2026 strikes me as the year in which AI’s fiercest critics began writing their critiques with the help of artificial intelligence. I notice because they use it clumsily and barely polish the output, though I could be wrong. The impression, however, has set me thinking about how perceptions of AI’s usefulness have shifted over the years.

A glance at the statistics on so-called Shadow AI makes the phenomenon unsurprising. According to a 2025 UpGuard report, over 80 per cent of workers say they use AI tools that have not been authorised by their employer, and roughly half do so on a regular basis; senior executives turn out to be even more inclined than others to use them on the sly.

This clandestine adoption marks the end of a phase in public debate. For years a fairly widespread form of denialism held that language models did not work, that they produced only hallucinations, that they were by definition useless or harmful. The position may have had a certain logic two years ago, when one squinted with some myopia at still rough-hewn tools; by now it has become simply untenable.

The intellectual roots of this denialism are by now well known. The “stochastic parrots” thesis, formulated by Emily Bender and Timnit Gebru in 2021, inspired a legion convinced that language models were mere manipulators of linguistic appearance, devoid of understanding. In The AI Con (2025), co-authored with Alex Hanna, Bender consolidates this position into a systematic argument: the entire AI industry is a fraud, a con orchestrated by Big Tech to sell a product that does not do what it promises. The authors propose renaming language models “synthetic text extruding machines,” an image that likens the process to the industrial production of plastic.

The book has merits that would be unfair to ignore. The chapters on labour and on the precarity induced by automation offer a well-documented analysis of how AI is being used to fragment creative professions and drive down the cost of intellectual work. The analysis of public services is perhaps the book’s strongest section: Bender and Hanna show how AI hype functions as ideological cover for further cuts to services, and how the promise of the robot-doctor or the robot-teacher both presupposes and normalises the lack of adequate human resources. Their critique of AI catastrophism is also persuasive, with its analysis of the TESCREAL ideological complex and of how catastrophists and enthusiasts end up being two sides of the same coin.

The problem is that The AI Con does not confine itself to these sectoral analyses; one of its load-bearing arguments is that language models do not work in any cognitively relevant sense. The authors write that “what is currently being developed as ‘AI’ does not work, nor is it helpful, for an overwhelmingly large portion of people living on the earth today, especially people in the Majority World”. Consistent with this stance, in a lecture at Harvey Mudd College in November 2024, Bender reiterated that when the output of an LLM happens to be correct it is so by pure chance, and that one might as well consult a Magic 8 Ball.

Yet the book itself undermines this premise. Bender and Hanna devote hundreds of pages to documenting the concrete harm caused by the very tools they declare useless: British authors, translators, and illustrators who have lost work because of generative AI; film studios replacing concept artists with Midjourney; newsrooms sacking editors to fill pages with synthetic text; Clarkesworld magazine forced to suspend submissions because it was swamped with stories generated by ChatGPT. The entire third chapter presupposes on every page that the technology works well enough to replace qualified workers in the eyes of those who commission the work.

This is a phenomenon worth verifying, because it is often difficult to tell when what is really going on is “AI washing,” that is, layoffs that employers attribute to AI in order to deflect responsibility. Anyway, if the output were truly garbage, literally nobody would want it.

When it comes to denying cognitive value, the authors declare these tools “inherently unreliable, being designed to make shit up”. When it comes to documenting harm, they implicitly presuppose that the models work well enough to be adopted at industrial scale, to take commissions away from professionals, to pollute the information ecosystem. The two claims require incompatible premises; the book alternates between them according to rhetorical convenience without ever acknowledging the contradiction.

The authors’ implicit answer seems to be that the problem is us: we project meaning where there is none, we let ourselves be fooled by the superficial fluency of synthetic text. But if AI does not work and we are simply too credulous to notice, the problem is less AI than the fact that the entire species is cognitively inadequate—which, among other things, makes it rather pointless to write a book to warn it of the danger. If, on the other hand, we admit that at least some users are capable of distinguishing useful output from useless output, and that they do so regularly by selecting, discarding, and correcting, then the usefulness of the models is an empirical fact and the thesis that they “do not work” collapses.

The chapter on art and science is the weakest in the book. Bender and Hanna define image generators as “probabilistic (aka ‘stochastic’) algorithms trained on piles of work stolen from creative people” and reduce the entire question to the formula of the “three Cs”—credit, consent, compensation. The question of artists’ rights is undoubtedly important (and predates AI), but the chapter does not discuss it; it merely asserts it, as if theft exhausted everything there is to say about a technology that also opens genuinely new creative possibilities. There is no distinction between predatory uses and exploratory ones, between the replacement of creative labour and the amplification of artistic practices already under way.

The gap is all the more serious because it ignores how ancient and recognisable the pattern is. Baudelaire’s dread before photography sounds virtually identical to Bender and Hanna’s dread before generative models: the ease of production multiplies mediocre works, the new industry destroys whatever remains of the divine in the human spirit. The same objections have been levelled at movable type, at lithography, at cinema, at the web. Baudelaire’s error was not in diagnosing the multiplication of mediocrity but in believing that the new instrument could not give rise to works of art. We know perfectly well that a camera does not make a photographer, just as a prompt does not make an artist.

