All walkthroughs · Paper briefs

Why Is AI So Contentious?

Martin Beraja (UC Berkeley) and Noam Yuchtman (London School of Economics)
Brookings Papers on Economic Activity, Fall 2026 conference draft (draft dated 9 September 2026; posted 23 September; presented 25 September 2026 in the Panel on AI and Policy)

Paper · Brief and rating · Dashboard entry

About 21 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of Why Is AI So Contentious?, by Martin Beraja of the University of California, Berkeley and Noam Yuchtman of the London School of Economics. It is a conference draft for the Brookings Papers on Economic Activity, posted on the Brookings site on September twenty-third, twenty twenty-six, and presented two days later in the conference's panel on AI and policy. This episode is an AI-generated walkthrough, not an Unjournal evaluation or decision. It is a short essay, about ten pages plus two figures. It asks why a technology that promises large growth gains, and in which tens of millions of people hold a financial stake through index funds and pensions, attracts so much hostility. Their answer is that AI raises the two kinds of concern that earlier transformative technologies raised, economic and existential, plus a third, moral concern that has no clear precedent: the models were built from people's words, creations and expertise, taken without consent.
Discussant
So the claim is about the nature of the backlash, not its size. Can you state the thesis in their own words?
Host
Yes. In the conclusion they write: "We have argued that AI is so contentious — and the conflict over it so visceral — because, on top of the economic and existential concerns that every transformative technology raises, it raises a moral one: that it was built from people's words, creations, and expertise, taken without consent." Earlier, in the introduction, they put the contrast more sharply. They write that this "moral concern, not the return of familiar ones, is why AI is so contentious and the conflict so visceral." There is a second strand that is more economic. They argue the taking had an economic logic, the same one they say has driven innovation at what they call the resource frontier throughout history, and that some of it may even be desirable. And the policy conclusion is that the remedies most often discussed, universal basic income, retraining, or pauses on development, will be incomplete because they address the economic and existential concerns and not the moral one.
Discussant
Let's take the evidence. What do they show to establish that people are negative about AI in the first place?
Host
Figure one has three panels. Panel A uses the Pew Research Center's US surveys: around half of American adults say they are more concerned than excited about increased use of AI in daily life, and only ten percent are more excited than concerned, a share the authors say has fallen by nearly half over five years. Panels B and C use the Ada Lovelace Institute and Alan Turing Institute survey of attitudes in the UK: sixty-three percent of UK adults see large language models, the text-generating systems behind chatbots like ChatGPT, as beneficial, while forty-seven percent see them as concerning. So many people hold both views at once.
Discussant
None of that speaks to why people are negative. Is there anything in the paper that separates the moral concern from job fears or safety fears?
Host
No. There is no survey of their own, no decomposition of stated concerns, no regression, no model in this essay. The method is conceptual argument plus historical analogy, illustrated with examples. The analogies come in two layers. The first layer establishes that economic and existential concerns are familiar. For the economic one they cite displacement across occupations from drivers to paralegals and programmers, faster than institutions can adjust, and the concentration of gains, noting that Forbes counted forty-five new AI billionaires in twenty twenty-six. They point to the first papal social encyclical, Leo the Thirteenth's Rerum Novarum from eighteen ninety-one, which they quote on "the enormous fortunes of some few individuals, and the utter poverty of the masses." For the existential concern they use nuclear technology and John the Twenty-Third's Pacem in Terris from nineteen sixty-three, quoting its remedy: "Nuclear weapons must be banned."
Discussant
That establishes precedent for two concerns. What is the second layer?
Host
The second layer is the resource-frontier framework, which comes from a companion working paper by the same authors, Growth at the Resource Frontier, dated July twenty twenty-six and available as an NBER conference paper. The idea is this. Many big innovations bring a previously unused resource into economic use: mining and metallurgy for minerals, drilling for underground oil, radio for the electromagnetic spectrum. Before the innovation, nobody had bothered to define property rights over the resource, because it had no value. In the essay they write: "A distinctive feature of resources outside the frontier is that property rights — who owns them and how they can be used — are typically poorly defined." That gives the innovator a chance to claim the resource in anticipation of large rents. Counterparties, sometimes the state, then contest the claim once they see the value, and the dispute is resolved through litigation and regulation, often in the innovator's favor after the fact.
Discussant
And the mapping to AI is that online text, images and work practices were the unclaimed resource.
Host
Exactly that. They describe developers scraping essentially the entire accessible web, including Common Crawl, which is a large public archive of web pages, Wikipedia, Reddit, news sites, blogs and, as litigation later revealed, pirated book libraries. They distinguish three things that were taken: people's own words, people's creations such as books, photographs, music and code, and increasingly people's expertise. On expertise, they cite Uber drivers being used to train the self-driving systems meant to replace them, Meta asking employees to record their work for training before meeting resistance, and firms such as Mercor paying experts to teach models their jobs. They argue that before the insight that this material could be training data, it was abundant, cheap and ignored, and that copyright's fair-use doctrine was the main governing rule. No market for training rights existed, no price, and no shared idea of what licensing language for training would mean.
