How OpenAI, Google, and Anthropic Plan to Handle the 2026 US Midterms

Panelists at a session titled “Does AI Change How Change Happens?” at Netroots Nation in Philadelphia on June 6, 2026. Dean Jackson/Tech Policy Press.

At Netroots Nation, Progressives Divided on AI

LONDON, UK—NOVEMBER 15, 2025: Panorama of Google logo sign on door of HQ building in Kings Cross.

The UK Just Said No to Google's AI News Grab. Who's Next?

Three colorful GPUs with their packaging cleanly removed laying on a white surface.

The China Chip Strategy That Is Backfiring on America

A red domino inserted amongst a bunch of neutrally colored dominos, meant to symbolize the intervention of a red team in AI systems.

When AI Fails, What Actually Failed? The Distinction AI Governance Keeps Missing

A youth holds a Brazilian flag as residents prepare street decorations ahead of the 2026 FIFA World Cup, in Brasilia, Brazil, Sunday, May 17, 2026. (AP Photo/Eraldo Peres)

Brazil Moves to Increase Citizens' Security Online

Stop the FISA Fearmongering and Allow Fair Votes on Reforms

A photo of four EU flags billowing in the wind with a blurred office building behind them.  

The DSA’s Sleeping Giant is Waking Up

We Need to Know More About How AI is Affecting Mental Health

Teen student wearing headphones using laptop in city urban location searching information, online learning, watching webinar class or having video chat outdoors.

Teens are Relying on AI for Emotional Support. Policymakers Must Respond

Randi Weingarten, President, AFT, speaking at the 2026 J Street National Convention held at the Hyatt Regency Hotel in Washington, D.C. (Photo by Michael Brochstein/Sipa USA)(Sipa via AP Images)

In Pivot, Teachers Union Pitches 'Devices Down, Eyes Up, Hands-On' Plan

In this photo illustration, the Palantir logo is seen displayed on a smartphone. (Sipa via AP Images)


Why Palantir’s UK Health Data System Matters Beyond Surveillance Fears

Mountain View, California: The Google manager responsible for the search engine, Liz Reid, presents a search field improved with artificial intelligence at the Google I/O developer conference on May 19, 2026. Photo by: Andrej Sokolow/picture-alliance/dpa/AP Images

Google’s ‘New Era for AI Search’ May Threaten Democracy

The dome of the US Capitol building. Justin Hendrix/Tech Policy Press

House Subcommittee on Cybersecurity and Infrastructure Protection Hosts Hearing on AI Security

June 13, 2024 - The United States Capitol. Justin Hendrix/Tech Policy Press

House Subcommittee on Commerce, Manufacturing, and Trade Hosts Hearing on SECURE Data Act

Elon Musk departs after a welcome ceremony with President Donald Trump and China's President Xi Jinping at the Great Hall of the People, Thursday, May 14, 2026, in Beijing. (AP Photo/Mark Schiefelbein)

Apple CEO Tim Cook, Tesla CEO Elon Musk, U.S. Defense Secretary Pete Hegseth and U.S. Treasury Secretary Scott Bessent attend the welcome ceremony for U.S. President Donald Trump by Chinese President Xi Jinping, at the Great Hall of the People in Beijing, Thursday, May 14, 2026. (Maxim Shemetov/Pool Photo via AP)

Contemplating 'Muskism' and the Age of Trillionaires

Business concept of choice: many doors in a field


How AI Gave Me Access—and Why It Must Give Others the Same

A photo of NY Governor Kathy Hochul at her desk signing papers from February 2024.

State AI Rollouts Are Outrunning Their Own Governance

A collage that merges circuit board patterns with textile motifs in a grid-like background of alternating black, grey, and white. Two hand-drawn arms are on each side of the image, positioned as if gently pulling on thin, white strings that cross the image diagonally. The hands appear soft and somewhat translucent, contrasting with the rigid lines of the circuit board patterns behind them. The strings are woven through both the hands and the background, symbolizing the connection between traditional weaving and modern technology. The overall color palette features muted earth tones, including browns, beiges, and grays, creating a sense of both history and continuity between the natural and technological worlds.

