For some reason, British universities don’t know how to end their academic year. As a result, I’ve found myself giving four papers at UK academic conferences in the last half of June, which should never happen, and at least one more in the 2^nd week of July. The summer World Cup will end before the British academic year does. British academia forgets that summer gives people time to stop the old thoughts so that they can have new ones. ISRF is no better: we’re launching a (very good) book, Generative Justice, on 16th July.

Luckily, other countries of the northern hemisphere have more sense. The 2025-26 academic year has been over for weeks in the US, Canada, Mexico, France, etc., or will end soon in the knowledge that it won’t start again until late October (Germany). So those academics are in a position to assess what happened to AI’s reputation over the course of their year.

ISRF’s research group, “Reframing AI,” is finishing a Green Paper this summer on the impact of Large Language Models (LLMs) on university-level learning and writing. It’s not easy: some AI trends are from Venus and others are from Mars. Here’s my short list:

1. The Rise of “Collapse”   

The things that AI can collapse have been multiplying. There’s model collapse, retrieval collapse, quality collapse, argument collapse, semantic collapse, cognitive collapse, and knowledge collapse.  A group that modelled knowledge collapse found an AI-use equilibrium in which short-term benefits destroyed the collaborative intellectual practices that “sustain long-run collective knowledge.” Other papers could have said “research collapse” as they discuss the enshittification of academic publishing through the infiltration of AI-generated papers and assessments. Outside of AI spin-world and capital expenditure rationalisations, the discourse is increasingly of unwanted decline with a future cliff.

2. The Personal Assistant Rapture Has Arrived

The models keep getting better at tedious admin, causing predictions of imminent superproductivity and lots of current relief. Many people are embedding PA uses into their daily lives.

For example, the editor of the Yale Review explains that, overwhelmed with conflicting duties one day, she turned to a chatbot for help.

Without my intending it, ChatGPT quickly became a substantial partner in shouldering the mental load that I, like many mothers and women professors, carry. “Easing invisible labour” doesn’t show up on the university pages that tout the wonders of A.I., but it may be one of the more humane applications. Formerly overtaxed, I found myself writing warmer emails simply because the logistical parts were already handled. I had time to add a joke, a question, to be me again. Using A.I. to power through my to-do lists made me want to write more. It left me with hours — and energy — where I used to feel drained.

My sense is that people love this PA function, which they have on a decent level of automation for the first time. This contrasts with their feelings about LLM’s uses for counterfeiting art, music, film, novels, poems, and university papers, which people mostly hate.   

3. PA Rapture Can Bankrupt AI Companies  

We have now learned that AI companies lose massive amounts of money powering through your to-do list for £15 a month. The service costs vastly more than that, has no prospects for traditional “economies of scale,” and runs on a trillion-dollar data infrastructure that is miraculously uniting the left and right in opposition. 

When companies had to move from “all you can eat” subscriptions to token-use charges, some, like Uber, blew through their annual budgets by spring and started rationing. Nearly half of firms in one survey said “they had scaled back use of AI agents because costs outweighed the benefits.” As AI master critic Ed Zitron noted on a Chapo Trap House episode, “Let’s Lose Lots of Money,”

These companies were like, ‘oh shit, wait, we didn't make sure that this was useful. We just told everyone to use as much AI as possible.’ And then everyone did that, cost them a bunch of money. And now everyone's suddenly going, ‘wait, is there an ROI for this shit? Wait, wait, wait, wait, is [AI] good? I didn't check before.’

Zitron is now joined by the likes of the Bank of International Settlements (BIS), that sobriety house for global investors. BIS warned in their Annual Economic Report 2026 of an AI double bind (not their term) that if AI isn’t transformative then investors won’t be able to pay back their enormous debts, but if it is transformative, labour will be too impoverished to buy AI services (see, for example. Box C, p 20). 

4. From Universal Good to Controlled Substance 

Concerns with cost and collapse are leading users to define hard boundaries between good and bad uses. Many academic administrations haven’t caught up with this one and are still offering free token burn to every community member for everything. But instructors and students have been drawing bright lines. They were analysing usage issues from the very start, particularly regarding the role of LLMs in replacing active thinking for writing.

Within a few months of the launch of ChatGPT in November 2022, a classic statement had already been written by the Columbia University undergraduate and LLM user Owen Kichizo Terry:

As any former student knows, one of the main challenges of writing an essay is just thinking through the subject matter and coming up with a strong, debatable claim. With [ChatGPT], one snap of the fingers and almost zero brain activity, I suddenly had one.

The author of another mainstream media piece, from September 2023, used chatbots to write university admissions essays and concluded that if a student wanted their essay to make them stand out among thousands of applicants, they should “just write their own—chatbot free.”

These earlier critiques were often tentative when faced with a powerful industry’s grandiose claims (imminent superintelligence) and withering tech-shaming.

Gentle pleas to “consider writing your own essay” have been largely replaced by AI writing bans and citations of research on AI-induced cognitive offloading that leads to cognitive loss and brain rot—measurable reductions in capabilities. Newspapers like The Guardian are covering dangers to thinking (for example, April 2025, November 2025); I summarised some of last year’s work in “AI’s Two Pathways.”

These concerns continue to grow. One new paper analyses “epistemic risks—[AI-based] threats to humanity's collective capacity to know things accurately, reason well, form beliefs, and maintain a healthy information environment” (Yang et al., 2026). The swelling demand to contain LLM (ab)use also involves increased academic policing, as in the announcement that “Princeton Introduces Proctoring, Changing Century-Old Honor Code.”

