And I would argue, in this moment it it makes a more sense for it to be increasing given the ramifications of of adopting something that is faulty. We still haven't figured that out in the legal system. I mean there are lots of lawsuits are going through the court right now but we haven't determined who's liable things go wrong. Is it is it the tool provider, is it the company that contracted with the tool who who's ⁓ responsible here. ⁓ ⁓
Jessica Talisman: Right. So
Thd AI Table: While that's up in the air, it makes sense for everybody to cover themselves. so
Jessica Talisman: There's a concept in in in libraries that's in archives, which is really interesting, which is ⁓ in the American College Research Libraries Association, they they they created sort of tenets of AI. But but one thing that's pretty preval prevalent is no human, no AI. And I really, really like that because if we don't have like humans in the loop and evaluating ⁓ before we get to sort of that emergency or or legal then course ⁓ when happens, it's too late. and ⁓ you so you don't have those things in front of, but ⁓ there are things in terms of copyright, right? I mean, there's there's other issues that extend further. and those just policies haven't been updated. We had the d the digital. Millennial Copyright Act. And that ⁓ was created when internet came around because it a format or a medium that hadn't existed before. So it was to protect, you know, people's work, but also extend a little bit further because they needed, you know, was this new medium. We haven't addressed this new medium, this new format. No.
Thd AI Table: No. We're we're still having conversations. We're still figuring it out. it's it's up in the air. ⁓ sometimes I think companies forget that part that it's up in the air. Like ⁓ doing all this stuff, but like the legal framework behind it is not solid. It it just because we have some people can do something one day doesn't mean that we can do it the next. so it'll be interesting to see how everything shakes out. We're we're kinda all along for the ride.
Jessica Talisman: So up here. It's not there.
Thd AI Table: But I do think that it's important that people like yourself in and and share insights their professional perspective ⁓ of we need to consider as we think about these new policies that'll need to come into place.
Jessica Talisman: I mean I I I do have like on respect, there are some ⁓ organizations, one that I'm working with, that are very highly regulated. And they are, you know, some of these organizations are being very thoughtful, careful, and mitigated. And and I think those are the ones that are gonna emerge. And they may be working more slowly, you know, good things take time.
Thd AI Table: Yeah, they're more thoughtful. That's my that's my experience as well. And I just want to give a shout out to the banks because ⁓ a lot of the banks a lot of the bank ⁓ executives are really thinking deeply about this. And they're not just out and adopting tools and using a checklist and say, hey, AI, we're doing it, we're doing it all. they're really thinking about how to it in ways to best serve their customers.
Jessica Talisman: Yes. Yeah.
Thd AI Table: ⁓ Okay, so my next question, how does a company maintain control over truth language and product names and industry regulations are changing frequently?
Jessica Talisman: gosh, that goes back to the knowledge management muscle. It's like if you don't have people that are managing, it's like a chain. I was thinking about this in terms of change management for ontology, taxonomy, knowledge management work. So working on a project right now at an unnamed company. ⁓ And is the most interesting thing is how it's education, education first and foremost. Most people do not understand. There's there's think that it's a one and done. I build this knowledge structure, this taxonomy, this ontology, this thing that's gonna help to reduce or eliminate hallucinations ⁓ and it's done. And just not the case. It's like, would you do that with your database? As fast as the data in your database changes is as fast and as as careful as you have to be about managing the knowledge that you give to your mom.
Thd AI Table: Mm-hmm. ⁓
Jessica Talisman: It is lockstep. so I'm, ⁓ know, I still see people saying, okay, well, can I just have AI generate this for me, the knowledge structures, you know, the taxonomies, the ontologies for me, and call it done. Not really. Like, no, you can't because if your data is continuing to change, then you're basically, of course it's gonna go stale. That's the secondary complaint. It's very interesting because I'm not sure if people register it. The first complaint is takes a lot of time to organize and to build and to curate. It's like okay. Yeah. ⁓ then they say, okay, ⁓ can AI do it for me? Is usually the second question. It's like if you want AI to organize your business, sure.
Thd AI Table: Yeah, that's a hard.
Jessica Talisman: But you're gonna get that quality of of understanding, which is, you know, introducing hallucinations because you haven't taken the time to curate and to determine what is meaningful for your business. then the third step is: okay, so I built it once ⁓ and it's going great. And you're like, okay, thumbs up, that's that's wonderful. It's going great. They they took the time to build it. And then you hear people complaining, it's stale. Ontologies are stale, taxonomies are stale. And it's like, yeah, because you didn't maintain it. You don't have people that are constantly curating and updating it at the rate of the business. Of course it's stale. So ⁓ something that's occurred to me is like a lot of this involves abstract logic and reasoning. A lot of this includes areas that we haven't cared about. And then there's the the adopted and adapted principle of build fast and break things, which we cannot apply to AI. is not the type of solution or implementation that we can build fast and break things. Like we're moving back to that data maturity. This is something that actually good things take time to build it needs to be a constant iterative process.
Thd AI Table: and I mean I would argue that we can build fast and break things, but we probably shouldn't. ⁓ I really want folks to like have a vision of of what that means when we say that phrase of what are the things we are breaking, ⁓ or that we can potentially break it comes to AI because those are serious things and the biggest thing that we can break is we can break trust.
Jessica Talisman: Yeah.
Thd AI Table: ⁓ we can break customer trust, we can break societal trust, we can break trust between job and job offers. and these are things that once it's broken it's it's hard to clean clean up. ⁓
Jessica Talisman: Yes. Yeah. Like have you ⁓ had a chat bot experience where where you're trying to reach customer service the company has allocated the to a chatbot, right? And and you can't it's just a little window and you can type whatever you want ⁓ none of the answers or none of your your questions or your entries actually match. ⁓ accurately. I mean that's like that's like a first line of trust and and a chatbot architecture is the simplest AI architecture you can engineer. It's like a closed sort of you know situation. So if you can't build like a reliable chatbot architecture, you're you know that's that's your first broken trust bond with your consumers or your customers.