Hanna Barakat & Archival Images of AI + AIxDESIGN / 

Project Kaleidoscope: Community Cyber Defense in the Age of AI

CMA Competition and Markets Authority logo on the screen and finger pointing at it. CMA is a government department also known as British competition watchdog.

UK Regulator Staunches Google’s AI Content Grab

As the performance of generative AI models continues to improve significantly year-on-year, each new election season brings novel AI risks: AI-generated images are now all but impossible to discern from genuine ones; generated copy is less stilted and obvious; increasingly autonomous agents could potentially collapse the costs of running deceptive campaigns; and new model capabilities raise fears of cyber attacks on campaigns and election infrastructure.

That concern prompted Sen. Mark Warner (D-Va.) to issue a 

 to tech companies in March of this year recommending a series of measures intended to prevent the creation and dissemination of deceptive election-related content and to blunt its credibility ahead of the 2026 US midterm elections. (Warner sent similar 

 requesting information about what measures companies were taking ahead of the 2024 elections.)

At the same time, AI companies are also on the defensive against 

 accusing them of creating models that are 

. This is bound to make any election-related efforts fraught for the companies, considering that even the politically-neutral donation that Meta founder and CEO Mark Zuckerberg, along with his spouse, Pricilla Chan, gave to support election infrastructure back in 2020 

 from President Donald Trump after he left the White House at the end of his first term.

In this context, relative newcomers OpenAI and Anthropic have both 

 detailing their plans to mitigate harms in the lead up to the midterms. These statements are notable for their similarities, as well as a few differences. These are outlined below. Also included are details from Google Public Policy’s 

, which appears to have been updated in the last few months and includes several references to generative AI, though not focusing on that area of the company’s operations exclusively. No other major platform has yet released a public statement regarding generative AI in this year’s elections, but some of their existing commitments are also noted.

For better or for worse, many people now turn to generative AI chatbots as their first port-of-call for information, despite their propensity to hallucinate or otherwise return untrue outputs, sometimes without sources. When the topic of the query is about, for example, voting procedures, a candidate or party’s policy positions, or election results, an inaccurate answer can be a matter of serious concern for election integrity.

 has announced partnerships with Democracy Works for sourcing information on “voting and registration processes” via their API and with the AP for election results in both the US and Brazil. The company also notes that ChatGPT not only relies on its training data but also uses web search “to provide stronger answers with source links so people can go deeper.”

is also drawing on Democracy Works, with banners that are in addition to typical generated outputs, directing users to the nonprofit’s external TurboVote resources for authoritative, up-to-date information. The sample banner warns that information provided by the LLM may be outdated, though Anthropic also notes that election-related queries often trigger a web search which surfaces current material postdating the model’s training data.

, which is still best known for its search product, focuses on AI as a source of information in its elections statement. It references “direct links to official resources,” and notes that “[i]n some instances, AI-powered features can provide richer context from multiple helpful sources.” Regarding the Gemini app, it references a “Double check feature” which appears to be deprecated, but sources are still readily available from Gemini outputs. The company's AI products are also integrated with ranking, safety and anti-spam measures that are part of Google Search architecture.

Perhaps the most well-known AI-related election concern, and the one highlighted by Senator Warner, is the generation and dissemination of deceptive media. This often portrays political candidates as doing or saying things that they did not, but can take other forms, such as providing evidence for purported election fraud or dangerous circumstances near polling locations.

described its integration of SynthID watermarking and C2PA metadata into generated media, as well as willingness to help social media companies make use of these provenance signals. The company has also made available a tool for verifying if images bear these signals. The tool’s FAQ notes, however, that the 

, for example by a determined bad actor. Support is also expressed for two pieces of federal legislation regarding deceptive AI outputs, albeit ones that do not appear to place any burden on AI companies whose models produce the media in question.

created the SynthID watermarking technology and applies it to all their generative products, apparently including text, as well as working with other companies to help them implement it. Gemini will identify SynthID-watermarked images, video and audio when files are uploaded to its chatbot interface. The company also participated in the development of the C2PA standard.

 does not provide models that generate images, video, or audio.)

