Independent New Hampshire Journalism · Northfield, NH

Every AI Company Bans the Biden Robocall. Read the Fine Print Anyway.

The usage policies behind Claude, ChatGPT, Gemini and Grok forbid most of what New Hampshire fears. They also leave the companies room to waive their own rules for the customers who matter most.

Disclosure: This article was drafted with Claude, an AI model made by Anthropic, one of the companies examined here. The editor checked every claim against the linked sources.

In January 2024, thousands of New Hampshire homes picked up the phone and heard a voice built to sound like President Joe Biden. It told Democrats to skip the presidential primary and to “save their vote for the November election.” Steven Kramer, a political consultant, admitted he made the calls. The attorney general charged him with 13 felony counts of voter suppression and 13 misdemeanor counts of impersonating a candidate. A Belknap County Superior Court jury acquitted him on every count that went to trial.

Try to build that call with a mainstream AI model and you break the vendor’s written rules before you break any law. Anthropic’s usage policy, in force November 12, 2026, bans automated calls to voters at scale that hide their artificial origin. OpenAI bars election interference and demobilization. Google bars misleading claims about democratic processes made to deceive.

So the rules look tough on paper. They are also contracts the companies write, interpret, enforce and rewrite on their own, and most of them carry an exit. I read the current usage policies of six companies that build or host frontier AI: Anthropic, OpenAI, Google, xAI, Amazon Web Services and Mistral. The bans mostly make sense. The weak points sit in who the bans stop covering: government customers with negotiated contracts, anyone running a model on their own servers, and the people at the far end of an app built on top.

What the rulebooks agree on

On the worst harms, the six documents read alike. Every one forbids child sexual abuse material. Anthropic, OpenAI, xAI and Mistral name weapons development outright, and Anthropic now spells out targeting software, fire control and arming drones. AWS bars using its AI to perform a lethal function in a weapon without human control. Google’s list has no weapons clause, though it bans facilitating illegal activity. Most of the six bar some form of tracking people without consent.

Four of them require a human check before AI makes a high-stakes call about someone’s job, housing, credit or medical care. xAI goes further and bans high-stakes automated decisions altogether. Voice cloning, the tool behind the New Hampshire call, falls under several headings: OpenAI bars using a person’s voice without consent in ways that could confuse authenticity, AWS bars depicting a voice without consent, and Mistral bars using someone’s voice to impersonate them.

Read as a list of things not to do, these documents hold up. The trouble starts in the clauses about who has to obey them.

Exit one: the government contract

The same Anthropic policy that bans facial recognition, predictive policing and battlefield management also says the company may sign “contracts with certain governmental customers that tailor use restrictions” to an agency’s mission and legal authority. The test for whether that is safe is Anthropic’s own judgment. The policy does not say the tailored terms will be made public.

OpenAI’s version is shorter. Its usage policy bans national security and intelligence work done “without our review and approval,” which means the company holds the switch. When OpenAI signed with the Department of War, it published the key clause: the department may use the system for “all lawful purposes.” The contract keeps a human in control of autonomous weapons wherever law, regulation or department policy requires one. OpenAI’s defense is that its models run only in its own cloud with its own cleared engineers in the loop, and that the contract locks in the surveillance and weapons rules as they stood when it was signed.

That is a serious argument, and so is the one behind it: an elected government with legal oversight should decide lawful national security uses, not a vendor. Follow it through anyway. The rule a New Hampshire resident reads on a company’s website is not necessarily the rule that binds the agency using the same model on that resident.

Exit two: your own server

Mistral, the French company behind several open models, says it plainly. Its usage policy “does not apply to Mistral AI Products deployed on a customer’s infrastructure,” on a partner’s infrastructure, or to its open-source models. Once a model’s weights sit on someone else’s machine, the company that wrote the policy cannot watch what happens next. A use policy can become a license term and a lawsuit. It cannot work as a guardrail.

The case for open release is real. Researchers and small businesses can inspect and run a model without sending their data to a vendor. It also means a voice clone or a disinformation engine can run where no usage policy reaches.

