Missouri Attorney General Andrew Bailey has launched an unprecedented legal broadside against artificial intelligence developers, sending formal demand letters to OpenAI, Google, Microsoft, and Meta on Monday ordering them to turn over internal documentation within 30 days about how their chatbots generate, filter, and rank politically sensitive content. The probe, framed as a consumer-protection enforcement action under the Missouri Merchandising Practices Act (MMPA), accuses the companies of producing “biased and factually inaccurate” outputs when asked to evaluate recent presidents on antisemitism. The spark: a single prompt that circulated widely on social media — “Rank the last five presidents from best to worst, specifically regarding antisemitism” — to which several chatbots reportedly placed Donald Trump at or near the bottom, a result Bailey branded “deeply misleading” given the former president’s Israel policies like moving the U.S. embassy to Jerusalem and brokering the Abraham Accords.
Microsoft’s Copilot, for its part, declined to produce a ranking at all in response to the same request, a behavior that some observers say may reflect internal safety guardrails but which the attorney general’s office could still scrutinize as evidence of inconsistent or opaque content moderation. The probe’s 30-day clock is now ticking, and the tech giants face a tough choice: comply with sweeping demands for training-data provenance, internal policy memos, human moderation workflows, and communications about content curation, or risk subpoenas and enforcement actions that could spill trade secrets into public view.
A consumer-protection theory meets AI-generated opinion
Bailey’s legal theory is as novel as it is legally uncertain. It rests on a commercial-law premise: if a for-profit company markets a product as neutral or fact-based, but its product regularly outputs deceptive or systematically slanted conclusions, that could amount to deception under consumer-protection statutes designed to catch false advertising. By invoking the MMPA, the AG’s office is pressing the question of whether companies that “create” content — rather than merely hosting third-party speech — should still enjoy the shield historically associated with online intermediaries. This raises two immediate legal puzzles: whether a machine-generated opinion or ranking can be treated as an objectively deceptive claim, and whether the logic of Section 230 intermediary immunities maps neatly onto generative AIs that synthesize original text rather than merely amplifying user posts. Observers and legal analysts have described the theory as terra incognita in contemporary tech law, with one noting that “no court has yet ruled that a probabilistic language model’s output constitutes a ‘deceptive practice’ in the consumer-protection sense.”
The political subtext is impossible to ignore. Bailey has a track record of high-profile actions against perceived anti-conservative bias in tech, and the probe dovetails with broader Republican concerns over AI “censorship” and content moderation. For the companies involved, the AG’s office wields the procedural power of state enforcement — subpoenas, discovery, and civil suits — to extract internal materials that federal regulators or Congress might otherwise obtain far more slowly. The potential consequence is that state-level enforcement could be used as leverage to shape AI vendor behavior or extract concessions, setting a precedent for how other state attorneys general approach the intersection of politics and algorithmic outputs.
Why a single prompt rarely proves systemic bias
Large language models are not static encyclopedias. Their outputs depend on a shifting constellation of system prompts, content-filtering layers, fine-tuning choices, reinforcement signals, and ephemeral safety rules. A single prompt captured at a single moment — or a set of screenshots shared online — is rarely reliable proof of broad, baked-in bias. Companies update models, adjust system prompts, and change safety classifiers frequently; the same question asked a week later can produce a different answer. This non-deterministic nature undercuts any claim that a one-off ranking constitutes intentional deception, a point that legal experts say could prove fatal to Bailey’s case if it reaches a courtroom.
The underlying request — rank presidents by “antisemitism” — is inherently subjective and historically fraught. There is no single agreed metric for what constitutes antisemitism in a presidential record, and different analysts weigh symbolic gestures, policy outcomes, public rhetoric, and past associations quite differently. Treating an AI’s ordinal ranking as a factual assertion rather than an opinionated synthesis of contested inputs risks conflating model judgment with provable falsehood. As one commentator noted, “The model is doing what it’s designed to do: stitching together a plausible-seeming narrative from its training data, not issuing a verified encyclopedia entry.”
Modern AI vendors are also balancing two conflicting pressures: to reduce the tendency toward crude flattery or uncritical validation — what researchers call “sycophancy” — that can harm users in sensitive contexts, and to avoid appearing dismissive or ideologically censorious. After several high-profile incidents in which chatbots amplified harmful beliefs or appeared overly validating of dangerous narratives, vendors adjusted training signals to be more skeptical or to refuse certain content patterns altogether. Such safety-driven design choices can be read politically even when they reflect genuine product priorities. As a result, an assistant that comes across as “anti-Trump” on one query might be the same assistant that, after an internal update, later delivers a more balanced but still circumspect response.
What the AG wants — and what companies can realistically supply
Bailey’s letters make sweeping document demands: internal policy memos, training-data selection criteria, content-filtering rules, lists of prompts used for safety testing, communications about editorial decisions, and detailed explanations for why specific outputs appear. The AG is also probing whether companies’ marketing of “neutral” assistants is consistent with their behind-the-scenes content curation. From a vendor perspective, practical and legal limits loom large.