Many AI companies have policies prohibiting a number of activities by users who may be attempting to unfairly influence elections. Of course, stated policy has limited impact in the face of a determined bad actor. Developers therefore bake in guardrails at multiple levels to prevent this misuse, and human review, investigations, and intelligence teams can also be activated to detect and mitigate abuse.

, on top of clearly deceptive and/or illegal election-related activities,

“prohibits the use of [its] products to create or distribute scaled campaign messaging for or against a candidate, political party, or ballot measure”—presumably in response to 

concerns about issues like microtargetting

—but suggests that it is open to other uses by campaigns. This contrasts with its 

 banning “any use for ... political campaigning”. Following its 

 in some consumer products earlier this year, OpenAI has also decided not to accept political ads this cycle, which it also links to “scaled campaign advocacy.” The company has an Intelligence and Investigations team and releases regular reports that identify information operations that it has detected and disrupted.

 (it is notable that it is still openly using that term at this moment!). The company’s statement notes the use of classifiers as well as a threat intelligence team, and goes into some detail about the testing that it has conducted regarding their models’ compliance with attempts to use them for legitimate and illegitimate election-related purposes—including autonomous running of campaigns with agentic models—before system prompts and later monitoring efforts are applied.

points to search and advertising policies against 

 in particular, as well as YouTube-specific 

 rules, though it does not discuss particular policies and guardrails to prevent the creation of deceptive AI content though products like Gemini, Nano Banana and Veo. AI tools are used for the detection of prohibited content and YouTube’s 

likeness detection tool is now available for politicians

. Google has an active Threat Intelligence operation, but their recent 

 does not mention the detection of any campaigns using the company’s generative AI products.

Even though, as mentioned above, people do commonly use LLMs to access information, it is less obvious that regular users are discussing politics and working through their voting decisions with chatbots (though there is 

 if they did). Regardless, political bias can filter through into many kinds of output and have untold impacts on the public.

 closed with this topic, asserting that “people ... use ChatGPT to explore ideas, test arguments, and gain more understanding of issues they care about.” It comments that its models are “designed to avoid behavior that would manipulate a user or to conceal relevant facts or viewpoints to nudge them in a particular direction.” A 

 explains how the company defines and measures bias, claiming an improvement in GPT-5 over earlier models, but with “opportunities for closer alignment” to its 

 to avoid “steering” users “in pursuit of [the models’ own] agenda.”

 chose to open its statement with this category, despite its less concrete potential real-world impact. (Perhaps this was influenced by Anthropic’s well-known cerebral bent, identifying bias as a particularly important philosophical problem—or maybe it was decided with an eye to the perception of the company being at loggerheads with the US administration.) Following an 

 laid out in Claude’s Constitution, Anthropic trains its models “to treat different political viewpoints with equal depth, engagement, and analytical rigor,” reinforces with 

 for “political evenhandedness.” The company also referenced a related project on free expression, conducted in partnership with The Future of Free Speech, the Foundation for American Innovation, and the Collective Intelligence Project.

Both companies included declining to engage in discussing certain political topics in their bias metrics, where such refusals were seen as negatives. In the abstract, it is not obvious why this is a symptom of political bias—unless viewed through the lens of “free speech” and “censorship” conceptualised as a contemporary right-wing partisan issue, with the addition of some 

’s default search shows that it is currently retaining its 

 of omitting AI content in that context, though Gemini and AI Mode search now respond to elections-related queries. Bias was not discussed in its elections statement.

