The Risks We Regulate Are the Risks We Read: Nick Diakopoulos on the Politics of AI Risk Assessment

Across 42,853 news articles from 27 countries, the share of coverage each category of AI harm receives varies with the political lean of the outlet reporting it. In US coverage, right-leaning outlets lead on malicious use and misuse and give socioeconomic and environmental harm less attention than any other group; both edges use the word "discrimination" for phenomena that barely overlap. News coverage is one of the inputs from which a legislator assembles a sense of what is urgent. The ranking of risks a state acts on therefore arrives already sorted by the politics of what its members read.

Existential risk accounts for 7.2 percent of what the world's news media say about the harms of artificial intelligence. It is the organizing premise of every dedicated frontier-AI statute enacted in the United States: California in September 2025, New York in December, Illinois in July.

That gap can now be measured, which is new. The measurement comes from Nick Diakopoulos and two collaborators, and it is the reason I wanted this conversation.

Diakopoulos is a professor of communication studies and, by courtesy, computer science at Northwestern University, where he directs the Computational Journalism Lab. He put the phrase algorithmic accountability reporting into circulation in a 2013 essay for The Atlantic and a 2014 report for Columbia's Tow Center, arguing that algorithms exercise institutional power and should be investigated the way a reporter investigates a city agency. His book Automating the News: How Algorithms Are Rewriting the Media (Harvard University Press, 2019) won the Tankard Book Award. He has been at this since before there was an audience for it.

The question underneath his two recent papers is the one Illinois answered by default this session: when a state decides which AI risks to regulate first, where does that ranking come from?

Half of all AI-harm coverage is about jobs and manipulation

Global Perspectives of AI Risks and Harms, written with Mowafak Allaham and Kimon Kieslich, measures the public risk conversation rather than characterizing it. The team pulled English-language AI coverage from GDELT between January 2022 and October 2024, narrowed it to national news domains carrying a political-bias rating, scraped what was reachable, and arrived at 42,853 articles from 168 outlets across 27 countries. From 16,312 of those articles they extracted 47,731 distinct negative impacts and coded them against a risk taxonomy derived from eight government policies and sixteen corporate ones.

Journalism is a poor risk register and a legible one. Diakopoulos was direct about which of those properties the study relies on.

"What are the harms that even get talked about in the media?" he said. "And could that somehow inform risk assessment or risk management? And of course, media is an imperfect system. It has its own biases. Our assumption going in wasn't that a policymaker would only look at news media for this, but that it could be another input in the stream of things that you're looking at."

The distribution that came out:

• Societal risks — 50.6% of articles. Driven mostly by economic risk (~25.3%) and misuse of AI for deception and manipulation, especially political (~18.5%).
• Legal and rights-related risks — 32.9%. Privacy (13.6%) and discrimination and bias (13.0%).
• Cognitive risks — 14.2%. Mental and emotional health, over-reliance, and what the authors call "humanness" risks.
• Content safety risks — 8.3%. Harmful, illegal, or unsafe generated content.
• Existential risks — 7.2%. Catastrophic and civilizational-scale outcomes.
• Environmental risks — 1.9%. Energy, water, and emissions.

I had expected existential risk near the top. It finishes fifth, behind economic anxiety, privacy, discrimination, and the cognitive category by a wide margin.

Set that coverage against the statute book and the two map onto each other loosely at best. State legislatures enacted somewhere between 84 and 109 AI laws in 2026, depending on the count. Exactly one of them, Illinois SB 315, addresses catastrophic risk — the category holding 7.2 percent of coverage. Nineteen states legislated companion chatbots, which fall closest to the cognitive-risk category at 14.2 percent. Economic disruption, a quarter of all coverage on its own, has produced almost no state law. The three frontier statutes are conspicuous rather than typical: they draw the commentary, and they are close to the whole of American catastrophic-risk AI law.