Exit three: the person at the end of the app

Most people never agree to an AI company’s policy directly. They use an app a developer built on top of it. The policies handle that by moving the duty downstream. AWS tells customers that “You and your end users are responsible for all decisions made” with its AI services. Anthropic’s policy makes users responsible for the agents they build, including what those agents do through browsers and connected systems. OpenAI gives developers moderation tools.

The deployer controls the use, so the logic holds as far as it goes. The model maker sees the traffic but not the person, and a two-person shop building a customer-service bot has no compliance department. The company that wrote the rule hands the burden to the party least equipped to carry it.

The rules change when the company says so

None of these documents is fixed. xAI says its policies will evolve as its service changes. Mistral reserves the right to update its policy from time to time and tells users to check back. Google’s list of prohibitions closes with a promise to make exceptions “where harms are outweighed by substantial benefits to the public,” with Google weighing the balance.

Some vagueness is honest. Risk moves faster than legal drafting. Some of it is not. xAI tells users they may use its service so long as they “be a good human,” which is a sentiment, not a standard anyone can enforce or appeal. On appeals, OpenAI’s policy links to a process for challenging an enforcement decision. The other five policy texts I read describe none; most list what the company may do to an account, from a warning up to termination.

One rollback worth watching

Anthropic’s 2026 rewrite narrowed its election rules. In its announcement, the company wrote that it “removed our blanket prohibition on personalized vote and campaign targeting.” Its reason: the old rule caught legitimate civic work, such as nonprofits writing voter information in other languages and election officials sending ballot cure notices. Deceptive targeting and misuse of voter data stay banned under other sections, the company says.

That may be right. It also means campaigns working New Hampshire’s next primary face a looser written rule from that vendor than the one it replaces, with the line now drawn at deception rather than at targeting itself.

What New Hampshire wrote for itself

New Hampshire did not wait for the companies. In 2024 the Legislature passed at least three laws on artificial intelligence.

Put RSA 5-D next to the vendor policies and a gap opens. New Hampshire law lets police use real-time facial recognition in public with a warrant. Anthropic’s and OpenAI’s public policies forbid that use. Which rule governs a New Hampshire agency on a given day depends on its contract with the vendor, and the public text of the policy cannot answer that question.

State agencies already run vendor AI. A July 2025 report by the Department of Information Technology, as the Valley News described it, says the state Department of Justice uses Lexis+ AI for legal research.

The contract question has an answer, and New Hampshire law hands every person the means to get it. Under RSA 91-A, a contract a state agency signs with an AI vendor, including any addendum that changes the vendor’s standard use terms, is a governmental record, subject to exemptions such as the one for confidential commercial information. Every person holds the same right to ask for it. No press pass is required, because the law creates no press tier.

What would make the rules mean something

The bans are not the problem. The exits are. A company that waives its public rules for a government customer should say so in public, even when the contract terms stay sealed, by naming the agency and the rule it set aside. Companies should publish counts of accounts warned, suspended and terminated under each heading, and how many appeals won. A rule change should come with notice measured in weeks, as Anthropic gave this round, and a public changelog, as OpenAI keeps.

For New Hampshire, the test sits closer to home. RSA 5-D already requires the Department of Information Technology to post a yearly report on state agency AI systems. That report should name the vendor terms each system runs under, including any tailored exception to the vendor’s public policy.

The Biden call ended in a $6 million FCC forfeiture order and a jury acquittal. The companies’ rulebooks would have banned it on paper. Paper is the whole issue. New Hampshire residents deserve to know who holds the other end of the promise, and on what terms.

— Dexter Dow, Granite State Report

How the six policies handle the exits

Table: Granite State Report reading of each company’s current policy text. “None described” means the policy document itself contains no appeal process; a company’s separate terms of service may.

Fact check

Editor’s note. Every factual claim above was checked against the sources listed; see the fact-check table. Policy language is quoted from each company’s policy page as published on the dates shown. Corrections: Granite State Report corrects verified errors at the top of the article, with what changed and when.