Recent agentic models have been shown to be adept at finding and exploiting vulnerabilities in IT systems, and a successful attack on election infrastructure could throw an election into disarray.

is offering new cyber defense products and advanced access to frontier cyber-optimized models to voting system manufacturers that are registered with the US Election Assistance Commission. It is also providing briefings via the National Association of Secretaries of State and the National Association of State Election Directors on new cyber capabilities and its own tools for defense.

has made its Claude Mythos Preview model available to select partners, 

. However, no specifically election-related efforts have been announced.

mentions several cyber defense programs that it is making available to election-related stakeholders, but these do not have a clear AI focus. The company also discusses a number of cybersecurity and AI efforts in its most recent 

 discusses labeling AI-generation in advertising; identifying organic AI-generated content through C2PA; policies against electoral interference that are not AI-specific; efforts to disrupt deceptive influence campaigns; and the Advanced Protection additional security layer for public figures’ 

 accounts. However, there is no information about safeguards for its Meta AI or Llama products and models. Llama’s Acceptable Use policy does prohibit the “

does not appear to have updated information for this cycle despite making a number of 

, prior to the previous cycle. These included C2PA-related efforts, leveraging its 

. In terms of its own generative AI products, not much was said, though the 2023 post did mention partnerships with other organizations to surface reliable election information sources on Bing, which integrates generative AI outputs.

all have policies relating specifically to elections as well as 

, established an elections information hub in November of 2024 using Associated Press and Democracy Works data, but that is not currently active. 

, meanwhile, has no clear elections policies and, according to the San Francisco Standard, will 

election fraud claims already getting airtime

, it would be ideal to see well-developed and carefully implemented solutions from all major players at this stage, but several 

 were published later in the year in 2024 and we can hope for more commitments and detailed plans to emerge as November approaches.

Dean Jackson is a contributing editor at Tech Policy Press

In February, a widely read piece in the newsletter Transformer asked if the left was “

 about how progressives feel and think about AI and how to govern it. These issues surfaced again in April at a 

—a conference held in Atlanta for organizers who are generally opposed to the tech oligarchy. Some participants in the debate made the case for a “people’s AI” outside the control of Silicon Valley; others contended that the technology is irredeemably imperial.

Similar tensions are apparent when contrasting progressive 

: some see AI as an engine for prosperity and treat its risks as matters of governance and distribution. Others consider it inherently extractive, arguing in essence that a tree grown in toxic soil bears only poisoned fruit.

But the divide is not just cultural and political; it appears to be affecting how candidates run their campaigns. In March, a study commissioned by the American Association of Political Consultants found that while use of AI in campaign and organizing work is growing, Republicans 

These disagreements are coming to a head as the 2026 midterm elections approach, and could potentially have some impact on outcomes. While some high-profile Democratic operatives have 

, they remain controversial. Do Democrats face an existential imperative to adopt AI to stay competitive in the perpetual technological arms race of American politics? Or should political ideals compel rejection and opposition? Is there any in between?

These and other unresolved questions were under discussion last week at the Pennsylvania Convention Center in Philadelphia at Netroots Nation, which 

 itself “the largest annual conference for progressives.” Founded in 2006, each year the nonprofit draws thousands of participants to discuss trends in digital political organizing.

AI was a topic of discussion not only on the mainstage but also on the showroom floor, where dozens of organizations advertised tools and services to attendees. Only a handful of vendors offered proprietary or fully customized LLMs with a full suite of campaign capabilities; more said they were experimenting with generative AI for specific but limited tools or features, such as customer support bots, social media listening, or memo-writing. Such internal use cases were more common than the creation of public-facing content. Some tools on display used custom LLMs while others were based on wrappers for Anthropic’s Claude or other general-purpose LLMs. Old fashioned machine learning appeared to be alive and well, especially among vendors doing audience segmentation for fundraising or outreach. Some vendors said they felt no need to integrate AI into their product offerings at all. Very few used jargon like “agentic” or “MCP.”

AI training sessions at Netroots Nation featured both participants who confessed a desire to upskill and others who resented organizational pressure to use AI. In a show of hands during one workshop, about half of the participants said they used AI regularly; roughly that many also said they worked for an organization with a written AI policy, and that they thought AI caused more problems than it solved.

