AI Politics 13: What OpenAI Stopped Studying
connecting the 'New Deal' policy paper, the killed studies, and the lobbyist who wants the public to weigh in only on his terms
Three weeks ago, OpenAI published a thirteen-page policy document called Industrial Policy for the Intelligence Age: Ideas to Keep People First. It calls for higher corporate taxes, a 32-hour workweek, a public wealth fund partly funded by AI companies, automatic safety-net triggers when displacement metrics hit preset thresholds, and a tax base shifted from payroll toward capital gains. The framing is progressive. Fortune‘s writeup compared it to Sam Altman’s “New Deal” pitch. (FDR would have been proud, but as we said in Part 9, Sam Altman is not FDR.)
Outlook Business led with the 32-hour workweek as the headline. The substance, on closer inspection, is non-binding throughout — proposals expressed as “the government should consider,” “we recommend,” “could.” When asked which specific legislation OpenAI actually supports, the company’s spokesperson declined to provide examples. The document is a signal that the firm has thought about the political consequences of its own technology, not a commitment to act on the thinking.
Three days ago, the excellent Jasmine Sun reported in the New York Times that the same firm’s lobbyist hire — Chris Lehane, joined OpenAI April 2024 — has been quietly deprioritizing the research that would have produced data behind such a policy. Multiple sources, speaking to Sun on condition of anonymity to discuss internal deliberations, said that Lehane killed or sidelined research projects on the environmental impacts of AI, the gender gap in ChatGPT usage, the urban-rural divide in usage, how ChatGPT steers users’ career decisions, and long-run economic forecasting.
OpenAI proposes solutions. OpenAI’s lobbyist deprioritizes the studies that would tell you whether the proposed solutions are addressing the actual problems.
The killed-studies claim is the only anonymous-sourced item in an otherwise heavily-named-source piece — Sun spoke on the record with Lehane himself, with Steven Adler from OpenAI’s former safety team, with Tejal Patwardhan from OpenAI’s frontier evaluations team, with Anthropic Institute lead Jack Clark, with Brookings’ Molly Kinder, with former Biden NEC deputy director Bharat Ramamurti, and with David Shor’s Blue Rose Research data deck. The piece is built on three months of reporting. Lehane on the record offered a more flattering version of the same pattern: “We want to do applied physics, not theoretical physics.” He didn’t deny that some categories of internal research were being deprioritized. He characterized the deprioritization as a methodological preference. The off-record version Sun’s sources gave her is sharper: “We’re not going to release something about a problem until we have a solution for it.”
Either reading is consistent with the same posture. The firm decides which problems are publicly available for the public to weigh in on, and reserves the right to keep problems internal until it has a solution it’s prepared to defend.
Consider what each deprioritized study addressed, if only for a moment.
The environmental-impacts study would have produced internal OpenAI numbers on water consumption, electricity load, embodied carbon, and the marginal infrastructure footprint of a query at scale. Whatever those numbers showed, they would have been the firm’s own. The data center moratorium fights I mapped in Part 11 (Maine, Festus, Pennsylvania’s Lehigh Valley, Ravenna and Shalersville) are running on assumptions, utility filings, and consultant projections. State legislators in California, Ohio, Utah, Maryland, and West Virginia passed cost-shifting bills last year using the public-record evidence available to them, in part because the firms’ own internal data isn’t.
If you’re new here, Sacred Cow BBQ is where I try to make sense of the politics of AI (and higher ed, and a few other things that bug me) without pretending to know more than I do—because it’s such a weird liminal (ah, there’s that word again) time.
The gender-gap study would have measured who actually uses ChatGPT at scale, disaggregated by gender. Not who the marketing materials say uses it; who actually does. Economic-displacement projections that don’t disaggregate by gender will systematically misfire on which workers’ bargaining power is most at risk. Anthropic researchers are running a different version of this work, with results pending. OpenAI started one and decided to let it drop.
The urban-rural divide study would have mapped the political geography of AI adoption. Which counties are running ChatGPT inside their workflows; which counties are adopting it personally; which aren’t. The map matters because the moratorium-wave coalition is in some places and not others, and the politicization curve runs through that geography. Without OpenAI’s internal usage data, the political map of AI adoption is a guess. The firm has the data and chose not to publish it.
