The Anti-AI Trap
Janet Vertesi gave a spicy talk at Princeton the other day, arguing that many academics working on AI Ethics or responsible AI are ignoring the political project of AI. Her point, as I understand it, is that world-building project: a network of firms, investors, infrastructures, regulators, researchers, and institutions being assembled around a particular vision of the future. The talk is associated with a new paper, where alongside her co-authors boyd, Taylor, and Shestakofsky, Vertesi argues that many of the debates we’re having in academia (what AI “really” is, whether it is inevitable, whether it will be disruptive, whether it is safe, whether it should be regulated) can distract us from the fact that this is a politico-economic project.
Interestingly, the wider public, unlike many academics, has understood this viscerally. There is a deep and growing distaste for AI worldwide (and, dare I say, especially in the US). Some of this stems from how AI has been sold. The dominant narrative from tech CEOs and their intellectual entourage is that AI will transform everything, automate jobs, dehumanize education, and replace creative labor. Unsurprisingly, the sales pitch has landed as a threat. And I think that this deep distaste is absolutely correct. Elites often deliver a future that is more dystopian than what is promised — so if their pitch is that people will be replaced and useless, what else should we expect?
Jasmine Sun usefully describes “AI populism” as “a worldview in which AI is viewed not only as a normal technology but as an elite political project to be resisted.” The term is very adequate to describe the many instances of AI resistance in communities all across the US, from “data center NIMBYism, AI witchhunts among creatives, and in the extreme, assassination attempts like what happened to Altman this week.” But this backlash also carries a danger: AI can become a convenient container for a broader social anger. I don’t think it is particularly effective to “rage against the AI” when people really are upset about precarious work, hollowed-out institutions, environmental degradation, etc. AI is now entangled with many of these problems, but as of writing, it usually plays a secondary role.
But the more important point concerns how AI populism typically interprets the phrase ‘normal technology.’ For the uninitiated, the “normal technology” reference here comes from Arvind Narayanan and Sayash Kapoor’s argument that AI should be understood as ultimately ordinary technology. In their view, AI should not be treated as a superhuman force outside society. AI’s effects will depend on applications, adoption, institutions, incentives, law, and organizational change. But I’d say that the folk understanding of AI populism misses the other half of Arvind and Sayash’s argument: that technology is kind of magical in its own way, as even “normal” technology can still be incredibly disruptive. Electricity, the automobile, and the internet have reorganized the world, and so can AI.
But the ranks of La Résistance (and the academics most attuned to it) too often default to mockery. AI is just “synthetic text extruding machines” and ridiculous poop analysis startups. Among the intelligentsia of the AI Resistance, there is too little serious engagement with the fact that the world’s most prominent mathematician is in real time, showing how AI will change mathematics forever. Or that agentic coding has changed how software developers work around the world, even though it only really started working last December. There is little acknowledgment that millions of people find LLMs profoundly useful.
And I’m sorry to break it to you, but AI is not a scam. There are scammy AI products, scammy AI startups, scammy AI evals, and scammy people trying to staple “AI” onto whatever they were already selling. But the underlying technology is real, and it will change the world.
A serious theory of resistance, therefore, cannot rest on pretending otherwise. You will not understand the impact of AI if you use only the free version of ChatGPT to call it “stupid”. If the current version of AI’s “Project” doesn’t sound appealing to society, I think it is our job to reshape it. But the current plot dismisses that there’s something fundamentally different about GenAI, in particular, LLMs. In the AI Con, Bender and Hanna “reject untestable ‘everything’ machines that are marketed as ‘general purpose technologies’.” And instead say that we need ‘specific tools geared towards specific tasks.’ This is opposed to the empirical reality of machine learning, sometimes referred to as “the bitter lesson.” And sure, we should be serious about validating systems in the application layer — but pretending generality itself is a scam will not help us govern the technology we actually have.