Discussant
Why didn't firms just contract? That is the standard Coasean question.
Host
They give three frictions. Innovators would be offering to pay for a resource with no market. They would either have to reveal details of their innovation, tipping their hand, or make offers the other side would find opaque. And they would have to negotiate repeatedly with many content producers and, for expertise, with millions of individual workers. They write: "Facing such contracting frictions, innovators decided to appropriate without contracts. None of it was done with consent." The essay then lists the conflict that followed: the New York Times suit against OpenAI and Microsoft in December twenty twenty-three, suits from actors, studios, Encyclopedia Britannica and Merriam-Webster, novelists and others, a consolidated multi-district case now in the Southern District of New York, and the Bartz versus Anthropic settlement of one and a half billion dollars, which they call the largest copyright settlement in history.
Discussant
On Bartz, my understanding is that the settlement turned on pirated books rather than training as such.
Host
The authors say so themselves, in a footnote: the primary legal basis was the use of training data from websites holding pirated copies of books, not violations of fair use generically. That is an honest caveat, and it limits how much the settlement tells us about the legal status of training on lawfully obtained material.
Discussant
Where does lobbying come in?
Host
It is the framework's second prediction. Appropriation must ultimately be ratified by the state, so appropriators invest in political influence. Figure two uses data from Issue One, a US advocacy group that tracks lobbying. Quarterly US federal lobbying spending by major technology firms roughly doubled between twenty twenty and twenty twenty-six, from about ten million dollars to about twenty million. In percentage terms the fastest growth in twenty twenty-six was at the model developers: Anthropic's spending quadrupled from twenty twenty-five to twenty twenty-six and OpenAI's nearly doubled. In the European Union they cite the Corporate Europe Observatory, which ranks technology as the largest lobbying sector, above energy and finance.
Discussant
That is a weak test. A company that grew from small to very large in two years would lobby more under any story, and these firms lobby on export controls, energy, safety rules and procurement as well as copyright.
Host
Agreed, and the essay does not attribute lobbying to data rights; it offers the figures as consistent with the prediction. Its EU discussion is mainly about data protection and privacy rules, a related property-rights question.
Discussant
Now the core move. Why is appropriation of data different from appropriation of oil or spectrum, which also produced intense fights?
Host
This is section three point two, titled The Original Sin. They concede that appropriation alone does not explain the tone. They write: "appropriation partly explains why there is conflict over AI; it does not explain why the concerns underpinning conflict are moral and the conflict itself visceral." Their evidence for the moral character is the vocabulary of the aggrieved. They write: "The vocabulary of this conflict is one of theft and violation rather than of adjustment and compensation." They also draw a contrast with past moral objections to technology, which concerned uses, like building a weapon or editing an embryo. In their words: "The moral objection to AI is about its inputs." And they cite Leo the Fourteenth's twenty twenty-six encyclical Magnifica Humanitas, on safeguarding the human person in the time of AI, as focused on dehumanization rather than distribution or survival.
Discussant
And why should the inputs carry moral weight that oil did not?
Host
Their answer turns on what was taken. Land, oil and spectrum can be bought, sold and paid for once rights are settled. Human creativity and expertise are, for many people, an expression of who they are. They contrast piracy with model training. A pirated book is still the author's book, and the pirate does not take the author's ability to write. They write: "What large language models did is not copying. They absorbed millions of works into a single system that retains no trace of who made what, and that learned from them to do what the makers do". What was taken, in their account, is authorship, and the belief that writing, drawing and knowing were things only a human could do. They illustrate with a long passage from the cartoonist Sarah Andersen in the New York Times, who describes seeing an AI image made in her style from her name as a prompt. She writes: "I felt violated." The essay ends the section by explaining the label. In Genesis the first transgression was taking from the tree of knowledge, a wrong whose consequences tradition calls fortunate, and the authors suggest AI may be similar, adding that "it was humans, not God, who were aggrieved."
Discussant
What is the normative position? It sounds like they are on both sides.
Host
They say so explicitly. They do not claim developers acted illegally; they say whether they did is unclear to them and is for the courts. But they argue the taking was not incidental. From the companion paper, they argue that poorly defined rights at the frontier are a reason to innovate, since the chance to claim a resource before its value is recognized is part of what makes pushing the frontier pay. They write: "Some degree of appropriation may therefore be the price of innovation at the frontier, and a society that forbids it altogether may find that the frontier stops moving." So they describe two mistakes. One is to ignore the taking and treat creators' grievances as confusion or rent-seeking. The other is to reject all appropriation and shut down frontier innovation. They write: "Each is half right." The proposed direction is private or public arrangements that restore to creators some measure of ownership and attribution.
Discussant
Let's turn to uptake. How has the paper been used or cited?
Host