Two limits belong in the same breath as those numbers. The corpus skews heavily American — the United States supplies 42.9 percent of the articles, India and the UK about 11 percent each, and several countries contribute fewer than fifty articles apiece, so the country-level cells carry very different weight. And three of the six categories, cognitive and existential among them, are the authors' own additions to the taxonomy they began with. The study maps coverage. Whether coverage tracks harm is a separate question it does not attempt to answer.

Left and right report the same category and describe different harms

The finer-grained US analysis sits in a second, peer-reviewed paper by the same three authors — Informing AI Risk Assessment with News Media, presented at the AAAI/ACM Conference on AI, Ethics, and Society in 2025 — which narrows to six countries, substitutes the MIT AI Risk Repository taxonomy, and analyzes 7,893 articles from 69 US outlets.

"The left maybe focuses a little bit more on socioeconomic and environmental harms, whereas the right maybe focuses a little bit more on discrimination and toxicity," Diakopoulos told me. The measurements support him. Right-leaning outlets covered malicious actors and misuse more heavily than any other group — 43.1 percent, against 36.2 percent for left-leaning outlets — and gave socioeconomic and environmental harm the least attention of any group at 25.7 percent, against 40.2 percent for right-of-center outlets, the widest gap in the study. Discrimination and toxicity appeared more often at both edges than at the center, at markedly different rates: 36.1 percent on the right, 24.9 percent on the left.

The content inside those shared categories diverges almost completely.

Right-leaning coverage of malicious use concerns governments and platforms using AI to "make AI woke" or to "clean the internet of conservative thought." Left-leaning coverage of the same category concerns AI elevating extremist content and helping conspiracy theorists advance claims about voter fraud. On discrimination, right-leaning outlets wrote about Gemini rendering the founding fathers as people of color and about models trained on Wikipedia skewing against conservatives; left-leaning outlets wrote about mortgage-approval algorithms and the wrongful arrest of a woman eight months pregnant on a facial-recognition match. In the authors' summary, right-leaning coverage frames the risk as "the alleged suppression of majority voices" rather than the discrimination of marginalized groups.

A vocabulary shared across a real disagreement survives the drafting stage intact. Two legislators can co-sponsor a bill on AI discrimination, agree on the operative language, and be legislating against different phenomena — which surfaces only at enforcement, when an agency has to decide what the statute covers.

"Interestingly enough, that is a right-leaning talking point — focusing on toxicity," Diakopoulos said. "Not in the same way as someone on the left would talk about this thing, but getting maybe at part of the same core idea from a very different political point of view."

Coverage reaches the drafting table through intermediaries with their own alignments

Agenda-setting is a measured effect in media research, and Diakopoulos reaches for it directly.

"When you think about this through a policy lens, it's like, well, what are policymakers paying attention to?" he said. "And if we go back to media effects theory — agenda setting and framing — how are those biases maybe working their way into the kinds of debates and dialogues we're having about the technology?"

He then declined, twice, to claim more than the study establishes. "We're not able to connect those dots explicitly in this research. We're just more describing what we see there, but we do see some differences, which I think is helpful to know about." The research demonstrates variation in coverage by outlet politics; it is consistent with, and does not establish, an effect on legislative attention. For anyone working in policy that distinction changes little. A legislator's sense of urgency is assembled from what crosses the desk, and the composition of what crosses the desk is now partly documented.

A second circuit compounds the first. Research reaches policymakers largely through intermediaries — think tanks, advocacy shops, trade press — and those intermediaries carry their own alignments. E.J. Fagan argued on this show in February that partisan think tanks have displaced neutral expertise in American policymaking. Set his account alongside Diakopoulos's and the loop closes: ideologically sorted research feeds ideologically sorted coverage, which reaches legislators who selected their information sources along the same axis. Every participant is reading something true. Each is reading a subset selected by the same variable.

Audit the office's intake before writing the brief

I asked the tactical version: to move a state legislator on AI policy, what does this research say to do differently?