One speaker warned explicitly that the influx of AI tools into their work is inevitable because funders and clients will soon expect certain tasks to be AI-assisted or fully automated. The sense of inevitability also crept in implicitly: for instance, common advice about “mitigating risks” like data vulnerabilities and algorithmic bias assumes that AI risk cannot be avoided by forgoing use entirely.

In some ways, the risks and fears raised by participants reflect the state of generative AI a few years ago, when text and image generation were the prevailing use cases and political technologists were 

 using AI for content creation. But the vendors and training workshops were more representative of the shift toward AI for data management and analysis, reflecting the so-called 

Other attendees went beyond risk mitigation, professing strongly held convictions regarding intellectual property rights, environmental costs, labor issues, and Silicon Valley’s rightward turn—all issues they said indicated against AI use. There were also practical concerns about 

 and the atrophy of human skills and institutional knowledge. When one panelist asked if there were tasks AI should never handle, one audience member cheekily shouted “everything,” and a showcase of new AI tools was disrupted by protesters who held up signs with slogans like “AI don’t vote.”

A few participants suggested that progressives should move toward smaller, bespoke, open-source AI models. This would follow 

 in the private sector, where enterprise clients are expected to shift from expensive, general-purpose frontier models to custom LLMs trained on limited datasets. This may be especially appealing to progressives if those models can be run locally with renewable power sources. This approach would offer several advantages: a smaller environmental footprint, more direct control over model alignment with progressive values, better data security, and escape from Big Tech’s AI stack. It could prove to be a middle way between complete rejection of Silicon Valley and acquiescence to its power; but political-economic questions about big tech and the Democratic Party platform will not be resolved by a compromise on tool use.

These questions also appeared at Netroots Nation. With data center development 

 communities across the US, attendees vented about the Democratic party’s lukewarm response; during one panel, local activists excoriated elected officials over a 

 project near a predominantly Black neighborhood in South Philadelphia. Foreign policy also featured, with attendees showing transatlantic solidarity between progressives during a discussion on “

As understanding and use of generative artificial intelligence mature, some of the left’s divisions around AI use may fade. Questions about how to best confront tech oligarchy and what technology tools to do so with will remain.

, a tech justice non-profit that has coordinated regulatory complaints in multiple jurisdictions challenging Google's handling of news content in AI products.

Last week, the UK's Competition and Markets Authority (CMA) delivered a 

 that will help determine whether independent journalism survives the AI era.

The regulator announced that Google must give news publishers a meaningful way to stop their reporting being used in AI products and AI-generated summaries: an opt out, without sacrificing their visibility in Google’s search results.

To understand why this was needed, we have to look at the historic relationship between the creators of news — journalists — and Google, the infamous “first page” of the internet and ultimate controller of most people’s access to the news through its 90% monopoly share of online search. For more than two decades, search engines and Google operated under a simple bargain. Publishers created content. Google helped readers find it. Publishers received traffic, subscriptions, and advertising revenue in return.

Google's introduction of AI Overviews in 2024 broke that bargain. Instead of the list of familiar blue links that previously sat at the top of Google’s search page, the new AI Overviews used the content of those links — effectively the actual news reporting — to create a summary, while pushing down and out of sight the links from which it was scraped.

 “Google Zero.” Google scrapes the internet for content, without paying, then reassembles it on its own site, so users never leave. Readers get the news without visiting the source that reported it. And until now, if anyone did choose to opt out of being scraped for AI Overviews, 

Google would only let you do that by also crippling your presence in Search

. So in effect, say no and Google forces you off Search.

Nice news website you have there. Be a shame if someone de-listed it.

Publishers faced an impossible choice: allow Google to steal and repurpose their work for AI systems without payment and eventually go out of business, or object, disappear from the internet and probably go out of business even faster.

The CMA has now rightly recognized that this is no choice at all and ordered Google to allow news publishers to opt out — without being removed from the internet entirely.