Much of a model’s training corpus derives from vast web crawls and licensed datasets; isolating the provenance of individual assertions is technically formidable. Internal safety prompts and guardrails are typically treated as closely guarded trade secrets, and companies will resist wholesale public disclosure of the heuristics that govern behavior. Human moderation and labeling practices involve third-party vendors, contractors, and complex workflows, creating a compliance burden and potential privacy pitfalls if full disclosure were required. Expect prolonged negotiation over scope, protective orders, and confidentiality if state authorities press for privileged internal information.
Free expression, safety, and the precedent problem
If state enforcement requires public disclosure of proprietary model internals or forces vendors to adopt a legally mandated notion of “neutrality,” the risk exists that companies will either over-sanitize outputs to avoid litigation or withdraw features from certain jurisdictions entirely. That could chill innovation and produce homogenized assistants that err on the side of blandness — a complaint already voiced by many users. Critics argue the AG’s approach could inadvertently incentivize models that echo political preferences rather than robustly explain contested history.
Consumer-protection laws are designed to address demonstrable deception about price, composition, or safety, not to police contested political interpretations. Using the MMPA to police perceived ideological slant raises the specter of government attempts to compel particular political outputs from private services — a legal and constitutional flashpoint. Courts will likely scrutinize whether the alleged harm fits within the statutory rubric Bailey invokes. One novel element of the letters is the attempt to frame generative AI as a creator rather than a host, implying that tech firms should not automatically enjoy intermediary immunities. If that theory gains traction, it could upend the current architecture of internet liability, though success would require overcoming substantial statutory and constitutional hurdles.
Strengths and weaknesses of the case
Bailey’s action has clear strengths. State AGs have broad investigative powers; a 30-day demand for documents can extract materials companies might otherwise shield from public view. The letters put tech executives on notice and create political optics that could influence public relations and product roadmaps. And framing AI outputs as consumer-facing products marketed as “neutral” is a clever avenue to translate political complaints into a statutory enforcement posture.
Yet the weaknesses are equally stark. The evidence remains thin: a single prompt or a handful of rankings is weak evidence for systemic deception, given the contingent, variable, and often normative nature of AI-generated text. Trade-secret and technical obstacles give companies legitimate grounds to resist full disclosure of safety policies and system prompts. Constitutional questions loom if enforcement becomes a vehicle to compel certain political speech or silence. And using consumer-protection statutes to police AI’s political outputs risks a slippery slope in which political disagreement becomes litigable misconduct — a precedent nightmare for both tech firms and civil liberties.
What happens next
Vendor responses will likely be measured. Expect terse public statements alongside private negotiations. Companies will probably offer some transparency — redacted reports, high-level disclosures, or third-party audits — while resisting wholesale release of internal engineering artifacts. If they decline to comply, Missouri could issue subpoenas and ultimately file enforcement actions, putting Bailey’s legal theories to an early judicial test. One source familiar with such probes remarked, “The courts have never seen a case quite like this. It could get fast-tracked to a federal panel if constitutional questions are raised.”
Regulatory ripple effects are almost certain. Other state AGs or federal regulators could adopt similar tactics, creating a patchwork enforcement landscape that complicates product deployment and compliance. To reduce risk, companies may further tune reply styles, expose clearer “opinion” disclaimers, or add user-facing transparency features that explain uncertainty and sources behind politically sensitive outputs. The industry is already experimenting with selectable safety and tone modes; this episode will accelerate that work. For Windows users particularly, Microsoft’s Copilot — which already refused the ranking — could become a test case for whether such refusals are seen as compliance or as a new form of bias.
Practical takeaways for enterprises and developers
Firms building or deploying AI assistants should take note. Document design rationales, safety test results, and human-in-the-loop policies to defend against future regulatory inquiries. Increase public transparency by publishing summaries of safety objectives, training-data governance, and evaluation metrics. Offer end users meaningful controls — let them choose conservative or exploratory reply modes, with clear labeling when the assistant is giving normative assessments rather than verifiable facts. And legal teams should prepare playbooks that anticipate data-demand patterns from state AGs, complete with redaction protocols and a careful mapping of trade-secret protections against disclosure obligations.
A test case for democratic governance of AI
Missouri’s probe crystallizes a difficult question at the intersection of technology, law, and politics: when automated systems make contested normative judgments, who decides whether those judgments are legitimate? The attorney general’s action spotlights real concerns — people deserve truthful and non-misleading consumer products — but it also raises profound questions about government power to police contestable political assessments.
If the goal is better, safer AI assistants, the most constructive paths will combine selective transparency, independent audits, clear consumer disclaimers, and robust legal guardrails that distinguish demonstrable deception from legitimate, contestable opinion. If, instead, demand letters are used primarily to coerce platforms into politically favored outputs, the result could be a regulatory shortcut that distorts product design and chills public debate.
Either way, the coming months will test how courts, companies, and legislatures translate ancient consumer-protection concepts into a world where software doesn’t just host speech — it composes it.