The AI Con, by Emily Bender and Alex Hanna

Reviewed by Justin Joque

Emily Bender and Alex Hanna, The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want (New York, NY, Harper Collins, 2025), xi + 274 pp.

Emily Bender and Alex Hanna’s recent book, The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want, provides a valuable and accessible overview of the harms of artificial intelligence and the nature of the speculative bubble fueling its rapid rise. The book covers the use of AI to shift employment to gig work; the replacement of social services with AI; its impacts on creative industries, journalism and science; and a chapter outlining the ways discourse around AI destroying the future actually help those developing and promoting the technology. The book concludes with a chapter outlining possible policy solutions as well as individual ways that people can organize against ubiquitous AI. In sum, the authors argue that there is too much hype around AI and they show all the harms hidden behind the unrealistic promise of a technology that will never achieve what its boosters claim. The book largely achieves these aims and as such provides a strong reference for the overarching problems with this technology.

The AI Con expands on earlier, influential critique of AI in two papers also co-authored by Emily Bender, “On the Dangers of Stochastic Parrots”^[1] and “Climbing towards NLU.”^[2] Bender’s critique of AI, in all of her work, depends on an intentionalist interpretation of language, wherein the primary role of language is to communicate an intent from one speaker to another who is able to understand that intent. According to such a theory, language produced by machine models is inherently meaningless as, according to proponents of this theory such as Bender and Hanna, a machine can have neither intent nor understanding of a human’s intent. One mark of this paper’s influence is how widespread similar assumptions are among the many critics of AI-produced language or art today, which leave almost unexplored the possibility of a critique of AI that would nonetheless start from structuralist or poststructuralist insights about the mutable relationship of language and subject or language and context.^[3] Moreover, this humanist critique of language machines tends to coincide, at least in Bender and Hanna’s writing, with a liberal critique of the political economy in which AI is embedded, as opposed to a left or Marxist critique that would call for more radical upheaval of the capitalist conditions driving automation and useless “innovation.”

In this vein, what is most interesting about Bender and Hanna’s new book is the overall structure of the argument, which is framed as a series of demystifications of a surface of superficial hype that obscures an underlying depth of complexity. The book is framed as a sort of how to guide: “How to resist the urge to be impressed, how to spot AI hype in the wild, and how to take back some ownership in our technological future” (20). The framing of hype implies that behind this surface of fascination, of impressive feats of automation, of Silicon Valley showmen who promise untold riches, this technology frequently fails to live up to the hype. In each of the areas Bender and Hanna focus on, large language models appear to succeed only on the surface of the problem they address, whether it is scientific research or the provision of social services, whereas the real problems they attempt to address have a depth and complexity that AI boosters readily poster over in order to make it to the next round of funding.

This seems an accurate assessment of the situation in these fields, but when the relationship of surface and depth appears again in their understanding of the technology and language it takes on a much more aporetic form. Here, that which hides behind human language (for them its ‘communicative intent’) must be both solid and real, but at the same time ephemeral and ungraspable. For if that which hides behind language were too real, were fully empirical, it would then allow itself to be computed and one of their central arguments in this book is that human language is fundamentally not computable. For them, no amount of data or statistical analysis can produce text that actually means something in the way that a human does when they speak.

In a very short section on the history of benchmarking in computing, Bender and Hanna mention early work at the Defense Advanced Research Projects Agency (DARPA), in particular work on machine translation and transcription. They outline how machine transcription can be tested against known audio inputs and text outputs to get the “word error rate” (i.e. what percent of words are transcribed incorrectly). With this example, there are two important considerations. First, that there exists a “ground truth” by which an algorithm can be measured, in this case the correctly transcribed text. Second, the creation of these benchmarks and ground truth against which systems are first trained and then later tested (usually two separate datasets or two parts of a dataset) necessarily reflect the social biases that influence their creation. These biases create a gap between the system’s performance on the test and in the real world, where, for example, a speaker may have an accent or speak in a dialect that was not included in the training data (81–82).

While of course these two considerations are not irreconcilable—and significant amounts of social theory have been developed to explicate the relationship between objectivity and social construction—they still express a tension that sits at the core of Bender and Hanna’s project. That is to say, it is important for them that behind words there exists an objective grounded reality, that words are fundamentally tied into what Leif Weatherby calls “the ladder of reference,” but also to maintain that all of this (language, systems, institutions, etc.) are shaped by a socially constructed reality, one that can only be fully accounted for and understood by a human.^[4] Language must be just real enough that its use can be objectively ‘wrong’ but also not too real that it can be described (and modeled) as a formal and computable system of reference.