Another watershed was announced by a Brookings Institution AI power user: “I’ve recently drawn a sharp line in the sand: no A.I. for writing.” The author, Rebecca Winthrop, referenced new studies showing chatbots can increase diversity of terminology even as they reduce diversity of ideas. She added, “research shows that A.I. has the largest homogenizing impact on students who are farthest from the mean and have unique perspectives, including neurodivergent students and those from racial and linguistic minorities.”

University managers, who are the dumb money that buys everything in every ed-tech wave including this one, now have on their hands a full-scale movement tailored towards “brain only” use. This is particularly the case for their writing professors. But the same problems affect maths instruction and computer science. As for primary and secondary schools, one long-time reporter has said, “I’ve never seen the kind of parent backlash about school tech that we’re seeing.” The same goes for resistance to sloppocalypse in popular culture. We’re starting to see something like the popular containment of AI on many fronts.

5. Chaos on Job Loss

Economists couldn’t disagree more on AI’s workplace effects. 

Everyone now seems to agree that jobs are bundles of tasks. Everyone agrees that some tasks will be “exposed” to AI replacement and others will not. Everyone agrees that AI, like other waves of automation, will “shift” a job’s “task content.” There is no agreement on which tasks will be exposed or how much and what the outcome for jobs will be. And there is no agreement on AI’s impact on productivity. 

I find Daron Acemoglu’s low estimates of AI's benefits to be convincing, and have explained why. I find the Erik Brynjolfsson group’s higher estimates less convincing, even though the FT is more likely to publish Stanford enthusiasm than MIT scepticism. One of my reasons is that Acemoglu does not assume that the typical worker is deficient and needs to be improved. Another is that Acemoglu does not invent big productivity boosts when agentic AI “reshape[s] the core human competencies, shifting from information-focused skills to interpersonal ones.” In other words, the sceptics are more parsimonious with the evidence.

That said, jobs are destroyed by business executives not academic debates, and executives seem bent on destroying entry-level jobs. The new interim report on “Young People and Work” finds that the share of young people in the UK who are not in education, employment or training (NEET) is the second highest in Europe. Economists do agree that starter jobs are being crushed, while the high-level skills that “complement” AI automation appear at risk.

My hunch is that executives will do more damage to their workforces in the continuing quest for elusive AI gains. Consultancies will tell them that they must: the failure to find big returns in tech disruption so far, they’ll say, means executives haven’t been disruptive enough of their people. Noting that 40% of AI projects produce improvements of 0% to 10%, Bain and Company are selling “deliberate investment in role redesign, new ways of working, and change management.” 

Gains are likely to remain elusive, but all the more reason to reorganise people’s workflows every few months and reengineer Brynjolfsson’s “core human competencies” while sacking some solid percentage and ploughing the wages thus saved into bigger AI capex. This is a chaos default worth fighting.

6. AI Derangement Syndrome 

A giant slice of humanity’s intellectual bandwidth has been hijacked to talk about AI. Much of it, as science and technology scholar Alondra Nelson has been explaining, is in the domain of agnotology, the deliberate creation of ignorance, as tobacco companies did with research on smoking. We have major issues to study and solve: unemployment, war, world poverty, climate change, racism, inequality, water shortages, fossil fuel dependence, the decline of universities, toxic masculinity, political polarisation, global democratic backsliding, cognitive decline, and migration phobia.

Though it’s not all AI’s fault, public discourse on these issues is stuck in low gear.  While we wait for AI to fix everything, we could spend  $93 billion per year to end world hunger, say half a trillion dollars by 2030. Instead, AI “hyperscalers” will spend $5 trillion on AI infrastructure by 2030—ten times the bill to end hunger.  

The job sectors I care about most, academia and the arts, have been two of the most dislocated, suffering job degradation or loss, process corruption (like writing), and the disparagement of their human skill—as well as non-investment in the new instructors, researchers, and artists that expanded, non-slop human production will require. AI is helping to keep major problems unsolved, in a pecking order for which no one has voted.

7. Hints of a Military Bailout

Silicon Valley was built by the US military before and after the Second World War. Its consumer popularity in the PC, internet, and smartphone eras may have been the aberration, as it becomes increasingly disliked while reverting to “move fast and kill things.” 

The biggest AI companies, Anthropic and OpenAI, are embedded in the military, as are Amazon, Google, Oracle, and the other hyperscalers. The dramatic break-up and make-up between Anthropic and the Department of War, OpenAI’s offer to step in, the constant warnings about the China threat, the proposed increase of the US military budget from under $1 trillion to $1.5 trillion next year, the drumbeat of strategic forebodings about the super-capabilities of Glasswing, Mythos, and other models with which only special entities can be trusted: all this may be the run-up for a public bailout of the industry via the military (equity stakes, debt purchases, direct funding, etc.) in the name of national security. Ed Zitron and others don’t think that AI is too big (or too important) to fail. My guess is that the industry has already made itself exactly that.

I wrote that paragraph on the 1^st of July.  On the 2^nd, the Financial Times reported that “OpenAI proposes handing Trump administration 5% stake.” Sam Altman is trialling his bailout, with the US government as the investor with the biggest pockets.

AI advocates would like you to believe that the total integration and unfettered use of AI is inevitable. Yet the cracks in this vision are clear. We need to stay in democratic charge of this technology.  We’ll need to organise together if this is to have a chance of happening.