Thd AI Table: ⁓ yeah, and the flip side of that, because I I wanna be clear that, you know, I'm not anti AI, I'm just anti stupid AI. ⁓ so as an example, know, similar to your chatbot example, but where AI is used for good.
Jessica Talisman: yeah.
Thd AI Table: my personal bank uses AI when I call them. ⁓ They information about me and my different accounts and the AI answers a phone and they say, Hey, and ⁓ using information that they have on me, they take a guess on why I'm calling. Nine times out of ten, they are correct about why I'm calling. ⁓ You ⁓ hey, haven't ⁓ used credit card in ⁓ this days, and they know that I usually use it every day. Did you lose your card? ⁓ And it really takes a weight off. Now this one time
Jessica Talisman: Yeah. Right, right.
Thd AI Table: They didn't know what I was calling for and they easily let me get transferred to a human. So that maintained the trust between me and this organization because one, nine times out of ten, they made my life easier. They knew what I was calling for, and we just got to the point. And the one time when they didn't, it was clear that they didn't just have this tool simply to keep me from a person. It was put there to like help make my day better. And it did. So, you know, companies can adopt tools.
Jessica Talisman: There you go.
Thd AI Table: ⁓ in ways that are intelligent, but it does take strategy, it takes thinking. And the thing ⁓ that there are firms that are doing things in this way and you'll have to compete against them. So that's why like it's important for everybody to kinda ⁓ get on because there are shining stars already that are starting to stand out that aren't, you know, implementing in ways that are sloppy
Jessica Talisman: Yeah, yeah.
Thd AI Table: so one other question. We're seeing companies th ⁓ more and more are starting to adopt responsible AI policies. What are of the things that you think should be addressed that are often overlooked when it comes to
Jessica Talisman: Something that happened is these models became commercially available, the adoption was based off of automation first, removing augmentation step. So what augmentation means is that you're augmenting human work before you trust it to automate. There should always be human in the loop. those that critical step is something I would love to see more organizations adapt to and adopt. ⁓ definitely slows ⁓ timelines down a bit, but it also gives you a certain amount of certainty and a lens into ⁓ aligned AI is with what you expect the automation to look like. the other one is the knowledge management. most organizations do not have dedicated knowledge management or ontologists or knowledge graph departments. I want to see investment. That's that needs to happen. if you do want to invest in AI, you also have to invest in knowledge. ⁓ and so that's missing. And and that means either you upskill your employees or ⁓
Thd AI Table: Uh-huh.
Jessica Talisman: You start to hire experts in those fields, information scientists, library scientists, ontologists. ⁓ start to hire. Now, the th the most compassionate thing, because that leads to another thing, is that is that there's a huge disruption of jobs. people either losing their jobs or ⁓ I've ⁓ seen and talked to people recently where they've just accepted demotions. And it's scary. I mean, like there needs to be a human-first approach as well. And so one of the most compassionate things a leader can do, and I've seen some great leaders do this through Graph Academy. We've implemented Knowledge Graph Academy internally within organizations. And that's upskill data scientists, data peoples, governance, content strategists in place ⁓ to understand the management piece and the knowledge infrastructure piece. And it is the most passionate thing an organization leaders can do is to invest in education of their people.
Thd AI Table: I would say not just compassionate but I it's also part of a good business strategy personally. we see taking very different approaches what this ⁓ looks like ⁓ I think be interesting ⁓ say five years the road to see the differences in outcomes. I know FedEx is a an example I always ⁓ go to who who's taken the route of ⁓ upskilling their staff and using ⁓ committing to it as as something that augments human potential and thinking about how do they change existing within the organization to be ⁓ in alignment the innovation that we are seeing become available. anything else you wanna share with our audience? Pet peeve, something you wanna get off your chest and you're like, I want I don't wanna see this anymore or I wanna see more of this. Something anything else you're passionate about?
Jessica Talisman: ⁓ gosh, dialogue. I mean this is more of a general. There needs to be dialogue and reading. ⁓ I seem really, really basic, but there are such you know, we're reading less. There's less And it's through like reading and dialogue and critical thinking and just being able to sort of, you know, break down the silos within these di these conversations that are happening online. there's a certain, you know, there's a certain competitiveness and guarded secrets of of how people are doing things or implementing and ⁓ ⁓ you know that where learning happens, you know, sharing each other's learnings, it's each other's wins, but also really trying to understand that and sit with problems rather than Focusing so much on solutions. So the problem solving effect. I mean, I write about this a lot on my Substack newsletter. because I see it as as larger than than ⁓ just immediately how we're doing AI within our organizations. There's like a a deeper cultural and personal impact. ⁓
Thd AI Table: Yes.
Jessica Talisman: Yeah. And I d I just really value those those tenants. And, you know, I have one kid in college, one kid that just graduated and and something that's really impressive with them and their friends is that they're, you know, there's this movement where they're actually swearing off AI and going into careers that have nothing to do with AI, trying to avoid it. and so there's, you know, it it's important we keep the dialogue open and to also realize that like what we're building into the future, our kids involved, you know, whether we like it or not, and they're impacted. So there's just a larger, heavier weight that's gonna have huge societal impact. So that's, you know, one of the reasons it's really important to care.
Thd AI Table: Thank you so much for for putting it in that broader context. ⁓ Thanks coming on.
Jessica Talisman: Thank you, dear. It was really fun.