So, it is noteworthy that they mention but quickly abandon the example of machine translation. In transcription, notwithstanding ambiguities which inevitably arise such as how to properly use punctuation, it is much easier to imagine the ground truth with which an output can be computationally compared. But, translation, with a significantly greater amount of freedom in what can constitute a ‘correct’ answer, makes it much more difficult to devise an objective metric against which to measure a given output. The reader—at the very least this reader—struggles to quickly imagine how one could algorithmically decide between right and wrong translations, especially given the flexibility in word order and synonym choice available to most translation.

It is clarifying to spend a moment outlining the solution to this problem that DARPA developed in the late 80s and early 90s as it elucidates some of what is at stake better than transcription. Their evaluation system was deeply dependent on a significant amount of human labor and involved two separate tests. The first provided the output of the translation system (alongside controls) to a set of educated adults who were unfamiliar with the source language. They read only the output and then were given an SAT type test of reading comprehension to evaluate how understandable the text was. The second evaluation was done by a team of skilled translators who knew both the source and output language. They graded the quality of the translations on a version of a standardized scale that was used by government agencies to evaluate human translators.^[5]

One can see here so much more clearly the latter point that Bender and Hanna would like to make: from the use of source material (which in the cited description above were all articles from the Wall Street Journal about business matters) to the use of government standards for evaluating translators, the evaluation was deeply enmeshed in a very specific cultural and institutional milieu. I would venture the reason it is an inadmissible example or at least one they choose to set aside is that to include translation requires doing away with the solid ground from which an LLM could be conclusively wrong; so, they instead turn toward transcription because it allows them to sidestep the problem of depths and surfaces that structures their project. Transcription extracts the true ground of text from out of audio, but translation moves simply from one surface to another—that is from one text to another text.

In an earlier article that Bender cowrote with Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” Bender and her coauthor there more explicitly grapple with the challenge that machine translation presents to their project. They state—after insisting that a machine can never understand language—that “If such models were to reach the accuracy of supervised translation models, this would seem to contradict our conclusion that meaning cannot be learned from form. A perhaps surprising consequence of our argument would then be that accurate machine translation does not actually require a system to understand the meaning of the source or target language sentence.”^[6] Translation here opens up the possibility that a system can successfully manipulate language only on the surface as the necessary reference is not to a “real world object” but simply another word in another language. To guarantee that success at this task cannot be used to argue that a machine has understood language then requires making understanding a metaphysical category that cannot be adjudicated by the performance of a system in the way the famous Turing test suggests.

We might then ask a similar set of questions about this relationship between depth and surface when Bender and Hanna discuss intention. For them an LLM can never capture what they call communicative intent, which requires a human to breathe meaning into language. “Simply modeling the distribution of words in text provides no access to meaning, nothing from which to deduce communicative intent. Language models thus represent nothing more than extensive information about what sets of words are similar and what words are likely to appear in what contexts” (30). But, let us imagine the now all too common activity of a person giving an LLM instructions for writing an email to a colleague. What would the relationship between that message and the intent of the one who prompted the LLM and sent the message be? Would it change depending on how closely they checked the LLMs work? Or, what if they even wrote an email and asked the LLM to make it more cordial? We quickly arrive at an e-mail of Theseus where it becomes very difficult to articulate the precise moment where communicative intent slips away and is replaced by “synthetic text” extruded by the machine.

While Bender and Hanna’s book offers a solid and helpful overview of the current state of AI and its uses, their metaphysical commitments to an unbreachable depth of human meaning and volition opposed to a machinic non-sense leaves their project ill equipped to address the harms they accurately identify. They end up left with a set of overwrought metaphors and liberal talking points that in an ironic way mirror the social averaging that critics rightly accuse LLMS of producing, what the artist Hito Steyerl calls “mean images.”^[7] For example, toward the end of the book as they discuss a future after the AI bubble pops, they state, “The residue of the bubble will be sticky, coating creative industries with a thick, sooty grime of limitless tech expansionism. This is the fallout of venture capitalists and tech entrepreneurs not pausing to think about who would be caught in the blast radius” (195)—a bubble, that is also a nuclear weapon, whose fallout is somehow a not thinking by the capitalist class or perhaps the fallout is the residue and is caused by the not thinking. Regardless, while a tortured metaphor is not itself an impeachable offense, and I am sure even in this short text I have allowed a few bad metaphors of my own, it speaks directly in this case to the slipperiness of language and the possibility that even humans can and often fail in their attempts to share a communicative intent.