The career-steering study would have measured what ChatGPT actually recommends to young users asking about credentials, careers, and job applications. We don’t know what the study would have found. We do know that ChatGPT is one of the most heavily used recommendation systems in the world for exactly these decisions, and that Stanford HAI documented software-developer employment for ages 22 to 25 down 20 percent from the 2022 peak. The connective tissue between that labor-market shift and what the model is actually telling kids to study is what the deprioritized study would have illuminated. The firm publishes a policy paper proposing a public wealth fund to address displacement; the same firm chose not to study whether its own model is accelerating that displacement. Both can be true at once.
The long-run economic forecasting study is the most methodologically fraught of the five. Long-run economic forecasting on a three-year-old technology is genuinely premature, and reasonable researchers can disagree about whether it should have been undertaken at all. But the firm’s policy paper rests on assumptions about long-run displacement and growth. Producing internal forecasts and choosing not to publish them is a different decision than not producing forecasts. Sun’s reporting suggests the former.
That’s five deprioritized studies. Each addresses a question that’s politically alive. We don’t know what the studies would have found, and reasonable methodological objections apply to at least the long-run forecasting one. What we do know is that the firm built up research capacity in five politically inconvenient directions and then decided not to pursue any of them to public release. The pattern isn’t subtle…and, of course, it makes the social scientist in me die a little not knowing those answers.
Lehane’s rule about not releasing problems without solutions is a research-comms posture. As such, it’s also a policy stance. It decides, on the firm’s behalf, that the public can weigh in only on problems for which the firm has identified its preferred solution. The firm becomes editorial board for the data its own users and infrastructure produce.
Selective publication of internal industrial research isn’t unique to OpenAI. Pharmaceutical companies didn’t publish failed clinical trials for decades, until post-Vioxx reforms forced disclosure, and even now compliance is partial. Most large firms in politically consequential industries operate some version of an internal research-publication policy that prioritizes the firm’s interests. What’s different here is the political stakes — and the speed at which the technology is producing political effects the firm’s research could illuminate. The firm-as-editorial-board pattern is older than OpenAI. The firm-as-editorial-board pattern operating on a technology this consequential, with this much velocity, is something newer.
Two weeks before Sun’s piece ran, Lehane was publicly pushing back against AI-safety advocates in a different forum. After the April 10 firebombing of Sam Altman’s home, Lehane told the SF Standard that “some of the conversation out there is not necessarily responsible” and that “when you put some of those thoughts and ideas out there, they do have consequences.” You can read these as two reactive moves by a lobbyist managing two different incidents. You can also notice the shape of the response. Researchers writing about AI’s downsides are not responsible in public when their material reaches a bad-faith reader. Researchers writing about AI’s downsides are theoretical physics in private when their material reaches the firm’s policymaking. Whether the consistency reflects coherent strategy or reactive PR, the firm’s preferred speech environment looks the same in both directions.
Steven Adler, a former OpenAI safety-team employee, told Sun: “The A.I. industry is engaged in cutthroat competition over truly world-changing technology. Unless we change their incentives, we shouldn’t be surprised when companies cut corners, even if they’ve said the right things.”
The line does the analytical work. It also names the thing the policy paper can’t name. The Industrial Policy paper exists to describe the world in which OpenAI’s technology gets governed responsibly. The research suppression exists to make sure the public conversation about that governance happens on the firm’s preferred terrain. Both are real. Both serve the firm’s commercial interests. Neither one alone tells you what the firm is doing. Together they do.
The suppression operates on conversations already in motion. Each deprioritized study corresponds to a debate happening right now, conducted with whatever evidence the firms have chosen to make public. Local moratorium hearings in Maine, Pennsylvania, Ohio, and Missouri are running on assumptions and consultant projections, not internal usage data. The cost-shift bills already enacted in California, Ohio, Utah, Maryland, and West Virginia set their numerical thresholds based on utility filings and industry claims rather than the firms’ own research. Congressional hearings on AI displacement, where the Stanford HAI report shows industry witness share tripled to 37 percent while academic witness share shrank to 15 percent over the past decade, are conducted on testimony from the firms whose research operations are deciding which questions are worth pursuing. The Bores race in New York’s 12th district is fought partly as a proxy battle over what the firms know and aren’t sharing.