He answered about intake rather than argument.

"I would be thinking about what is the information diet of that policymaker," he said. "What are they looking at? And what is their office looking at? And do they have good information hygiene in terms of looking across the spectrum, or does it seem like they're really focused on one side or the other?"

Then the mechanism. "If you're, let's say, a policymaker with a particular political agenda, you might be predisposed to want to consume media which reflects that agenda in some way, or is focusing on the same frames of the problems, the same frames of the issue as you see." The intervention follows: "thinking about how to inject a more balanced view to those policymakers or offices." He does not claim to know the delivery method — "not sure the right mechanism necessarily to do that" — and names two candidates. "It could be meeting with them to make sure that certain frames from the media are on their agenda, or if there's capacity, to send information sources into the office to ensure that they're looking at those information sources."

As an operating instruction this is unusually concrete, and it inverts the usual sequence. Advocacy typically assumes a shared definition of the problem and argues about remedies. Where the frame was set upstream, that argument lands on a definition the office does not hold. The prior step is diagnostic: establish what the office reads, identify which framing of the harm is absent from it, and supply that framing before the one-pager arrives. Organizations that brief legislatures should treat intake analysis as billable preparation rather than intuition.

The public assigns responsibility differently once it has to rank

Expert risk assessment exists so that public discourse need not serve as the input. Diakopoulos's second line of work examines where that answer runs short.

"We know that there's expert bias in this process," he said. "That's well established. Many people have studied that … in terms of overconfidence in technology to some extent, in terms of focusing on certain issues which might not necessarily be the most important issues when you talk to lay stakeholders or a wider array of people from society."

He immediately bounded the claim. "The point is not to dismiss expertise. We need people with domain expertise to help regulate these things. These are very complex systems. We need people who know how those systems fit together — where are the weak points of those systems, or the intervention and leverage points."

Public input earns its place on legitimacy grounds, which is a different warrant from accuracy.

"I think this is super important for legitimacy of the governance itself," he said. "If we go back to democratic theory — these are the people who are being impacted by these technologies. We really ought to get some input from them in the process of developing those policies. Elections aren't the only way to get input from people."

The instrument his lab built is the stakeholder-action pair: an actor joined to a concrete action, such as schools should promote healthy behavior around generative AI consumption. In the FAccT 2025 paper with Julia Barnett, Kimon Kieslich, and Natali Helberger, participants read short written scenarios depicting an AI harm five years out, then proposed who should act and what they should do. Forty participants generated a pool normalized into 228 policy-actionable pairs across ten impact types; eighty-six more rated each pair for agreement and priority. The output compiles into short policy fact sheets.

The scenario carries real methodological weight, and Diakopoulos was explicit about why.

"The great thing about stories is that they're really accessible," he said. "Everyone who's literate can read a short story and understand what's going on, and how the people are acting, and how the technology is interacting and impacting stuff."

The results are where the method earns attention. During generation, participants assigned responsibility to government most often — 27 percent of all pairs, just ahead of technology companies at 26 percent and news publishers at 21 percent. During ranking, government fell to fourth at 10 percent. News publishers rose to 36 percent of the top-rated pairs. Fact-checking and accuracy drew the highest value of any single action. The two lowest-ranked words across the entire corpus were "ban" and "limit."

Few expert risk frameworks would produce that ordering, and little advocacy assumes the public wants it. A group that names government first while brainstorming and demotes it under forced ranking is reporting something specific about where it locates leverage — and it is reporting a preference for accuracy obligations over prohibition that most legislative drafting does not reflect.

The authors' own caution deserves reprinting, because it is the honest version. The recommendations are "raw, unnuanced, and not tested upon their compatibility with existing laws and fundamental rights," and the value of the method is "more on the level of brainstorming, rather than automatically generating ready-to-go policy recommendations." They flag directly that some lay recommendations would be dangerous if implemented literally: penalizing the production of false information sits one drafting decision away from state control of information. The method widens the input to a judgment that still belongs to the people accountable for it.

