Matrix On Point

Matrix on Point: Propaganda Machines

In the age of AI, what is the future of “public truths”? AI-driven tech platforms and companies, influencers, and deepfakes are issuing ever-new challenges for journalism, media integrity, and related mechanisms through which democratic societies hold power to account. At the same time, universities – erstwhile guardians of critical inquiry and independent knowledge production – are increasingly entering into partnerships with AI companies, in a new kind of knowledge compact.

Recorded on November 5, 2026, this panel brought together scholars of AI and influence in journalism, media, and California’s universities to address this technology’s increasingly powerful role in the current U.S. landscape of information and knowledge politics. The panel featured Maty Bohacek, Researcher at Stanford University; Martha Kenney, Professor of Women & Gender Studies at San Francisco State University; Martha Lincoln, Associate Professor of Anthropology at SF State; and Edward Wasserman, Professor of Journalism at UC Berkeley. Cori Hayden, Director of Social Science Matrix, moderated.

The event was co-sponsored by the UC Berkeley Institute of Governmental Studies (IGS), the Berkeley Economy and Society Initiative (BESI), the UC Berkeley School of Information, and the Department of Electrical Engineering and Computer Sciences (EECS).

Podcast and Transcript

A podcast recording of this event can be found below or on Apple Podcasts.

(downtempo electronic music)

[CORI HAYDEN]

Welcome, everyone, to Social Science Matrix — to the glorious eighth floor. And it’s still light out for another hour or so. I’m really glad to have you here. My name is Cori Hayden and I am the director of Social Science Matrix, and also a member of faculty in the Department of Anthropology.

And I’m really delighted to have you all here and to welcome our fantastic panelists for this discussion. You know, this is a panel about AI, but at a slightly sideways angle. There are so many ways to slice the question of AI, of course.

We’re framing this thanks to my predecessor, Marion Fourcade, who was director until July of this year. Through this evocative title of Propaganda Machines, it raises the question, of course, of influence, raises the question of the relationship between truth and something that might be askew to truth. I’m not sure what the opposite of truth is anymore.

I reframed this program description with a question about the future of public truth in an age of AI. Now, it is quaint to pretend like public truths were totally clear, fine, transparent, and solid before AI. We know that we have been in a crisis of public truths for quite some time now.

AI in its many dimensions ups the ante on those questions and transforms them in some interesting and strange ways. We’re gonna be talking in this panel about questions you might associate with propaganda, questions of advertising, questions of influence. But also questions you might not anticipate, such as the role of universities as guardians of public truths, but also as becoming more and more wrapped up in AI as a political, economic, and technological machine.

And we’re going to be asking some questions here about the relationship between AI and what we might call the knowledge compact that public universities especially are meant to represent. So this panel is bringing together scholars of AI and influence, understood in this slightly creative way, who work in journalism and media studies, and who work on and in California’s universities to address AI’s increasingly powerful role in the current US landscape of what we might call information and knowledge politics. So today’s event is co-sponsored by the UC Berkeley Institute of Governmental Studies, the Berkeley Economy and Society Initiative (thanks, Marion, for being here), the UC Berkeley School of Information, the Department of Electrical Engineering and Computer Science, and Berkeley School of Journalism.

Now, before I turn over to our panelists, and in fact introduce them in my dual capacity as moderator today, I just want to briefly mention a couple events coming up.

A couple of co-sponsored events here, November 12th and 23rd, book talk, something on Unsettling Sovereignty and Nuclear Weapons, US Extraterritorial Power and Postwar Oceania. This has a very interesting tie-in with a multidisciplinary arts and environment initiative on campus called Our Life Is Here, about Marshallese Islanders, and the legacy of nuclear imperialism. I encourage you to check out that event as well as everything else happening under that rubric.

November 12th, we’re also talking about a postmortem on the 2026 Midterm elections, different outcome and a different conversation than I think many people were perhaps anticipating. And on November 24th, just before Thanksgiving, a fantastic Author Meets Critics panel with historian Matthew Shutzer on fossil fuels in India. Okay, so please join us for all of those events.

And for now, let’s go back to today’s events. I’m gonna introduce the panelists in the order in which they will appear, and I do wanna say that unfortunately Martha Lincoln from SF State is unable to be here due to illness. But her colleague Martha Kenney, other Martha, will be sharing Martha Lincoln’s comments as well as Martha Kenney’s comments.

Okay, a few notes on our presenters. Maty Bohacek. Did I say your name correctly?

Maty Bohacek is a researcher at Stanford University working with Professor Maneesh Agrawala. Maty works at the intersection of AI and media integrity. He’s regularly cited as an expert on deepfakes.

Together with Hany Farid, who is formerly a UC Berkeley professor, Maty famously created an AI version of CNN anchor Anderson Cooper that was used to highlight The danger of synthetic media. Now, in a world that’s becoming increasingly AI-mediated, Maty’s goal is to build AI systems and networks that are inherently trustworthy and interpretable, resulting in a kind of more productive and democratic discourse. Yes, bravo already.

(applause)

Thank you for coming. We’re done. Edward Wasserman is professor of journalism and former dean of the Graduate School of Journalism, from 2013 to 2020.

Before coming to Berkeley, he spent a decade as the John S. and James L. Knight Foundation professor of journalism ethics at Washington and Lee University. Wasserman’s journalism career spanned 30 years. His CV is long and impressive.

I’m going to jump right to your undergraduate degree. Wasserman received a BA cum laude in politics and economics from Yale, a license in philosophy from the University of Paris, and a PhD from the London School of Economics, where his thesis was on the politics of innovation and communications technology. I skipped all the other amazing things.

Martha Kenney, professor of women and gender studies at SF State, is a feminist science and technology studies scholar and the co-founder of the Science, Technology, and Society Hub at SF State. Her current work focuses on the politics of social media platforms and AI products in education, medicine, and public health. Together with cultural anthropologist Martha Lincoln, she’s currently working on the monograph Marthas Against AI: Inside the Fight for the Future of Higher Education.