There is a tendency among many liberal critics of AI, including Bender and Hanna, to focus too heavily on metaphysical concerns, arguing about fundamentally what AI is or is not capable of. Meanwhile, tech companies, large and small, invest increasing amounts of capital in the technology, force chatbots on users whether they want them or not, and hand over decision making to these systems. While Bender and Hanna accurately assess the harms of these technologies, their liberal humanist insistence on an ultimate ground truth behind it all and a need to return to it leaves us with an ineffective political program. What is needed instead is a much stronger, radical program that recognizes first and foremost the artificial nature of all language, institutions and knowledge production. The liberal position offers us the fantasy that if we just fix or do away with AI we can return to a more reasonable, less harmful form of capitalism. Against this we must instead recognize that capitalism has always been a crushing, extractionist artificial intelligence.

To my mind, the most telling pronouncement of the book is the moment when they declare: “We need the ability to compel the owners of the means of automation to tell us when we are viewing the outputs of a media synthesis system or [are] subject to consequential decisions that are due to an algorithm” (182). In this paraphrase of Engels’ 1877 Anti-Dühring, which would go on to become a radical watchword, we witness on both the level of form and content what is at stake in this play of surfaces and depths.^[8] On the level of content, we observe the abandonment of any radical position: the owners of the means of production are to be left their riches and social position and these technologies will be left in place but by law we will be alerted to their usage. On the level of form, we see the desire to reuse language; this citation without citation to a much more radical position which is abandoned but also spoken in clear reference. Thus, this reference reveals a poignant criticism of Bender and Hanna’s position performed almost automatically by the (non)citation to Engels, suggesting the question of why not simply seize the means of automation. We witness, in short, their communicative intent at odds with itself.

Despite all there is to commend their schematic analysis of everything that is wrong with AI, in attempting to reveal what is really going on behind the curtain of AI hype we risk mistaking our clear view of the situation as insight or worse yet as an effective political program. Within their project of dispelling hype, underneath all of the marketing slogans, there exists the hard kernel of truth—an ultimate inability of this technology to ‘work’ that will one day catch up with the con artists selling it. As such there is no need for a radical response, no real possibility of a different world, all that is to be done is to reveal what lies behind the hype then human nature and the market will take care of the rest. The bubble will pop; you and I, trained in identifying the hype, will be proven right. We may eventually have to scrape off the residue of the fallout of the unthinking, but beyond that everything will unfold as it should, perhaps even as though it were automatic.

1. Emily M. Bender, et al. “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, 2021, pp. 610–23, https://doi.org/10.1145/3442188.3445922. ↑

2. Emily M. Bender and Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (Association for Computational Linguistics, 2020). ↑

3. For a Derridean critique of AI, see: Víctor Betriu Yáñez, “Is a Derridean Critique of Generative AI Possible? Writing Machines and Logocentrism without Subject,” The OLR Supplement, Jan. 15, 2026, https://olrsupplement.com/2026/01/15/is-a-derridean-critique-of-generative-ai-possiblewriting-machines-and-logocentrism-without-subject/. ↑

4. Leif Weatherby, Language machines: cultural AI and the End of Remainder Humanism. (Minneapolis, University of Minnesota Press, 2025). ↑

5. John White, “The DARPA Methodology,” MT evaluation: basis for future directions. Proceedings of a workshop…2-3 November 1992, San Diego. Available: https://mt-archive.net/90/AMTA-1992-White.pdf ↑

6. Bender and Koller, “Climbing towards NLU,” 5193. ↑

7. Hito Steyerl, ‘Mean Images’, NLR 140/141, March–June 2023, DOI: doi.org/10.64590/uhm ↑

8. Friedrich Engles, Anti-Dühring: Herr Eugen Dühring’s Revolution In Science (Moscow, Foreign Language Publishing House, 1954). “The proletariat seizes political power and turns the means of production in the first instance into state property. But, in doing this, it abolishes itself as proletariat, abolishes all class distinctions and class antagonisms, abolishes also the state as state. […] The first act by virtue of which the state really constitutes itself the representative of the whole of society – the taking possession of the means of production in the name of society – this is, at the same time, its last independent act as a state” (388–89). ↑