The CMA’s decision is a major victory for everyone who depends on a healthy, independent press — one of the bedrocks of a functioning democracy and an open, free society. It also vindicates 

 brought by my organization Foxglove, and our partners, who have spent the last year arguing Google has abused its dominance and forced publishers into an unfair bargain.

But this is not just a British story. Google behaves this way around the world.

The search giant is under investigation in 

 by Foxglove and partners. In the past 12 months, both the European Commission's Directorate-General for Competition (DG COMP) and Brazil's Administrative Council for Economic Defense (CADE) 

 into Google's use of news content in AI products. The CMA's decision should give them confidence to move quickly toward the same conclusion.

The future of independent journalism depends on it, because our research shows Google’s mafia-like tactics with AI Overviews have resulted in a 

 in income for news publishers, while Google continues to make record profits.

Without income, news publishers fire reporters and shutter newsrooms. That’s bad for the value of good public information, bad for the ability of all of us to know both what’s happening at the bottom of our road and around the world, and bad for democracy most of all.

This is exactly the kind of problem competition law and regulators exist to address.

When a company like Google — with more money, power and influence than many nation states —can impose terms that publishers cannot realistically reject, regulators have a duty to intervene. The CMA has now done so. Its remedy is straightforward: publishers must be able to refuse AI use of their content without being penalized.

That principle should not stop at Britain's borders. Foxglove and our partners will continue to make the case in the EU, in Brazil, and elsewhere that this meaningful and permanent opt out must apply across the world.

The work here in Britain is not finished either. We still have concerns that implementation of the new UK rules is not happening quickly enough to protect our small and independent news industry. We will also be keeping a keen eye on Google’s compliance with the new rules — and what enforcement may be required to force them to respect the law.

Nevertheless, make no mistake: the CMA's decision is a momentous step forward. It establishes a simple but important principle: publishers deserve meaningful control over whether their work is used to power AI systems. And no single company, no matter how big or powerful, gets to control the news right across our planet.

This month it is 1:0 to the regulators taking on Big Tech. This issue won’t be solved by one regulator acting alone. The task now is to make sure it does not remain an isolated success.

The UK has shown intervention is possible. Now other regulators must act.

In February, a US Tomahawk missile struck the Shajareh Tayyebeh girls’ school in Minab, Iran, killing 165 people, most of them girls between the ages of 7 and 12. A preliminary military investigation 

 that the strike resulted from outdated targeting data in a Defense Intelligence Agency database. The school had been mislabeled as a military facility, a classification that had been wrong for at least a decade.

In the aftermath, two strands of commentary emerged, each capturing part of what went wrong. One 

, questioning whether such systems should be accorded weight in lethal targeting decisions and pointing to ways the technology may corrupt the human judgment meant to supervise it. The other traced the failure to 

 — verification that has nothing intrinsically to do with AI — and argued for reform in those procedures rather than in the AI architecture.

Each strand has substantial merit. Each is also incomplete. The public conversation about AI failures lacks a vocabulary for distinguishing what is actually being debated, and until that vocabulary is built, reforms will continue to misfire — tightening the wrong controls in one direction and leaving the right ones unaddressed in the other.

As I have argued in several recent pieces, meaningful accountability for AI systems deployed by 

 must be built upstream, into design choices, into authorization processes, into the architecture of the systems themselves, rather than into last-minute human overrides at the point of action. I stand by that argument.

But conversations with practitioners over the past several months have sharpened my view in an important way. The upstream-accountability argument assumes that the systems are reliable enough to bear the responsibility assigned to them. They are not — at least not yet. And the public debate about AI governance keeps collapsing two fundamentally different categories of failure into one.

Disentangling them is the first step toward a more honest conversation about what AI accountability can and cannot do given the current state of the technology.

The first problem is as old as governance itself. Decisions are only as good as the information on which they are based. When data is faulty, outdated, or incomplete, any decision-maker, human or algorithmic, is likely to get it wrong.