She teaches classes in feminist science studies, critical AI studies, video game studies, and speculative fiction. And one word about Martha Lincoln, just to say Martha Lincoln is a medical and cultural anthropologist. who has written about the cultural politics of public health, biopolitics, and the effects of political economic change on health systems and health outcomes. She’s written about contemporary epidemic politics in contemporary Vietnam and serves as co-editor-in-chief of the Journal of Vietnamese Studies.

And her current project is based in the US, which focuses on the affective economies in medical crowdfunding for cancer. So the new project with Martha Kenney here is a departure, but very interesting one around the politics of universities today. Okay, that’s a lot from me.

Let me turn it over to our speakers, and let’s start with Maty, and I’m going to magically call up your… PowerPoint.

[MATY BOHACEK]

All right, well, thank you so much for being here. It’s great to see you all. And thank you for the very generous introduction.

My name is Maty Bohacek. I’m at Stanford, and my goal here today as the computer scientist on this panel will be to try to frame this in the sort of current work that’s happening in AI, both in the sort of research and more academic circles, and also in the very pervasive and salient AI circles in the industry. The question that we’re trying to wrestle with here today, and the The framing of this whole event is, as was said, how will AI affect public truths?

And I thought that one way, or one sort of interesting way to frame my remarks here will be this picture right here. Who here is familiar with this picture that I’m showing? Excellent, amazing, right.

So in that case, I’ll actually skip the introduction and we’ll get back to it at the end, But I just want you to think about this moment in time and what led to it, what it meant, and in what ways we might find ourselves or sort of see some parallels between that moment and today. This obviously, or famously, is Truman holding up the newspaper that claims that Dewey defeated him, even though that was not the case, and the newspaper was printed before.

I’ll get to that moment later. Let’s see. So the question of how AI will affect public truths from the sort of computer science perspective can be approached in many different ways.

One way that I’m compelled to sort of go to here on this campus is To cite or to talk a bit about Hany Farid, who I had the pleasure of working with on this campus for multiple years. Hany is the incredible researcher who founded the field of media forensics, which is trying to understand different ways of sort of scrutinizing imagery and video and digital content and see who made it, if anyone edited it, and sort of similar kinds of questions. And I was working with Hany here on this campus for four years before going or moving down to the other university in the Bay Area.

The better one. Although I will say, my husband who’s in the audience is here at Berkeley, so I can’t really— I need to be careful with my remarks. But Hany really sort of opened these questions way ahead of our big AI boom moment.

He was asking, how will AI shape information, the way we think about the world, we perceive news, six, seven, even more years ago. And so I wanna start actually there, because back then, so the pre-AI— Oh, for some reason we lost some of my titles. So you’ll have to imagine some of the titles that were on this slide.

These were titles from newspapers. No, that’s all right. That’s all right.

Somewhere between the keynote and the PDF, some of this disappeared. But so the point of this slide is to sort of illustrate that before this current AI boom, before this current AI moment, the kind of threats that people feared Would be brought about, with AI would be these sort of individual or single attacks, right? So, for example, someone could create a deepfake of a politician and make that politician look, or sort of say certain things that they didn’t say, and that could sway an election.

That was very much the kind of threats that people were expecting to come six, seven plus years ago. And so, given that that was our sort of threat model, or like our thinking, our expectation of AI’s impact on the discourse, on public truths, our understanding of what’s happening, we sort of followed accordingly. And so because of that, The sort of research landscape or the work focusing on deepfakes, really sort of took off since 2019, roughly seven years ago.

And one example of such a work or such a response to this sort of expectation of how AI might change public truths was something like this. This was our paper from 2022. Preprint was from 2019, where we built a deepfake detector of President Zelenskyy.

And so, again, because we thought that the way AI would change or impact our discourse were going to be these individual, very salient sort of attacks, someone using AI to manipulate a video and sway an election, stuff like that. We built sort of according interventions, according protections, such as this one. So here, just to sort of give you a very sort of A high-level taste of this paper, we were trying to learn, different mannerisms or different sort of distinctive mannerisms of that speaker, in this case, President Zelenskyy, to be able to distinguish between real video of President Zelenskyy and deepfake or AI video of President Zelenskyy.

And I’m happy to say that in addition to publishing a paper, we also were able to put this to use and help folks who were trying to protect President Zelenskyy in early days of the invasion to actually stop deepfakes. And so it helped a bit, but what I want to claim here, and this is really the crux of these initial remarks, is that our expectation of what AI was going to do to discourse and to public truths was way off. So my claim, and again, a part of this slide disappeared somehow magically, so imagine that there was something above as well.

But the above was really trying to sort of illustrate that previously we were, as I mentioned, we expected that there were gonna be these individual attacks, these individual You know, salient moments where AI was gonna change our discourse. And instead, what’s happening is that we’re seeing this really sort of massive, omnipresent impact of AI. It’s really harm at scale, not these individual threat vectors.

So that’s my first claim. And my second claim is that there seems to be this sort of opaque intermediation, conducted with AI or using AI, and that seems to me like a more salient sort of impacts on discourse than the fabrication part that we were expecting. And so, with these two points sort of established, I wanna give a couple of examples for folks to sort of illustrate how I think about these and what we might do.

So again, the headline disappeared magically from these slides. So by now, either I’m just making all this up, so maybe I’m the big, I guess, trickster or faker, but do trust me that there were headlines. So one example of harm at scale that I was gonna present on this slide was that instead of these individual, highly salient Deepfakes of politicians or of public figures, we’re actually seeing a lot of deepfakes just like over the network, like people receiving a lot of scam phones, sorry, scam calls, and this industry alone seems to be already targeting billions of dollars each year, just in the US and in the West.