It is very new, and we found little. It was posted on September twenty-third and we could not find it in OpenAlex, so there are no tracked citations. Brookings published a summary page. Its session, the Panel on AI and Policy, also included papers by Luis Garicano and Stijn Van Nieuwerburgh, and the program lists no formal discussants. We found no press coverage specific to this paper. One caution about our own materials: the dashboard brief links an Inside AI Policy story titled Brookings paper warns of AI backlash fueling future regulation. That story is from May twenty twenty-five and covers a commentary by Darrell West of Brookings. It should not be counted as attention to this paper, and the brief's line that the paper argues the backlash will shape future regulation fits West's argument better than this essay, which says resolution will come after the fact through lawsuits, licensing, legislation and new institutions. A search miss is uncertainty; BPEA papers often draw commentary over the following months.
Discussant
Is the framing of training as dispossession new to this essay?
Host
No. Computer scientists have written about training-data dispossession, for example Precel and coauthors in twenty twenty-four. The essay's contribution is to place the grievance inside an economic theory of frontier appropriation and to argue that the moral element is what distinguishes AI from earlier technologies.
Discussant
Who would make decisions differently because of this paper?
Host
Regulators weighing copyright, text-and-data-mining exceptions, collective licensing and provenance rules would hear that compensation alone may not end the conflict, and that attribution matters in its own right. Labs deciding whether to license data, offer opt-outs or build attribution into products would hear that the backlash is partly about consent and authorship. Funders of AI governance research would be pointed toward the moral channel. For an Unjournal evaluation, the relevant fact is that the claims are empirical in principle though the essay does not test them. Whether grievance over inputs explains a meaningful share of negative attitudes, relative to job and safety fears, can be checked with existing surveys and new survey experiments.
Discussant
Then let me give the critiques. First, the central claim is comparative and causal: that the moral concern is why AI is so contentious. The essay shows that some aggrieved parties use moral language. It does not show that the moral concern drives the aggregate attitudes in figure one. The Pew and UK surveys ask about a range of specific concerns. An evaluator should ask what share of respondents cite consent or use of creative work, and whether that share predicts concern after controlling for job exposure and safety views.
Host
That is the main gap. The UK survey does ask about specific concerns, so a first pass is possible with public data.
Discussant
Second, selection on the dependent variable. The vocabulary evidence comes from plaintiffs and a cartoonist, people who are by construction the most aggrieved. Litigants have strong incentives to frame a dispute as theft; that is what copyright claims are. Moral language from lawyers and creative professionals is not evidence about the median survey respondent, most of whom have not had their work scraped in any way they would notice.
Host
The authors might reply that expertise appropriation is spreading to ordinary workers, which would widen the aggrieved group over time.
Discussant
Possibly, but that is a prediction, and an evaluator should note it as one. Third, the novelty claim. The essay says past moral objections concerned uses, not inputs. There are counterexamples worth checking: objections to enclosure of common land, to the patenting of indigenous traditional knowledge and genetic resources, to the use of tissue from people like Henrietta Lacks, to photographic likeness without consent, and to sampling in music. These were partly about inputs taken from people. If those cases produced moral, visceral conflict too, AI may differ from them mainly in scale.
Host
The Luddites, whom the essay does not discuss, also framed their objections in moral terms about the degradation of craft.
Discussant
Fourth, the theoretical link. The companion paper's model says appropriation rents can incentivize frontier innovation. The essay leans on that to say some appropriation may be the price of progress. But the counterfactual is not shown for AI. Licensing deals between developers and publishers or platforms did appear once the value was clear. An evaluator should ask how much the breakthrough depended on unlicensed scraping at the margin, and whether a regime with opt-outs, collective licensing or statutory rates would have slowed it much.
Host
They do concede a counterfactual with contracting is imaginable and only argue it would have been hard.
Discussant
Fifth, the policy conclusion is stated more firmly than the evidence supports. The claim that universal basic income or retraining will at best be incomplete depends on the moral concern being large and independent of economic concerns. If the moral grievance is partly an expression of economic fear, which is plausible for illustrators and translators whose markets are shrinking, then compensation may address more of it than the authors allow. Conversely, if attribution is what people want, it would help to know whether attribution is technically feasible for models that, as the authors themselves say, retain no trace of who made what.
Discussant
So what should an evaluator actually do with this?
Host
Four checks. First, judge the argument as an essay, while flagging that the central causal claim is untested. Second, check the descriptive facts against the Pew, Ada Lovelace and Turing, Issue One and Corporate Europe Observatory sources, including what counts as a major technology firm and as lobbying spending; the draft shows these series without definitions or uncertainty. Third, look for survey evidence that separates concern about training-data consent from job and safety concerns. Fourth, test the novelty claim against the counterexamples we discussed. The expertise needed is a political economist, a specialist in public opinion about technology, and ideally an intellectual property scholar. The data are mostly public; the final BPEA version may add replication materials.
Discussant
My summary: a clearly written hypothesis from strong authors, with a useful theoretical frame from the companion paper, and an empirical claim that remains to be tested. Its value for decision-makers depends on whether the moral channel turns out to be large in the data, and the essay does not tell us that.

Sources