The Minab strike is a devastating illustration. The school had been clearly separated from an adjacent military compound for at least a decade. The compound's security posts had been removed years earlier. The school was the only operational facility at the site. Yet the targeting database carried a classification from a different era. A human operator consulting that database in real time would have reached the same catastrophic conclusion. The failure was upstream, in the intelligence gathering, verification, and database maintenance processes that preceded any targeting decision, whether made by a human or an algorithm.

The same pattern recurs in civilian AI deployments. Epic Systems' sepsis prediction model, a machine-learning tool implemented in hundreds of US hospitals, was 

shown in a 2021 JAMA Internal Medicine study

 to have missed roughly two-thirds of sepsis cases at Michigan Medicine while generating clinically unmanageable rates of false alerts. The breakdown occurred upstream of the model: the electronic health record data the model relied on at the deployment site was incomplete and inconsistently coded, with key lab values, vital signs, and chart updates not entered in time or entered in fields the model could not parse. A clinician working from the same EHR would have been working from the same compromised information. The failure was not the inference. It was the data that the inference was asked to read.

A clarifying note from the engineering side is worth making here. "Data" in AI systems means two distinct things: the curated datasets used to train and validate a model, and the runtime inputs that feed into a deployed system. Minab was a failure of the second kind. Bias embedded during training, which I address below, is a failure of the first. Both are, in a literal sense, data problems, but they call for very different remedies.

The lesson here is about data, not about autonomy. Whether the targeting decision is made by a human analyst, by an AI system, or by some combination of both, the quality of the underlying information bounds the quality of the decision. Improving the quality, timeliness, and verification of intelligence data is an urgent priority in its own right — a question logically prior to, and separable from, the debate over whether and how AI should be involved in such decisions at all. Better methods of information gathering — improved satellite imagery analysis, pattern recognition across large datasets, faster cross-referencing of intelligence sources — may improve data quality. But no analyst and no algorithm can compensate for fundamentally flawed inputs.

The second problem is different in kind. AI systems themselves — the models, algorithms, and architectures that power them — are not yet perfected. They are subject to limitations inherent in the current state of the technology, not in the quality of the data they receive. This is the part of the conversation that public-policy writing tends to gloss over, often treating AI systems as more capable and more stable than they actually are.

These limitations take several forms. Models trained predominantly on data from specific operational environments can fail catastrophically when confronted with genuinely novel situations. This brittleness or inability to adapt reliably when presented with inputs outside the training distribution can produce outcomes that are not merely wrong but systematically wrong in ways that are difficult to anticipate.

 that a single unlawful use of a civilian vehicle by combatants could lead a system to classify all similar vehicles as legitimate targets. This is the mirror image of Minab. There, the data was wrong and the system processed it faithfully — a human consulting the same database would have reached the same conclusion. Here, the underlying observation is accurate, but the system draws a categorical inference no human analyst would make. Minab is a failure of information. This is a failure of the system itself.

, even when the data appears neutral on its face, through feedback loops, proxy variables, and flawed logical reasoning. Systems trained on surveillance footage and behavioral patterns 

 profiling based on race, religion, or geography. In government contexts more broadly, predictive systems used in 

 have been shown to reproduce and, in some cases, amplify biases present in the institutional data on which they were trained.

Beyond bias and brittleness lies the deeper challenge of opacity. Many high-performing AI systems operate in ways that are insufficiently transparent — even to their designers — to support reliable attribution of outcomes or prediction of behavior under operational stress. When a system produces an unexpected outcome, it may not be possible to reconstruct the chain of reasoning that led to it. That problem becomes especially acute when the outcome is a lethal strike or a denial of critical public services.

 that has long preoccupied nuclear strategists: accountability depends on the ability to identify the source of a particular outcome. Opacity undermines AI accountability for much the same reason that imperfect attribution defeats nuclear deterrence. And some degree of this unpredictability is not a bug to be patched but 

an inherent property of statistical inference itself

 — the trade-off for capabilities that exceed what rule-based systems can deliver.