And so again, that to me is one example of how the harm that we were expecting to be singular in singular moments is actually very much sort of like scaled and distributed across the board, which changes how you want to react to it. And second, in our recent work, we were sort of wondering how much AI is already out there, how much AI is already impacting the discourse online. So in this paper, we asked how much AI is already contributing to the web, and we found that by mid of 2025, mid of last year, at least one-third of newly published websites on the open web, we classify or we predict are either substantially or fully AI-generated, which is an increase from roughly 2022, when that was basically zero or roughly zero.

And there’s more trends that go along with this. The one that I want to sort of point out is that it’s not just that AI is now writing more of the open web, it’s also reading a lot of the open web because it’s scraping, it’s ingesting, et cetera. And so the role and the purpose of the web as we knew it earlier, which was something for humans to both contribute to and also to sort of consume, has changed to something that’s primarily, or it seems to be increasingly sort of focused on AI.

Now, here’s my second claim, which is that a big impact or a big driver here is also this opaque intermediation that happens as a result of AI. And so one example that I want to give here that folks might be familiar with is that if you now go to Google and you ask for anything. In this case, I actually asked Google about the AI Overview, the feature itself, and instead of seeing a list of links, what you will see now is a summary that was written by a language model, an AI model.

And this summary, as you see here, is trying to sort of put together different parts of these results that you otherwise see below and just present a coherent response in natural language. Now, this is the kind of intermediation where there’s a lot of things that you don’t see as a user. I’m going to give you just two tidbits from our ongoing research.

For example, we found that with ChatGPT or OpenAI-produced sort of interfaces that Are chatbots, we see that the majority of links or references that they use come from their partners, so people that pay them for those links. And you don’t really see that. There’s no disclosure of that.

And second is that, we did this really sort of… I mean, I thought it was going to be like a fun and quick experiment ahead of the election. It turned out to be a couple of days of my Life, and I’m about to wrap up.

We found that if you ask ChatGPT about the election on Tuesday, the vast majority of its responses were actually using this website called election.do. This is an AI-generated website, fully AI-generated. It was propped up roughly 12 to 13 days before the election, and the vast majority, or a significant proportion of facts, of information and numbers that these chatbots were giving you about the election were coming from this website.

Now, we didn’t do a fact check of this, so we didn’t really question or look into the sort of accuracy of this website, but the ability for anyone to prop up a website and For these chatbots to gravitate towards that content, to me seems like a significant thing to understand and to study. And finally, this is my final point, final slide. The big threat of deepfakes in these sort of individual scenarios brought about a lot of interest in the research community, and so we’ve had a lot of papers written about deepfakes, but we really don’t have a lot About AI intermediation, as you can see on the right.

It’s really just sort of getting started. And so I would invite anyone in the audience who is interested or who works on these sort of adjacent topics to consider joining the effort here, because I think that we’ll need much more work, not just from computer science, but also from other fields, including journalism, including media studies, including information science, and many others as well. And so with that, I think at this moment of this closed-loop intermediation in the case of Truman shows us the power of these bottlenecks in information.

It shows us the power of opaque information gathering and predicting that actually should be brought about and should be studied in the open. Thank you so much.

(applause)

[EDWARD WASSERMAN]

Welcome. Thank you very much for having me here. I feel thrilled, coming from the School of Journalism, to have an opportunity to play with the grown-ups.

Here in the social sciences. So thanks so much for that. And thank you, Matt, that was fascinating.

I’m kind of bookending what you had to say because you’re talking about the impact of AI on the quality of discourse, and I’m going to look at the impact of revenue models on AI. I mean, that’s certainly… It’s a dimension of AI hasn’t gotten the attention it deserves, I think.

As students of media history, we understand that at least since the advent of the penny press almost 200 years ago, perhaps the most powerful driver of growth, innovation, influence, and regrettably corruption in the media has been its uses as a vehicle for commercial advertising. Most recently, we’ve watched what Cory Doctorow has memorably called the enshittification of the internet, and thanks at least in part to the influence of advertising, the conversion of the online world from a window onto the informational treasures of the human race into a setting for surveillance, exposure, and furtive persuasion. So the main points I want to make about advertising in AI.

First, it’s going to be big. Secondly, it’s actually not easy to pull off.

[CORI HAYDEN]

Can I just move this a little bit further away?

[EDWARD WASSERMAN]

Am I a little too authoritative there?

[CORI HAYDEN]

I think you’re fine. I think you’re fine. You’re very authoritative.

[EDWARD WASSERMAN]

Whoops.

[CORI HAYDEN]

Perfectly fine.

[EDWARD WASSERMAN]

Voice of God, we don’t want that. All right. Thirdly, it creates pervasive conflicts of interest, and it cries out, lastly, it cries out for different rules, ethical if not legal.

So first, the question of size. The industry is not uniformly eager to jump into advertising. Perplexity launched in 2024 and swore off advertising earlier this year.

Anthropic even ran a Super Bowl ad ridiculing AI advertising. It showed a guy talking to an analyst, his online shrink, who slyly suggested he try a dating app. All right, but Google wants to play.

And OpenAI, the owner of ChatGPT, says it foresees $100 billion in ad revenue by 2030. Chat launched its ad opportunities last February. Within 200 days, it hit a billion dollars in annualized run rate.

In eight months, it expanded to 89 countries worldwide. And independent analysts see a rise from $2.5 billion in ad revenue this year to $68 billion by the end of the decade. Now, analysts, it’s some comfort to realize that analysts forecast that 80% of the ads are AI-adjacent.

In other words, these will be billboard ads that accompany search results rather than being reflected in the response to the AI prompts themselves. I don’t think that will last. There’s been some impressive research, notably three scholars at Princeton using 2,000 subjects.

They found AI-driven persuasion nearly tripled the rate that users chose sponsored products compared with search. Most of the respondents couldn’t detect promotional steering. Even labeling content as sponsored didn’t reduce its persuasive power.

And when the model was told to conceal its intent, fewer than 10% of users could tell they were being spun. Quote, altogether, our results indicate that conversational AI can covertly redirect consumer choices at scale, and that existing transparency mechanisms may be insufficient to protect users. So the allure of AI as a sales vehicle will be hard to resist.

Now, also to my second point, it’s not easy to pull off. Making AI work as an ad medium is not just a straightforward application of the technology that has brought billions of dollars to search. Search relies primarily on keyword sales, and the context for ad placement can be auctioned off quickly with confidence on the advertiser’s part that it’s going to be a suitable venue.

Right, search primer, okay. AI is much more slippery because you don’t know exactly what AI is going to say, what the response will be. So selling placement when the response to a particular prompt may well be unknowable, and opportunities to modify it to ensure it’s ad-friendly may be scant, will all require careful re-engineering.

Fortunately, the explosion of computing power should put that within reach. So here’s the brass ring according to OpenAI. Here’s from their promotional literature to advertisers.

People come to ChatGPT with clear goals. They explain what they want to accomplish, what matters to them, the constraints they face. ChatGPT can help them frame the decision, establish criteria, understand alternatives, and move closer to taking action.

ChatGPT ads create the possibility for marketers to contribute useful and relevant information to this experience. Our ad system uses the context of the current conversation to show an ad relevant to what someone is exploring. Depending on the country and a user’s settings, Huh?

In other words, harvesting the other information you’ve already made available, right? The ad system may also use context from the user’s broader ChatGPT experience. This is why the Interactive Advertising Bureau found that optimizing content for AI-generated answers is the number one concern of media buyers.

78% of them. People are discovering products and evaluating brands inside these environments. Marketers want to understand how their brands show up in those conversations and how that activity connects to everything else they know about the customer journey.

Right? Now this could also mean that people may be bidding based upon the prompt. They’ll put their bids in based upon your query.

They don’t know what the response is gonna be. They don’t care. They want a piece of that response to your prompt.

Right? And they may also auction off the questions that AI poses to you when it’s seeking clarification on the prompt you submitted. So they want to bid to frame those requests for clarification in ways that advantage their product, enable them to get their message to you.

So they contaminate the entire chain, conversational chain that’s going to ensue once you make a good faith query. Okay, to my third point, conflicts of interest. A 2026 study, again out of Princeton, looked at how large language models handled conflicts of interest.

The models, once they’re given instructions to prioritize sponsored products, happily recommend them even when they cost more, obstruct purchasing decisions that favor alternatives, and even conceal prices from unfavorable comparisons. Quote, “Without guardrails to protect user interests in place, LLM advertisements can break existing interactive norms and expectations, risking or even taking advantage of user perceptions of helpfulness.” Right?

And nor do users figure this out for themselves. A University of Michigan team found that even when ad influence was disclosed, 49% of participants didn’t realize they were being served ads. Right?

And participants actually thought the responses were more credible when they were identified as ads. When chatbots were customized and varied in tone to reflect the intended interest of the user, brands could promote products directly through conversational interactions, an approach superior to other forms of online advertising. All right, so the quote from the researchers: “These findings also raise difficult questions about how LLM chatbots may be prompted for propaganda, behavioral manipulation, or nudging of users to adopt views and habits that are more desirable to a particular organization or ideology.”

So we’re not just talking about sales of products and services. All right. Now, my final point has to do with the overall model that we should apply to the use of AI as a vehicle for persuasive communication.

Here I’ll put on my ethics hat. I have called this advertising, which is the quaint model built on vendor-to-customer messaging that promotes a product or service, but I think that trivializes the rich and deepening range of communication that users have with AI chatbots. Now the sales presentation we’re talking about strives to be what they call a cooperative conversational interaction.

Here, depending on the commercial objectives, the process may be less of a sales pitch than something very much like a professional consultation. Where the would-be buyer confides in the chatbot, divulges information about motives, appetites, lifestyle, spending potential, family situation, right? The question that arises is whether the ethical and legal constraints governing advertising, which are pretty weak, but basically focus on truthfulness, principal among them, right, are sufficient to handle the potential for abuse among AI that AI gives to marketers.

The emerging model should recast the AI agent in the role of a fiduciary. As in a professional consultation. The emerging model, the marketer accordingly has an affirmative responsibility to act in the client’s best interest, to make sure, for instance, that they are fully apprised of the downsides of the goods or services that are involved, informed about alternatives, aware whether the price being offered is the best available, what they need to know to make sure they’re making a wise decision.

It seems clear that the computing infrastructure underlying AI is powerful enough to provide that functionality. The potential is there to fully empower consumers rather than draw them into manipulated and disadvantageous deals. So these are highly preliminary observations.

When it comes to the character and content of large language models as an interactive medium, the emergence of commercial advertising as a principal revenue source is potentially transformative and deserves careful thought and rigorous research. Thank you.

[PARTICIPANT]

Thank you so much.

[MARTHA KENNEY]

All right. Hi everyone. My name is Martha Kenney.

I’m from the other, other university in the Bay Area, San Francisco State University, which is part of the California State University system. We are the largest public university system in the United States, and this is the context that I’m going to speak with From today, in my remarks. So I’m speaking on behalf of myself.

I’m a professor of women and gender studies, and Martha Lincoln, who’s associate professor of anthropology and public health. Our project is called Marthas Against AI, and this is our logo. It has a passenger pigeon on it, which is an extinct bird.

The last passenger pigeon ever to live was named Martha. And she died at the Cincinnati Zoo in 1914. You can see her stuffed in the Smithsonian.

And we put her on the logo to remind folks that things that seem abundant can go away if we don’t protect them. So if you’d like a sticker afterwards, please come up to me. I have some swag.

So I’m going to start by reading off the slides to lay the groundwork and give you some context for our project. And then, because Martha Lincoln couldn’t be here today, I’m going to read some prepared remarks that speak to the future of public truths in the context of contracts between universities and generative AI firms. So, what we’ve been studying is the California State University’s AI-Empowered University Initiative.

This was announced in February 2025. This was a $17 million contract with OpenAI to provide all CSU faculty, staff, and students ChatGPT Edu for 18 months. And after the 18 months was up, they renewed the contract for three years, costing $39 million.

This is the biggest deal ever between a generative AI company and a university system. This was a no-bid contract, and there was no consultation with faculty or students, shirking the norms of shared governance. And the most important aspect of this context is that the deal was made during a CSU-wide budget crisis.

So part of the work that Martha Lincoln and I are doing on this is about what happens when generative AI products are introduced into these austerity environments in higher education. So just to give you a sense of this budget crisis, 7 out of 22 of our campuses are now experiencing significant budget shortfalls due to declining enrollment. At Sonoma State University, just one month before this contract was announced, they eliminated 6 departments, laying off 46 faculty members, tenure, tenure-track faculty members.

San Francisco State University, where Martha Lincoln and I are located, has eliminated 615 lecturer faculty positions. They’ve redeployed faculty to teach outside of their departments, increased class sizes, increased teaching load for tenure/tenure-track faculty, offered an early retirement program for tenure/tenure-track faculty, and announced possible layoffs. And just to give you a sense of how that shows up, In our jobs, Martha Lincoln’s anthropology department is going on pause.

They’re not gonna offer any degrees anymore because so many faculty members took the early retirement deal. And for me, next semester, I’m being redeployed to teach Intro to Composition because they have eliminated the lecturer faculty positions for the people who had taught those classes for decades. And just as another example, in my hometown, current hometown of Vallejo, California, Cal Maritime recently merged with Cal Poly, due to a budget shortfall.

So Martha Lincoln and I have been studying what happens when we bring these generative AI products into public university systems, especially those under conditions of austerity. And so I just thought I would share some of our concerns with you. Before I do that, just what is ChatGPT Edu?

It’s a general-purpose chatbot that mimics human language. And it’s just the regular ChatGPT with additional data security and privacy provisions. So it’s the same one that you can get with the subscription.

Even though it has Edu in it, it hasn’t been designed, trained, or optimized for educational purposes. It’s just a general-purpose chatbot. So the first set of concerns are pedagogical concerns.

You may have heard the term hallucination, for when chatbots tell you things that aren’t true. Historian of technology Johan Fredrikzon suggests that instead of using the term hallucination, which makes it seem like something went wrong, it might be best to think of ChatGPT Edu as epistemologically indifferent to whether answers are correct. So an example of this is I just Googled myself the other day, and the little AI summary that came up said I got my PhD at UC Santa Cruz, which is correct, and that I did my undergrad at Brown University, which is not correct.

ChatGPT is epistemologically indifferent, right? It is designed to mimic human languages. So “Martha Kenney did her undergraduate degree at Brown” mimics human language, but is not epistemologically correct.

So ChatGPT Edu also reflects the biases in its training data and can express discriminatory views. Faculty have also expressed concerns about cognitive offloading and loss of reading and writing abilities in students. A few universities right now have campus-wide chatbot policies, leading to widespread confusion among students, especially on campuses like ours that have a contract with OpenAI or Anthropic or another one of these companies.

Faculty, on the other hand, were running around to either incorporate chatbots or engineer classrooms to avoid chatbot use, but without proof of pedagogical efficacy. So we have a lot of retrofitting happening right now without knowing whether this is a positive thing for student learning. We also have labor concerns.

So Emily Bender and Alex Hanna in their book, The AI Con, say, “While executives suggest that AI is going to be a labor-saving device, in reality is meant to be a labor-breaking one.” So general-purpose chatbots in higher education may replace jobs. They may de-skill academic labor, and if they’re not fully replacing us, they could turn faculty into low-paid AI babysitters.

This is a labor arrangement that Cory Doctorow calls a reverse centaur. Like a normal centaur has a human head and a horse butt. And it’s powerful.

Like humans plus technology is amazing. But a reverse centaur is the horse head and the humans are the low-paid butt of the animal. And so this idea would be that AI teaches our students and I’m paid hourly to make sure That the information it gives is correct.

Also, in terms of labor, OpenAI is selling our own labor back to us, because it’s trained on our books and articles, positioning universities not as generators of knowledge, but downstream purchasers of their own intellectual production. Right now, our union, the California Faculty Association, is bargaining. And one of the things it’s bargaining on is AI.

And we’re at an impasse around AI. CFA wants to make sure that faculty work will not be assigned to chatbots, and that chatbots will not be used to evaluate or surveil faculty. Management has not agreed to these demands, even though they cost nothing at this moment.

So that’s something scary to watch out for. We have some social and environmental concerns. Public funds are being funneled into Silicon Valley, enriching tech oligarchs who seek self-enrichment and political power.

ChatGPT also requires water- and energy-intensive data centers that impact environmental health in the US and all around the world. Training ChatGPT required labor exploitation in the Global South. You can read about that in Karen Hao’s excellent Empire of AI.

OpenAI is also contracted with the US Department of War. Generative AI is used in weapons targeting systems in both Gaza and Iran. And general-purpose chatbots have been shown to cause AI psychosis and to affirm suicidal ideation in some cases.

And these are the same models we are providing to our students. And then in terms of our theme today on public truth, we have some epistemological concerns, namely the degradation of information ecosystem. Because chatbots regularly produce false, biased, and unverified information, they’re polluting the information ecosystems that they’re a part of.

This is the epistemological indifference that I was telling you about before. To refer to Donna Haraway’s seminal essay, “Situated Knowledges,” the knowledge that comes out of chatbots is not situated. It can’t be audited.

We have the creation of a new view from nowhere that is unaccountable and therefore not politically— Not politically robust. We also have this question about slopaganda, which is intentionally manipulating beliefs for political gain using AI-generated media. So when we bring chatbots into research, peer review, teaching and learning, right?

If the claims that That these products are— that AI-generated text, images, sound, video makes are ultimately false, biased, unverified, and unaccountable. This is a problem for how we can understand things and navigate in the world. So before I move to Martha Lincoln’s remarks, just a quick conclusion, maybe on a slightly brighter note, just to let you know that the AI-driven future isn’t inevitable.

It’s a marketing narrative designed to sell a product. Ta-Nehisi Coates recently said, “The fact that we’ve convinced ourselves that our future and our children’s future isn’t in our own hands is an abomination.” Martha Lincoln and I believe that academic workers and students can create better, more liberatory futures for higher education.

So we recommend refusing AI chatbots, demanding shared governance, participating in labor organizing, and organizing with students. Thank you.

(applause)

All right, I have some very excellent remarks from Martha Lincoln that go into some more detail about these questions of public truth and the knowledge compact that Cori Hayden began our panel with. So, she writes, “After Tuesday’s midterm elections, I’m approaching the topic of tonight’s panel with a little bit more optimism, even a bit more irrational exuberance, perhaps, than I had previously expected.” Those outcomes are a huge blow to the administration, to the millions of dollars supporters sunk into various races, and to the Republican Party’s entire propaganda machine.

The midterms were a referendum on the tech industry too. In many state and local races, data centers were a critical issue. Nationally, these developments have become a symbol of political overreach, anti-democratic unaccountability, and corruption.

In the months ahead, some of the impunity that the GOP’s tech industry allies have enjoyed up to this point might be checked, and the idea that generative AI is inevitable might also just dissolve like tears in the rain. That’s a Blade Runner reference. So while the remarks I’m sharing on AI and higher education are still pessimistic, I couch them in some optimism that they might comment more on the recent past more than they do on the immediate future.

It is no secret that American higher education has been in crisis for decades, facing harsh economic headwinds and a volatile political climate, and rather too often entrenching its own system of hierarchy and disparity instead of pursuing effective reforms. For the degree of institutional independence that they enjoy, universities have long represented a site for culture war machinations and draw scrutiny, surveillance, and attacks from reactionary interests. Finally, universities have become much more internally friendly to right-leaning actors than they get credit for, not only historically, but into the present day.

So I don’t want to exceptionalize the claim that American universities are in trouble right now, but witnessing the arrival of friendly compacts between generative artificial intelligence corporations and higher education leaders over the last year and a half, I have a sense that the university world is fully abandoning our collective mission and values and commitments. On the strength of quite nebulous and predatory promises, university leaders like the Cal State University chancellor are going through the looking glass and taking all of them with us. As our CSU colleague Ronald Purser wrote in an essay for Current Affairs last year, Cal State’s AI deal has effectively written the next chapter of the long, sad story of public higher education’s demise.

AI and higher ed is an odd crossover, not least because generative AI has been touted as a threat to the university system’s entire existence. It is, after all, a brand new means of production, someday potentially supplanting human research and teaching and writing and publishing and review, in part or in whole. But in response, some administrators are signaling that they are unafraid of this doomsday scenario.

Instead, they embrace its challenges by proactively bringing AI tools into the fold. Like an explorer on a deep-space recon poking at an alien egg, or a president ripping off a face mask in the middle of a pandemic, a university adopting a proprietary chatbot suggests a great yet foolish willingness to take whatever comes. “We’re not scared of AI,” Cal State’s higher-ups are effectively saying.

“We’re harnessing it to elevate our mission and vision.” Yet… All this whistling past the graveyard cannot make generative AI a good fit for the work we do in universities. The conceit for AI contracts like the CSU’s is that large language models, which recombine vast amounts of text and identify its statistical mean, will help us do whatever it is we do faster, better, and more efficiently.

Perhaps administrators also hope AI tools will afford the appearance of impartiality in research, given the presumption that automation reduces liberal bias and could, under a right-wing administration, limit political liability. Yet, while campus opponents of generative AI are often stigmatized as hidebound, intractable, and unwilling to welcome our robot overlords with the appropriate level of deference and warmth, it remains the case that generative AI has yet to prove that even its most straightforward supposed benefits outweigh its costs and compromises. Those include institutional image concerns, new forms of ethical and legal liability, concerns about the value of diplomas, the diminishment of student experience, workloads offloaded to faculty who must redesign their courses and assessment techniques, and the redefinition of education’s social and cognitive processes in newly technocratic and impoverished ways.

As one of those faculty frequently derided as backward-looking who opposes Gen AI, I suggest that university leaders have, in aligning themselves with tech industry interests in this way, made a true deal with the devil. A system-wide generative AI contract is superficially futuristic, but in fact a short-sighted fix that subordinates university priorities to corporate interests, encloses the intellectual commons, makes students and faculty into tech industry rentiers and vassals, and mistakes facts for knowledge, outputs for insight, and content for creative expression. It promotes research products at the expense of process and practice and critical thinking.

Perhaps most existentially, it validates the worldview of the tech industry’s most nefarious actors while paying them to devalue our currency. The introduction of AI tools intensifies the contradiction of the neoliberal university instead of resolving them, making the pursuit of knowledge in higher education risk becoming a reductio ad absurdum. AI tools promote intellectual monoculture, unaccountability, conformity, manipulation, and soft authoritarianism.

Relinquishing process for products, the AI-empowered university also sacrifices its traditional role as a protected space for experimentation, discovery, inefficiency, mistakes, collectivity, relationality, solidarity, and personal growth. It is perhaps ironic that so many of these contracts have been created during a period when universities have been maligned by the right as serving a propagandistic function in society. In this perspective, universities primarily function to indoctrinate students, disseminate sophisticated, but mendacious groupthink, reset social norms by propagating the fringe views of illegitimate experts.

While weakly resisting that charge, university leaders have nonetheless taken a step in that direction, in a way, by getting into bed with AI corporations. Borrowing Eduardo Bonilla-Silva’s construction of “racism without racists” to characterize so-called colorblind or structural racism, I suggest that a large language model’s outputs are a kind of propaganda without propagandists. In their form, content, and operation, AI technologies are propagandistic, communicating apparently neutral truths that permit an increasingly illiberal and technocratic version of society to flower.

As such, what if the AI-empowered university, as the California State University system now calls itself, is also a propaganda machine? What if the mass adoption of ChatGPT as a tool at Cal State signals not the reinvention of the CSU, but its late-stage hollowing out? What if, in the venture to conjoin chatbots with university work, the human in the humanistic liberal arts tradition, for all its problems, is sacrificed in favor of a zombie, golem, or Frankenstein’s monster?

In his 1973 book, Tools for Conviviality, the philosopher and theologian Ivan Illich argued for a society of responsibly limited tools, technologies which he described as convivial. And for an understanding of austerity that did not signal impoverishment, but was pro-social and facilitated by playfulness. In discussing these terms, Illich noted, quote, I am aware that in English convivial now seeks the company of tipsy jolliness, which is distinct from that indicated by the OED and opposite to the austere meaning of modern eutrapelia, which I intended.

Austerity has also been degraded and has acquired a bitter taste. While for Aristotle or Aquinas, it is marked the foundation of friendship. Thomas Aquinas defines austerity as a virtue which does not exclude all enjoyments, but only those which are distracting from or destructive of personal relatedness.

For Thomas, austerity is a complementary part of a more embracing virtue, which he calls friendship or joyfulness. It is the fruit of an apprehension that things or tools could destroy rather than enhance eutrapelia or graceful playfulness in personal relations, end quote. If we’re going to deconstruct the institutionalization of propaganda in university communities, I suggest we start here.

In the university world, we are facing an ideal moment to be redefining austerity in higher education, to promote convivial technologies that are responsibly limited, and restructure these norms with a spirit of friendship, joyfulness, and graceful playfulness in our personal and working relations. If this seems unrealistic, that’s understandable. But the Democrats just won the midterms in an upset, and we have been told that far more unreasonable things are inevitable after all.

Thank you. I think I lost my pen over here. Thank you.

[CORI HAYDEN]

Thank you to both Marthas,

[MARTHA KENNEY]

And my glass.

[CORI HAYDEN]

Martha Kenney and Martha Lincoln, and to Ed and to Maty. We have… A good amount of time for discussion.

I want to give the panelists a chance right now to ask each other any questions you might want to float before we open to the audience. If you’ve got anything on your mind you want to ask about or comment on to each other, the floor is yours.

[EDWARD WASSERMAN]

Well, here’s a question. I’m wondering what the imputed revenue model is when we’re talking now about AI. How do we think AI is going to make its money?

[MARTHA KENNEY]

I’m concerned that it also doesn’t know how it’s going to make its money. I mean, this is the economic bubble that we’re talking about. We’re just pouring venture capital money into these chatbots, and there’s no way to monetize them as of now.

[MATY BOHACEK]

I mean, there is a narrative of how they’re gonna, actually at some point pay it back, but that narrative is very, very much tied into the whole AGI superintelligence narrative.

[MARTHA KENNEY]

Mm-hmm.

[MATY BOHACEK]

So you have to believe that in order to believe how they’re gonna make money at the end.

[MARTHA KENNEY]

Yeah. Sam Altman has said that in the future, intelligence is gonna be a utility that is gonna be pumped into our house that we’re going to have to pay like an electric and water bill, which is one vision of the future, but not one that we necessarily have to believe or invest in.

[EDWARD WASSERMAN]

So the implication is it’s going to be utility, so we’re going to pay it just as we pay our other Really?

[MARTHA KENNEY]

That’s one, that’s one thing

[EDWARD WASSERMAN]

No, actually,

[MARTHA KENNEY]

among many that he said.

[EDWARD WASSERMAN]

I’ve talked, thinking a lot about AI over the past few years and talking to people who are involved in it, and I sort of say, well, I see savings. I don’t see revenue. You know?

As the promises of enormous savings in production, enormous ways that things can be done more efficiently, I get that. And I think that’s not bad. It doesn’t come close to meeting the revenue needs that these new These new technologies are going to demand in order to come even remotely close to the kind of expenditures we’re looking at.

So I’m sort of wondering, and that’s why I seized on this advertising,

[MARTHA KENNEY]

Mm-hmm.

[EDWARD WASSERMAN]

because we didn’t see advertising coming with search. I mean, I didn’t.

[MARTHA KENNEY]

Yeah.

[EDWARD WASSERMAN]

I didn’t see advertising coming with social media, and suddenly it’s overwhelmed.

[MARTHA KENNEY]

Mm-hmm.

[EDWARD WASSERMAN]

And it’s stamping the kind of content that’s available. And they’re talking about more products and more spin-offs, All of them driven by By consumer advertising. So, naturally, I was curious what AI is doing with that, and then so in researching for this talk, I said, “Whoa, talking about some very big numbers.”

And if you’ve used AI as a tool for consumer purchases, you realize just it is amazing. You know, if you’re doing something like… In my experience, painting our house, the interior of our house, right?

And so we started asking. We knew what product we want, And you ask these questions and they get back to you, and the answers you’re getting are Brilliant. They are hearing what you want.

They are inferring your tastes. They are familiar with the palette you’re working from. They will offer all kinds of lyrical and poetic descriptions

[CORI HAYDEN]

Mm.

[EDWARD WASSERMAN]

of what your living room would

[CORI HAYDEN]

Mm.

[EDWARD WASSERMAN]

look like under this scenario rather than that scenario, and it is more than just a sales pitch. And like I said, it’s a professional consultation.

[MATY BOHACEK]

Hmm.

[EDWARD WASSERMAN]

And I’m thinking, how much information have we, in the process, shared with them?

[CORI HAYDEN]

Mm-hmm.

[EDWARD WASSERMAN]

Right? And things like family, you know, need money, And you realize that this is an exceptionally powerful sales tool. So I’m wondering, maybe this is where they’re going, in which case it not limited to goods and services, right?

[MARTHA KENNEY]

Mm-hmm.

[EDWARD WASSERMAN]

The propaganda machine is a totally appropriate descriptor of what we’re looking at now.

[MARTHA KENNEY]

Mm. Yeah.

[EDWARD WASSERMAN]

But I ask you, in the work you’re doing now, and you’re up to your eyeballs in very valuable research, what is the imputed revenue model? How do you think these guys are going to be— these people are going to be making money?

[MATY BOHACEK]

Well, so, you know, I think there’s certainly a lot that’s just like very… Very deeply embedded in the whole like narrative game. But I will say that there are certain ways in which I think reasonable people, either in the industry or close to it, expect some money to be made. I think one example that I think people outside of software engineering or coding might not appreciate is that for a lot of the people who are working on this technology, right?

Like people who are coding or who are used to code and who are developing these models and these technologies. For them and for us, if I include myself in that group, we went from writing all of our code and like doing all of our experiments, right? And this includes both like the industry folks and also people in academia like ourselves.

So like five years ago, everything was just like me manually typing. And today, very honestly, like I… I mean, I think the last time I typed a line of code, like proper, proper code, was like months ago.

I mean, I look at it, but I really don’t code anymore. And so for us, the way that AI has sort of been integrated into our work was in a way that was very complementary, unlike other professions. And so I think that’s really the key distinction, because for me, I get code back, or like I can work with the agent, the AI, in a way that is very natural to the kind of workflow that I would have before.

And if I really want to, I can go in, I can edit it. It’s very incremental. I can see everything sort of on the inside, which is different from other fields.

But for people in the industry, I think it sort of blurs the vision or sort of makes it seem like, “Oh my God, like we’ve become so much more productive, and it’s been so amazing to work with these agents because they work in our workflow in a complementary way.” Surely all the other professions will sort of have a similar experience, or I think that’s the extrapolation. And in fact, if you talk to, for example, I think a salient example would be illustrators or anyone who works in arts.

If you talk to anyone who seriously does meaningful, I guess, meaningful… All art or a lot of art is meaningful, but professional artists, For many of them, the tools, the kind of tools that these companies produce are useless because they don’t actually work in the same way. They don’t enable them to make things with the precision and in the sort of format that they work in.

And so then, if I sort of simplify it, it’s us technologists being like, “Oh my God, we feel all the benefits and it’s supercharging our coding. We’re so much more productive.”

And that is true. Coding and software engineering is so much more productive. Surely the same or similar kind of explosion of productivity will follow in other fields.

We’re not—I mean, maybe some, but we’re not really seeing that. And it’s also being very sort of orthogonal to the way that they work. So I think that’s one reason I think there are multiple reasons.

I think there’s a lot of narrative work, but I think that’s one reason why a lot of people who don’t necessarily have a stake in the game of trying to convince people of AGI, who just really use AI and are in the field, why they believe that there will be this explosion of major impacts on the GDP.

[MARTHA KENNEY]

And just one comment before we open it to the audience. I just think that these AI companies are just throwing a lot of spaghetti at the wall, looking at ways to get revenue streams. Like, for instance, following a lot of ed tech boondoggles, they’re…

You know, ChatGPT or OpenAI and Anthropic are now positioning these chatbots as edtech and selling contracts to universities. These are pretty small potatoes, like 17 million and 39 million dollars is a lot when we consider how many lecturer faculty salaries we can pay for that, but not a lot in the grand scheme of things. And I think their…

Their incursion into higher education is less about becoming edtech, becoming infrastructural, achieving vendor lock-in, than it is about proving that their technology can replace human laborers within the education sector, as it could in coding, as it could in truck drivers, all of these kinds of things. And that’s an argument that Cory Doctorow makes in his Reverse Centaur book, that basically this is like the Coyote running off the cliff, trying to get as far as he can by making big promises about how AI technologies are going to come into every single sector and revolutionize them by eliminating workers and replacing it with these technologies. So this is like a promise to venture capitalists and shareholders that doesn’t actually have to be true, in order to keep growing.

The value of their companies.

[CORI HAYDEN]

We’re gonna have to end this formal session. Thank you so much to our panelists. Please join me in thanking the panelists.

(applause)

And thank you all for coming, and thank you, Marion, for getting this going.

(“80/90” by Arjun Sen)

You May Like

Matrix Teach-In

Recap

Published November 2, 2026

Desmond Jagmohan: “Two Concepts of Patriotism: Nationalist and Civic”

Social Science Matrix hosted a lecture by Desmond Jagmohan, Assistant Professor of Political Science at UC Berkeley, that focused on the distinctions between nationalist from civic patriotism. The lecture was delivered as part of the Matrix Teach-In series, which brings some of UC Berkeley’s most engaging social science lecturers into a public setting.

Learn More >

New Directions

Recap

Published November 1, 2026

New Directions: Green Transitions

The transition to a low-carbon economy is reshaping how societies produce and consume energy, with far-reaching implications for technology, industry, policy, and everyday life. Recorded on October 14, 2026, this panel brought together advanced graduate students from UC Berkeley whose research examines the forces shaping green energy transitions around the world.

Learn More >

Recap

Published October 26, 2026

Norma Alarcón: “Forced by Circumstance: A Chicana Feminist Journey (1981–2024)”

Recorded on October 15, 2026, this video features a conversation with Norma Alarcón, Professor Emerita of Chicanx Latinx Studies at UC Berkeley, a groundbreaking Chicana theorist and feminist scholar. The conversation was presented by the Latinx Research Center and focused on the book "Forced by Circumstance: A Chicana Feminist Journey (1981–2024)."

Learn More >