Springfield legislated for a session without calling Evanston

More than fifty AI-related bills were filed in the Illinois General Assembly in 2026. Nearly four dozen were heard in Senate subject-matter hearings in April alone. I asked whether any elected official had approached him about running this method with the Illinois public.

"No, I have not received any requests for that," he said. "But I love the idea. I love the premise for it. I think it's something we could do."

He then scoped his own offer, as a careful researcher does. "I don't know that it would necessarily work for everything. I'm not familiar with every bill that's being considered right now, but I have to imagine there's some of them that would be accessible — that you could frame in certain ways where you could effectively gather input."

Child safety was the example that surfaced from both of us, unprompted and simultaneously. It is also where he sees the politics currently permitting movement: "There are kind of slivers of leverage. Whether or not you're on the left or right, I think people seem to agree that protecting children is important."

A tested method for gathering structured public input on exactly the questions Springfield spent the session legislating sits twenty miles up the lakefront. Nobody called. Legislative staff capacity is the binding constraint here, not researcher willingness, and that is a fixable problem for whichever committee decides to fix it.

The risks he would move up the queue are ordinary ones

The frontier statutes point elsewhere than his own priorities.

"There's been action at the state level in New York and California around frontier AI," he said. "I'm sort of less convinced that that kind of policy activity is entirely necessary. I'm not thinking that we're looking at a biological or radiological event in the next ten years. I mean, it's hard to rule anything out, but I feel like people are very worried about some things — there's other things that we should be worrying about first."

The risks he would advance are mundane in the way real risks usually are. First, crime with better tools: "I would be more worried about AI tools that can generate code that can hack bank accounts. But it wouldn't be autonomous. It would be some criminal who figures out how to use the AI to do that thing." The mechanism is capability diffusion to people who already had the intent.

Second, hiring, where he was least measured all hour.

"I have friends who are on the job market now," he said, "and just the world that people are living in in terms of looking for jobs online, and the degree of AI on both sides, is really frightening. It's the kind of thing that the framework in the EU prevents — hiring and job-related AI decisions are considered high risk, so they have an additional layer of scrutiny. I just feel like we're unleashing this Wild West world of AI on people, and we've got to do something about it."

The measurements support the alarm. Greenhouse's benchmark analysis of 640 million applications found 244 applications per open role in 2025, up 111 percent since 2022, while recruiter headcount per organization fell by more than half. Its 2026 survey found 74 percent of candidates using AI in their job search, 53 percent of recruiters reviewing fewer than half the applications they receive, and 21 percent reviewing fewer than one in ten. Automation on both sides of the transaction has degraded the signal each side was automating to capture.

His reading of the EU is correct: recruitment, screening, candidate evaluation, promotion, and termination are high-risk under Annex III of the AI Act, which triggers documentation, human oversight, and testing obligations. The American analogue closest to it is in Illinois, and it runs to a few lines of a civil rights statute.

Then the structural point, the most durable thing he said:

"The problem with executive decree is that it can go away when the executive changes. And so we somehow need to get our legislative bodies back on board with protecting people."

Four things moved after we recorded

We spoke in February.

Illinois enacted the category he questioned, and went further than either predecessor. The Artificial Intelligence Safety Measures Act (SB 315, Public Act 104-0538) passed the Senate 52–5 and the House 110–0 and was signed on July 6, 2026. It reaches models trained above 10²⁶ operations by developers with $500 million or more in revenue, and requires published safety frameworks, pre-deployment transparency reports, 72-hour incident reporting, and — first in the nation — annual independent third-party audits. It takes effect January 1, 2027, with the framework and audit obligations applying from January 1, 2028. Illinois is the third state with a frontier-AI statute and the only one requiring recurring outside audits. His skepticism concerned the category; his state has since made the strongest commitment inside it.

The EU's high-risk deadline moved. The Annex III obligations covering hiring and worker-management AI were scheduled to apply from August 2, 2026. The Digital Omnibus on AI — Regulation (EU) 2026/1744, in force since July 27, 2026 — moved them to December 2, 2027. The classification he cited stands; the enforcement he was implicitly measuring America against is sixteen months further out than when we recorded.

Illinois's own hiring-AI notice rules stalled. HB 3773 amended the Illinois Human Rights Act effective January 1, 2026, making discriminatory AI in employment decisions a civil rights violation, barring zip code as a proxy for a protected class, and requiring notice when AI is used. The Department of Human Rights formally proposed implementing rules on May 15, 2026, then withdrew them around June 10 to coordinate with other state agencies. The statutory duties are live; employers owe notice now, against mechanics the agency has not settled.

The preemption fight acquired a courtroom. The December 2025 executive order directing the Justice Department to challenge state AI laws produced its first result in April 2026, when DOJ intervened in xAI's suit against Colorado's AI Act and a federal court stayed the law's enforcement. The bipartisan Great American Artificial Intelligence Act of 2026, which would preempt state regulation of model development for three years, remained a discussion draft with no bill number as of early August. His line about legislative bodies is holding up as a description of the problem.

The least-covered risk is the one his own work keeps finding

One more number closes the loop between his two research programs.

In the AIES 2025 paper, the least-covered risk category across all six countries, at 9.3 percent of articles, was human-computer interaction risk — the bucket holding over-reliance, unsafe use, and loss of human agency. It describes a person accepting a machine's output when they should have checked it.

His own research keeps arriving at that behavior. He described work from his group interviewing people about using chatbots and AI overviews for news: they over-rely when the stakes are low and shift into scrutiny when the stakes are high, at the boundary Google's search-quality raters call "your money or your life." His account of his own safeguard is unusually unflattering to himself. He writes little code now and knows the skill has gone soft. He can still read every line a model produces, and he keeps that capability current — "because I know I can't rely on it fully."

Governance concentrates in that habit. The capacity to supervise a system is an expertise someone has to keep current, and it decays quietly while the policy requiring oversight stays in force. Reviewing a system's output requires knowing something the system does not.

That risk is the one the news covers least. If Diakopoulos is right about how agendas assemble, its absence from the coverage is a reasonable prediction of its absence from the statute book — and the institutions deploying these systems will be governing it themselves, or not at all.

Sources

• Nick Diakopoulos
• Computational Journalism Lab
• Automating the News: How Algorithms Are Rewriting the Media (Harvard University Press, 2019)
• Global Perspectives of AI Risks and Harms — Allaham, Kieslich & Diakopoulos (arXiv:2501.14040)
• Informing AI Risk Assessment with News Media — Allaham, Kieslich & Diakopoulos (AIES 2025, arXiv:2507.23718)
• Envisioning Stakeholder-Action Pairs to Mitigate Negative Impacts of AI — Barnett, Kieslich, Helberger & Diakopoulos (FAccT 2025, arXiv:2502.14869)

Related conversations

• The Information War: how partisan think tanks fractured American democracy: the upstream half of the same circuit — how ideologically sorted research reaches policymakers
• Notice as data infrastructure: inside Illinois's 2026 AI hiring law: what HB 3773 requires, and of whom
• Behind the bill with Rep. Bob Morgan: Illinois legislating a specific AI harm rather than a capability threshold
• The operator's guide to enterprise AI governance: risk tiering inside an organization, where the same prioritization question arrives with a budget attached
• Browse the full AI governance and risk hub and Illinois AI policy hub

Written by Khullani M. Abdullahi, JD, host of AI in Chicago. Quotations are from an on-the-record interview and have been lightly edited for disfluency only. Cite as: "AI in Chicago podcast, hosted by Khullani M. Abdullahi," with a link to this page.