My concern is that if opposition follows this path, which in my view is blinded by this visceral distaste for the “Project of AI,” we risk doing little to help shape our future. This deprives us of having the right discussions about open-source models, government involvement (like in Switzerland!), democratic oversight, labor protection, and all things related to rethinking society in a world where radically disruptive technology exists—whether you like it or not.
Thanks Manoel for sharing your views! I enjoyed reading about your perspective, and I'm glad also to have learned about Vertesi et al.'s interesting paper via your blog post.
I agree that it can be unhelpful for "Responsible AI" broadly to center on a position that AI is a "scam" or completely ineffectual. AI is effective at some tasks (like coding) while ineffective at others.
However, I also noted that you mentioned disagreement with some specific authors and arguments, as well as with "La Résistance (and the academics most attuned to it)", the "opposition", "intelligentsia of the AI Resistance", "the current plot". I was a bit confused about the throughline between your specific critiques of specific arguments against AI leading to a general criticism of the "resistance" or the "opposition".
For example, is it clear that the "opposition" or "La Résistance" adheres to what you attribute to be the "current plot" or the arguments you specifically contested (of the AI Con or "normal technology")? It's not clear to me that the "opposition" is coherent in general, and is at best a collection of wildly different distinct voices and contingents, certainly where may be repeating talking points.
I think you're responding broadly to patterns in the discourse where some voices (especially with platform) share their positions that you disagree with, that you mentioned.
However, I found it confusing who was the intended "you" mentioned in the post, and I wonder if more clarity around the specific object of your critiques would be helpful. I'm also not really sure what "this path" refers to or "opposition following this path refers to", is it specifically the idea that AI is a "normal technology" no different than any other?
I'd love to sign up to join the intelligentsia of the AI Resistance of which you speak :)
Lastly, I agree with you that the job is to reshape the project of AI.
But I wanted to go back to where you started on the post, and make a pitch that strategic awareness of the politico-economic project of AI is needed to reshape the project of AI (and different from the "AI is a scam" arguments you critiqued).
I think that the important things you mention at the end about open-source models, democratic oversight, government involvement, are all 100% politico-economic infrastructure. I'd go so far as to say that the current politico-economic project of AI is exactly why we don't have these nice things you mentioned. The current politico-economic organization is so reliant on venture capital's hype cycle and selling a vision of necessary scaling that doesn't match up with the reality of what's needed to achieve impact on the ground. These nice things (democratic oversight) are against the entrenched interest of big AI developers.
To be a bit more concrete, I think there's space for a "normcore" and technology-centered La Resistance that also recognizes the politico-economic project of AI. I would argue we can start with a couple premises:
- AI models are good or really good at some things, but they're bad at some things too.
- AI companies have huge compute costs, and need to sell hard to convince folks they could eventually cover those costs. The compute and energy costs are so intensive (because the big companies have huge valuations and they don't face any pressure to become more compute-efficient) that this in turn affects broader swathes of the economy, including giving various CEOs cover to do mass layoffs as a short-term myopic thing.
From there, a measured perspective acknowledges that both these things are simultaneously true. For example:
- AI models are not good at everything out of the box. But a lot of companies need the world to believe that they are good at everything and magical, in order to make enough money to cover heavy compute costs, and so people stop protesting their data center builds.
- There's a lot of ways to achieve important impact with AI. Some are more efficient than others. Some ways are PR for companies. Companies starting leaderboards to reward employees for using the most AI is not the most efficient way to achieve impact with AI, and is probably so that they can brag about "innovation".
More importantly - I'm guessing here - but I would peg the modal academic in responsible AI, who is not explicitly a critical theorist, somewhere in this middle ground rather than at these extremes ("AI is a scam"). I'd actually love to see a survey or a more comprehensive poll of what people are thinking.
Thank you for sharing your thoughtful perspective!
I agree that more careful strategizing is needed to realize responsible improvements from AI. I'd love to see more community discussion like yours hashing out these difficult but crucial questions for moving forward. I think it's even more helpful for us all to be super precise about what exactly we disagree with; it's the only way out of the current schisms.