The studies, had they shipped, wouldn’t have settled any of these debates. They would have entered the record. That’s a different thing than determining outcomes — it’s setting the terms whoever else shows up has to argue against. A cost-shift threshold of 20 megawatts versus 50 megawatts is a different policy with different distributive consequences. The firm with the most-detailed internal data has a policy of not releasing it. That makes Lehane’s research-comms posture a quiet veto over what the policy conversation can include, wielded by the firm whose technology is producing the conversation in the first place.
Compare Anthropic. The firm publishes papers showing Claude users delegating their most personal decisions to the model and later regretting those decisions (”You made me do stupid things,” one user told Claude in research Anthropic released voluntarily in January). It publishes experiments showing junior engineers using Claude Code complete tasks not much faster but understanding the work less when quizzed afterward. Sun calls this the “light and shade” framing — bad results alongside good. Anthropic also has its own selective gating: Project Glasswing decides which fifty partner organizations get access to Mythos, the Mythos cyber-capability dispute with the Pentagon is being managed actively rather than disclosed openly, and the April 16 internal-employee-survey leak reached the public via leak, not voluntary release. Anthropic isn’t a transparency outlier in some pure sense. The firm’s brand is partly built on publishing bad results, which is itself a strategic posture.
Even allowing for that, Anthropic and OpenAI are visibly running different practices on the specific question of whether internal research that would embarrass the firm gets published. Anthropic publishes the user-regret paper and the engineer-skill-erosion paper voluntarily. OpenAI is reported to be killing the analogous studies before they ship. Two firms, same political environment, two different practices on this one specific dimension. The contrast doesn’t tell you which firm is more credible overall. It tells you which one will publish a research finding that’s bad for the firm.
Part 9 read OpenAI’s Industrial Policy framing as accelerationism with a New Deal coat of paint. Sun’s reporting is the closest thing to empirical confirmation we’ve had. The policy paper and the research suppression are both real, both produced by the same global-affairs apparatus, and the simultaneity is the data point — whether or not the two outputs were coordinated in any narrow sense. The Industrial Policy paper is what OpenAI says when it wants to be in the room with policymakers who care about labor displacement. The research suppression is what OpenAI does when it could put data in the room that would complicate the policy paper.
When you ask what this firm actually believes, picking one of those outputs as the real one is the wrong move. Both happen. The simultaneity isn’t necessarily strategic; it might just be the firm’s two operating modes running on different tracks. Either way, the political effect is the same. The public conversation about OpenAI’s technology gets conducted on terms the firm has shaped from both ends.
The Adler line ends with a useful framing for whoever’s doing this work next: “Unless we change their incentives, we shouldn’t be surprised when companies cut corners, even if they’ve said the right things.” The incentives question isn’t going away.
Not because OpenAI is uniquely bad. Because OpenAI’s posture is what’s rational under the current incentive structure for any large AI firm with significant federal-policy exposure. The mechanism: regulatory standing, talent retention, investor confidence, and policymaker access all reward a firm that arrives with a tidy proposal more than they reward a firm that arrives with an ambiguous research finding. The lab that proposes solutions gets the meeting; the lab that publishes problems without proposed solutions gets the lecture about being unhelpful. Lehane’s “applied physics, not theoretical physics” is the rational response to that environment. That doesn’t make Lehane personally innocent of the choice. It does suggest that fixing the choice requires changing the incentive structure, not just changing Lehane.
What would a credible firm look like? The honest answer is that there isn’t one yet, and that the credibility frame may not apply to firms-as-firms — it applies to firms as political actors, and the political-actor question is about delivering on policy commitments, not just about which research gets published. On the specific question of internal-research transparency, Anthropic publishes more uncomfortable findings than OpenAI does, and that difference is real even allowing for Anthropic’s own selective practices. The broader question of what a credible AI lab looks like is unanswered. Sun’s piece ends without answering it. This one does too.
What we know, after Sun’s reporting, is that the firms know what they’re suppressing and what they’re proposing, and that the same global-affairs apparatus is doing both at the same time. Knowing that doesn’t tell us what to do. It tells us what the room actually looks like when policy gets made.
These are things we need to know, and need to press to know, in the society that is to come.
If this clarified anything, or muddied things in a useful way, send it to someone who’s been reading the AI policy debate at face value. Comments are open below.
Seems like Sun isn't "excellent" if she couldn't unpack this herself. Was decidedly unimpressed with her piece, as I often am.
Interesting approach on your part tho, taking these balls and running with them.