The authors write that “if text and image synthesis tools can write or draw something that is both plausible-seeming and would be technically difficult for people, that passes for creativity. But these tools do not speak to the human condition”. The argument is identical to the one once made against photography: the machine cannot reach the impalpable. But the avant-gardes of the twentieth century long ago untethered art from artisanal skill or direct authorial intervention: from Duchamp’s readymade to Warhol’s silkscreens, from conceptual art to Manzoni’s Merda d’artista, the canon has repeatedly expanded to include means and procedures that, at the moment of their entry, seemed unworthy. The authors’ criticisms would carry more weight if they had engaged with the work of the many creative professionals and artists who actually use AI.

Kate Crawford has traced a different but convergent trajectory. Her essay Eating the Future: The Metabolic Logic of AI Slop, published on e-flux in September 2025, describes AI as a metabolic system that devours culture in order to reproduce it in degraded form. The metaphor is effective, and the infrastructural concerns she cites—energy and water consumption, the displacement of externalities onto disadvantaged communities—are legitimate. But the prophecy of self-destruction clashes with the most recent research: a study by Gerstgrasser et al. showed that model collapse occurs when synthetic data entirely replaces real data, but can be avoided if data accumulates rather than substitutes. Other work (Dohmatob et al., ICLR 2025) is more pessimistic. The question is open and under active investigation, which makes a prophecy of inevitable collapse at the very least premature.

As I have argued elsewhere, the notion that AI by its very nature produces a cultural residue devoid of value falls apart the moment one looks at history. The bulk of human production has always been slop. The canon we revere is the tip of an immense iceberg of forgotten, derivative, or simply boring creations. Every time a tool becomes accessible, production multiplies—and mediocrity with it. But from mass mediocrity, every single time, new forms of excellence have emerged. Crawford, like Bender, applies to AI a moral standard that she does not apply to other technologies. Nearly all of our technologies rest on analogous systems of extraction, consumption, and pollution, often on an enormously larger scale; yet we rarely subject intercontinental flight, video streaming, or 3D rendering to the same critical rigour.

The conclusion of The AI Con reveals the structural limit of the entire approach. Bender and Hanna propose “strategic refusal” as a political horizon: saying no to AI, drawing inspiration from Luddite movements and feminist struggles. The invitation is suggestive, but the question left unanswered is: refusal in favour of what? If AI is a technology that, as the book itself documents, is already reshaping work, healthcare, education, and cultural production, refusal without an alternative project risks being a symbolic gesture that leaves the field open to those who govern and deploy these technologies. The authors envision specific, narrow-scope tools designed in collaboration with communities; but this vision sets a very low bar and fails to reckon with the possibility that general-purpose language models have already produced widespread, imperfect, and improvable benefits—benefits that a politics of outright refusal would end up surrendering entirely to those with fewer scruples.

The position I continue to hold is that these tools have concrete potential, and that precisely for this reason the politically relevant battle concerns who controls them and how they are governed. The denialism of the early period has delayed this discussion, sustained by the illusion that we are dealing with nothing more than an ephemeral and overrated fad. It is time to abandon what Benjamin Bratton has called the first stage of grief towards AI: denial.

This means demanding public, open, and open-source artificial intelligences whose code and weights are inspectable and modifiable by the public at large; it means international regulation of military use, transparency about energy consumption, targeted taxation of tech companies, real protections for workers, and educational pathways that forestall privileged fast lanes. We should also criticise AI by using AI, without question, but without hiding it any longer.

Sources.

Bender, E. M., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021). “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21), 610–623.

Bender, E. M. & Hanna, A. (2025). The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want. Berkeley: University of California Press.

Bender, E. M. (2024). Lecture at Harvey Mudd College, November 2024.

Bratton, B. (2024). “The Five Stages of AI Grief.” Noema, 20 June 2024. https://www.noemamag.com/the-five-stages-of-ai-grief/

Crawford, K. (2025). “Eating the Future: The Metabolic Logic of AI Slop.” e-flux, September 2025.

D’Isa, F. (2025). “The Idea of ‘AI Slop’ Is Slop.” The Philosophical Salon / Los Angeles Review of Books, December 2025.

Dohmatob, E., Feng, Y., Yang, P., Charton, F. & Kempe, J. (2025). “Strong Model Collapse.” Proceedings of the International Conference on Learning Representations (ICLR 2025). arXiv:2410.04840.

Gerstgrasser, M., Schaeffer, R., Dey, A., Rafailov, R., Sleight, H., Hughes, J., Korbak, T., Agrawal, R., Pai, D., Gromov, A., Roberts, D. A., Yang, D., Donoho, D. L. & Koyejo, S. (2024). “Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data.” Proceedings of the First Conference on Language Modeling (COLM 2024). arXiv:2404.01413.

UpGuard (2025). The State of Shadow AI Report 2025. https://content.upguard.com/hubfs/resources/The-State-Of-Shadow-AI-Report-2025.pdf