These are not fringe concerns. They are the everyday operating conditions of contemporary AI. To the engineers who build and deploy these systems, the distinction between 

 is so basic as to be almost trivial. In the public policy debate, somehow, it has remained invisible. Part of the difficulty is that the underlying technology is moving fast enough that policy debate struggles to characterize a moving target — but that is an argument to sharpen the categories, not to blur them.

Conflating these two categories of failure produces three distinct governance errors.

. When a catastrophe results from faulty information, as in the Minab school attack, accountability properly falls on those responsible for maintaining the integrity of the intelligence chain: the analysts, database managers, and verification processes that failed to update a decade-old classification. When a catastrophe results from a system performing in ways its designers did not anticipate, accountability falls on those who designed, validated, and authorized the deployment of the system itself. The actors are different. The failures are different. The remedies are different.

. Imperfect information calls for better intelligence processes, more rigorous verification protocols, and more frequent data updates — reforms that are largely organizational and procedural. Imperfect systems call for continued research, more rigorous pre-deployment testing, greater transparency in model design, and more realistic assessments of whether the technology is mature enough for specific applications. Treating all AI failures as information problems leads to underinvestment in system validation. Treating all failures as system problems leads to unnecessary resistance to technologies that may perform effectively, with better data.

Third, the distinction demands intellectual honesty about the current state of AI. The principle that accountability belongs upstream is sound, but its practical implementation depends on continued progress in making AI systems more robust, more transparent, and more predictable. It also depends on two diagnostic capacities that the field has not yet built.

The first is the proper classification of failure types. The second is a solution to what I have called 

: the field's near-absence of basic measurement infrastructure for counting AI harms against the total number of opportunities that produce them. Without these capacities, even our most ambitious accountability frameworks cannot tell us whether they are working.

Until they are developed, frameworks must explicitly account for both categories of shortcomings, demanding better information and better systems, rather than treating either alone as sufficient. And the distinction this piece draws — between bad data and brittle systems — is itself only the first cut. A mature governance vocabulary will eventually need to separate distrust rooted in bad data, distrust rooted in system brittleness, distrust rooted in opacity, and distrust rooted in the irreducible uncertainty of statistical inference itself.

And there is increasing urgency to this work, because the federal scaffolding that might once have shared it is being withdrawn. A recent Presidential 

 has taken federal preclearance explicitly off the table and adopted an innovation-first posture, with the practical effect that the burden of distinguishing data failures from system failures — and of building the measurement infrastructure that would credibly demonstrate the distinction — now falls on deployers themselves rather than on the agencies that might once have regulated them.

First, AI governance frameworks should establish separate accountability tracks for data integrity and system integrity. A failure investigation should not be permitted to conclude with a finding on one dimension alone without also evaluating the other.

Second, when the data is bad, the algorithm makes it worse. This may sound like common sense, and it is — but policy debates continue to treat AI capability and data quality as separable investments. They are not. A flawed input into a manual decision may produce a single flawed outcome. The same flawed input into an algorithmic system can produce many flawed outcomes, faster and at scale. Algorithms amplify whatever the data gives them. Improving systems without improving data does not produce better decisions; it produces worse decisions, faster.

Third, data-integrity reform and system-integrity research compete for the same budgets and the same political attention, and treating them as substitutable is a false economy. An agency can fund a data-modernization initiative, point to it, and leave the underlying system failures wholly unaddressed — or pour money into model research while the data pipeline quietly decays. Neither investment discharges the need for the other. Both must be funded, and the public conversation must learn to ask for both.

The public debate about AI in government would benefit enormously from this distinction. It is the difference between asking “was the data right?” and asking “is the system ready?” Both questions matter. They are not the same question. Until we can ask each one separately, we cannot answer either one well.

Thanks to I-Jeng Wang of the Johns Hopkins Applied Physics Laboratory for helpful comments.

The World Is Already Resisting AI. Now, There is a List to Prove It.

Brazil's 2026 Elections Are Its First Real Stress Test for AI Regulation

Tech Policy Press - Technology and Democracy

Tech Policy Press is tracking laws and regulations, along with government investigations and litigation, that will shape the rules and accountability for tech companies.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe