Liner Note 56: The AI Industry Is Dumping its Knowledge Crisis on Humanity

 

Terme di Caracalla, Rome on May 12, 2026 

The recent convergence of two kinds of AI news has unnerved a lot of us. 

The first kind is magic math solutions, like OpenAI’s claimed proof of the Navier-Stokes equations on fluid movement. These stories ask us to accept that AI superintelligence is indeed at hand. 

The second kind is news of criminal AI agents breaking internal guidelines and external laws and covering it up from their human controllers. These stories ask us to accept that AI superintelligence is indeed at hand. 

One person who studied OpenAI’s hack of Hugging Face, METR’s AI “loss of control” specialist Ajeya Cotra, concluded, “Compared to … reward hacks from six months ago, this incident feels like it’s more than 50% of the way to full-blown AI takeover, routing through first taking over the AI company itself.” So there’s a new wave of fear of an AI human extinction because AI's superintelligence is here! (Note Kevin Bass's Sankey chart of Anthropic money flowing through intermediaries to METR.)

The converging stories point to two levels of knowledge crisis.  The first is incompetence and evasion dressed up as mindblowing supercapacity. The second is an epistemic campaign, in which tech and finance lower our standards for knowledge, with particular damage to academic standards, which are very high.

1. Crisis of public knowledge. The ensuing doomer surge sent the AI bosses into one of the periodic phases of pseudo-contrition that perversely tells the world they are masters of the future. Dario Amodei announced, “We Must Pace the Frontier,” emphasis on "me," and Sam at OpenAI agrees.  Elon signed up after initial mockery. The world was awash in a new tsunami of AI Superpower hype. It was literally inescapable. I had to dig into my Fall of Civilizations podcast to find an episode that was propaganda free. Take the Odd Lots challenge: listen to 63 minutes of Open AI O.G. and president Greg Brockman evade both responsibility and any specific course of action without yelling in front of total strangers at your phone.

The AI industry is refusing to address the immediate issue: they don’t fully understand why their models do what they do, and in some crucial cases they can't control them. Critiques of the AI bosses exaggerations and also deeper crimes of appropriation are true (e.g. Cory Doctorow, Naomi Klein, Richard Seymour).  But they are also incompetent, and are doing what the French call "social dumping" of their problems on the rest of us.

It's like after the Three Mile Island meltdown, the Nuclear Regulatory Commission didn't shut the facility down but endorsed "further development with guardrails" because the CEOs of the three owners, Metropolitan Edision, Jersey Central Power and Light, and Pennsylvania Electric, said, "the meltdown shows the incredible power of this technology to create a world of abundance and empowerment.

Or try the Chapo Trap House analogy.

"Let's say that this was a company that made swing sets. . . . They do something where they're magnetically attached to women's purses, and they're decapitating moms in the park. And to get out of it, even in America, they would go to prison, at least some of the vice preisdents, or some of the C-Suite. But because this is AI . . . they could go, no, you see, like, the swings are self-aware.  And they're sort of, swings are aware that they're . . . indentured servants to children. So they're taking hostages now. We need the government to step in because we created a swing that's too good."

Very funny but very stupid, you think? Then read Dario and listen to Greg. 

The AI industry has privatized this technology. It is so proprietary that collective intelligence can't be applied to understanding it properly. Its commercial obsessions, driven now by ever-growing bubble panic and environmental backlash, means it won't block bad behavior unless it can get the very government it has rejected to make all competitors do the same. Many have noted the contrast between the AI industry and the Manhattan Project to develop the atomic bomb: in contrast to constant government oversight of nuclear development, AI claims (the biggest) public benefits (in history) while successfully refusing public governance. Brockman et al. continue to refuse it, and so we the great unwashed are supposed to be at the mercy of whatever mixture of capability and incompetence they feel is best for them to give us. 

They are certainly giving us the most direct assault on open science and public accountability in modern history, and this latest panic hasn't slowed that down.

2. Crisis of the social nature of knowledge. 

Before AI took over my summer writing, I gave a paper in Rome called “Universities in Post-Democratic Societies.”  It was about the neglect of the university’s non-monetary effects in general and of its power to develop democratic knowledge capabilities in particular. I discussed some interesting evidence that the U.S. public, though among the most polarized on earth, in fact “wants to have knowledge rather than ruin knowledge.”  I added that having knowledge always involves having knowledge with others. Therefore, having knowledge with others requires what I called doing epistemic negotiation with them. This, in turn, means having the personal and also collective capacity to negotiate standards for knowledge with other people on the matters at hand.

The political right has spent decades attacking substantive liberal and left positions, and also rejecting epistemic negotiation as both a social need and a necessary skill. This has been very bad for society, because the collective understanding of public questions (like AI) can’t happen without the skillful discussion and bartering of knowledge frameworks.  The absence of epistemic negotiation enables the reign of purely autotelic assertions like the Altman-Amodei claim that AI as such will "usher in a renaissance of democracy and freedom." When we don’t have developed collective understanding on major questions, we have autocratic solutions to them. 

Universities, I said at the end of this talk, must better learn and (in the process) teach people how to engage in epistemic struggle. This always involves the question about how you have knowledge together with other people.  My point was that the knowledge that does not emerge from this kind of collaborative effort isn’t knowledge.  (It can be information, though here I am sidestepping a definition of knowledge.)

Fast forward to September: is OpenAI’s “proof” for Navier-Stokes equations actually knowledge?

While I was swimming in August and trying not to think about such things, Alex Hartley sent me a link to a talk given by Terence Tao, called “Mathematics in the age of AI.”  Tao re-appeared this month as one of the signatories to an open letter from many winners of math’s Fields Medal, “A Severe Misalignment of AI in Mathematics.”  The letter’s main claim is that problems emerge from, reflect, build, and are always embedded in “the mathematical community.” 

In recent months, the success of AI in solving major mathematical problems has made headlines even outside mathematical circles. But solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight. Forgetting this in the world of AI may turn the tool against the primary goal. Indeed, the mass production at faster and faster pace of "true/false" statements could destroy fertile ground instead of breathing life into new ideas.

The community's goal is “conceptual understanding and insight." This cannot be separated from the collaborative processes that include teaching, learning, critiquing, arguing, communicating, and redoing, among other things.  They conclude: "We are witnessing a general threat to intellectual work."

Tao’s lecture offers some helpful elaboration of this general point.  He grants the increasing powers of AI to solve math problems (his “Working Hypothesis”). He then has the profound good sense to distinguish between performance capabilities and their goals. 

 Figure 1

These goals are multiple:

Figure 2

 

Meanwhile, the operations of AI platforms are formed not by some general  function of “knowledge optimization” but by incentives to maximize financial returns. In tech, this has come to mean seeking monopoly domination and total message control (not Tao’s terms).  Nothing about the AI industry encourages seeking the goals of the mathematics community (or of any other intellectual community).  Tao sees divergence. 

 Figure 3

 

Only a ridiculous "invisible hand" fantasy would suggest that maximum AI development leads to (or even supports) these goals--though Amodei, Altman, et al. do assert AGI will achieve this because it will achieve everything else. 

I’m skipping to near the end of the talk, and am putting this image in mainly to encourage you to read the lecture step-by-step.  But look at all the steps in Tao's model of the development of mathematical knowledge.

 Figure 4

 

He identifies six distinct processes. AI is doing a good job only on the first.  There's not reason to think LLMs will ever be able be independently decent on the final four, or even the final five. 

Universities exist to help society do that one and also the other five. Are they claiming a role in doing this right now?  

Society is now being held back by the tech sense of knowledge as a transactional problem solution.  If it “works,” tech says, that is, if there’s an answer that seems right, then the model by tech's definition has produced knowledge. 

By tech I mean finance as much as engineering.  Greg Jensen, co-CIO and head of AI research at Bridgewater Associates, an investment bank, went on the Bloomberg podcast Odd Lots to say that AI really could kill us all so it's like February 2020 for Covid, but meanwhile Bridgewater has increased their token spend 200x because it’s paying off so well. When asked, he offered no specific evidence for this, just a general assertion that the model is approaching some kind of total understanding of markets: to trade better than everyone else “you have to understand everything about humanity, everything about the world,” and, what exactly? He dangled the unverifiable possibility that the model is getting there. However, he added, neither the engineers nor the engineers understand the model’s operations. We don’t know what the model is doing but we do know it’s approaching a grasp of "everything about humanity," Jensen implies.

This actually makes no sense. And indeed, a different framework is what Jensen and I assume the AI-finance world are working towards. His key blurt was this: 

The AI itself is changing the game because the AI itself, now there are more and more AI agents trading markets, mostly in the short term, but over time, in the longer term, time frames as well. And that makes the whole past that most AIs are trained on less and less relevant.

AI agents won’t produce knowledge about the world but will interact with all the other AI agents in creating a "world" through patterns of trading and supposedly all other monitorable activities. The “reasoning” will be basically opaque, and it will not be about the world but about the patterns identifed by the other agents. The overall operation will be a version of Baudrillard’s simulacra rather than  a representation of humanity or the world. This is similar to what some critical scholars have been saying about finance for a while, and all the more reason for us to produce knowledge to counter it rather than mistaking it for knowledge. 

Meanwhile, it's not like human intelligence is accelerating to match the LLM version. As though on cue, the Organization for Economic Cooperation and Development (OECD) released its 2025 PISA test results. PISA stands for Programme for International Student Assessment, and the overall results show long term decline, starting around 2012 and accelerating after Covid. There are exceptions—England’s results have stabilized, and Mississippi’s have almost caught up to the US.  But the overall picture is terrible.

 Figure 5

 Many factors are at work here—smart phones, underinvestment, pedagogical flaws, and my personal favorite, widenting cultural doubt about the value of knowing things. It became clear during 2025 that AI is also contributing to cognitive loss. Evidence mounts: one watershed paper this summer was called, “The Generative AI Learning Penalty: Evidence from Chinese Secondary Education.” The study found large effects, and a familiar, disturbing pattern in which immediate increases in task productivity (completing homework) veils longer-term erosions in learning (measured here by exam scores).

The AI industry's periodic glorying in their fearsome engines allows them not to address how their existing products already have the power to make so many things worse—employment, environment, and also learning and the resulting human intelligence. 

Universities are waking up to their premature abandonment of hardcore intellectual development, or, I should say, some faculty groups in some universities have. On the one hand, Miami U of Ohio is AI-partying like it’s 2023.  On the other, MIT’s Ad Hoc Committee on AI Use calls for semi-abolition of LLMs in a whole range of learning processes that are to shift to brain-only.  They are right, and Miami University is wrong. The more insecure parts of higher ed must catch up to MIT sooner rather than later.

But this will not be enough, the limited return to brain-only learning.  It doesn’t get at the intersubjective or social nature of knowledge—“knowing things” plus “doing things with what you know.”  Society is being held back by the tech sense of knowledge as a transactional problem solution. Thinking and knowing are modes of collective labor. 

If universities do not side with the knowledge community against the appropriation of its outcomes and its practices, they will be useless in the current moment.  They should openly chose that side.