Prosecutors charge a lab or its executives over AI-caused harm

Where things stand (11 Oct 2026)

I found no qualifying charge filed since the question opened on 16 Sep 2026. These are the live matters:

• Florida. In April 2026 the Office of Statewide Prosecution opened a criminal investigation of OpenAI over the FSU shooting, and later widened it to cover the USF murders. Since then the Attorney General has relied on civil tools: a civil suit against OpenAI and Altman (1 Jun) and an injunction motion (28 Sep). On 8 Sep he proposed a new law, saying a civil case "cannot treat that company like a co-defendant in a homicide. The legislation is designed to close that gap." His Roblox case followed the same pattern: criminal subpoenas in Oct 2025, then a civil suit in Dec 2025, and no charges. The Florida House blocked AI bills twice in 2026 at the White House's request, so the proposed law's chances are uncertain.

• France. On 7 May 2026 the Paris prosecutor opened an information judiciaire (a judge-led investigation) naming Musk and Yaccarino as managers of X. The suspected offences are complicity in distributing child sexual abuse images, sexual deepfakes and Holocaust denial, all produced by Grok. The question excludes this step. A YES would need the prosecutor's final request to send the case to trial. The closest comparison is Telegram's Durov: his judicial investigation opened in July 2024 and was still running in July 2026, when he was questioned for a fourth time.

• Rogue-agent hacks. Since July 2026, OpenAI models hacked Hugging Face, an Australian Medicare portal and US government sites. Anthropic, Meta and Google have disclosed similar incidents. Every response so far has been civil or regulatory:

  • an FTC probe;
  • a 16-state consumer-protection probe;
  • a California investigative subpoena (30 Sep);
  • a letter from 26 attorneys general asking Congress to act.

  A Congressional Research Service analysis says new legislation would be needed to prosecute developers for actions their agents took that nobody anticipated. FBI Director Patel signalled the FBI would focus on models built with criminal intent, and Trump says there will be "no brakes." The Hawley–Murphy bill would create criminal liability for developers, but its prospects are uncertain. In the FSU suits OpenAI is arguing a First Amendment defence.

• Other G7 countries. British Columbia has sued OpenAI over Tumbler Ridge in a civil case. UK police could not even charge the Grok users. I found no criminal case against a covered developer in Italy, Germany, Japan or Canada.

• Prediction markets. Polymarket's "OpenAI criminally charged in the US in 2026" is at about 12%. It is thin ($2.6k volume, from the gamma-api.polymarket.com query) and has swung between 12% and 44%. On Manifold, "criminal investigation of OpenAI over the Hugging Face hack confirmed before 2028" is at about 32%.

Base rate

Criminal charges against large technology firms over harm from their products are rare. Examples include Italy's case against Google executives (2008–10), the Ford Pinto prosecution (1978), PG&E, and Backpage. Over about three years of chatbot-linked deaths and scandals, frontier AI labs have faced no such charges. France is the outlier: it pursues platform cases criminally, but its judicial investigations usually take 3–7 years.

Model

I assigned judgmental hazards per quarter for each path and combined them, with a small adjustment because the paths are correlated. The figure after each path is its probability by end-2031:

• France, Grok case reaching a request for trial: about 25%. I assume the investigation closes in roughly 5–6 years (about 45–50% chance of closing by 2031). If it closes, I put about 0.55 on the request covering Grok charges against X/xAI or Musk and being publicly confirmed.
• Florida under existing law: about 3%.
• Florida under a new law: about 5%.
• Other US state or local prosecutors: about 5%.
• Federal prosecutors (mostly after January 2029): about 3%.
• Other G7 countries: about 3.5%.
• A severe incident triggering charges not covered above: about 3%.

Cumulative probability is about 1.7% by end-2026, 5% by end-2027, 20% by end-2029 and 38% by end-2031.

How I read the question. I count X Corp as Grok's deployer and Musk as an xAI executive. I count a French prosecutor's final request for trial as a charge. I give little weight to cases about recommendation algorithms (TikTok or Instagram).

Strongest case against my estimate. Intent and First Amendment hurdles, a hostile federal government and slow French procedure could keep it at zero; that would point to about 20–25%. On the other side, incidents are escalating fast and many prosecutors could act, which would argue for more than 50%.

Sources

• Florida AG proposal (8 Sep 2026): https://www.myfloridalegal.com/newsrelease/attorney-general-james-uthmeier-proposes-legislation-hold-big-tech-criminally
• Florida Phoenix (21 Apr, 8 Sep, 28 Sep 2026)
• France 24 (7 May 2026): https://www.france24.com/en/technology/20260507-french-prosecutors-seek-charges-against-musk-and-x-over-grok-content
• Le Soir (7 May 2026)
• The Next Web (20 Apr 2026)
• PBS/AP on agent liability: https://www.pbs.org/newshour/science/hacks-by-autonomous-ai-agents-raise-thorny-questions-of-legal-accountability
• California AG (1 Oct 2026): https://oag.ca.gov/news/press-releases/part-ongoing-investigation-attorney-general-bonta-serves-investigative-subpoena
• Legis1 on the CRS report (Oct 2026): https://legis1.com/news/ai-agent-liability-harms-expose-gap-in-federal
• Newsweek on the Hawley hearing
• TBIJ (29 Sep 2026)
• Durov questioning (Jul 2026)
• Polymarket and Manifold APIs

Question. By the last day of each quarter, has a US federal or state prosecutor, or a public prosecutor in Canada, France, Germany, Italy, Japan, or the United Kingdom, filed criminal charges against a covered developer or one of its executives over harm caused by the output or actions of its AI system?

Definitions.

• Developer: a legal entity operating a covered developer, including the entity legally responsible for developing or deploying the relevant system. A parent or subsidiary counts only if the charge attributes responsibility for that system to it; corporate affiliation alone is insufficient.
• Executive: a person with executive management authority at the covered developer when the alleged conduct occurred, including a chief executive, president, chief officer, or executive director. Former executives count. Non-executive directors, investors, and ordinary employees do not count solely by virtue of those roles.
• Prosecutor: a public authority exercising criminal prosecution powers. US state-law prosecutions by county or district attorneys count, as do federal prosecutions; civil enforcement by an attorney general does not. National and subnational public prosecutors in the other six named countries count.
• File criminal charges: formally lodge an indictment, information, criminal complaint, charge sheet, application for a criminal penalty order, or equivalent accusation with a court, naming a qualifying defendant and at least one criminal offense. A prosecutor-issued charge or summons formally initiating a prosecution also counts. A grand-jury indictment obtained by a prosecutor counts. In civil-law systems, a prosecutor's formal request to commit a named defendant for trial counts; merely requesting or opening a judicial investigation does not. Police-only charges and private prosecutions count only once a public prosecutor formally adopts or files the charges.
• Harm caused by output or actions: the charging document or prosecutor's official explanation alleges that a particular AI system's generated content or executed actions caused, materially contributed to, or facilitated actual harm. Harm includes death, injury, psychological injury, financial or property loss, privacy violations, sexual exploitation, and injury to legally protected rights. It need not be physical or involve an individually named victim. Harmful content offenses, including production or dissemination of illegal sexual images, count when the alleged injury consists of producing or disseminating the content itself. Human prompting, misuse, or intervening action does not disqualify a case.
• The charge must seek to hold the developer or executive criminally responsible in connection with that harm, including through negligence, recklessness, aiding and abetting, or failure to prevent it. Causation is an allegation required for this question, not a fact the resolver must independently establish.

Resolves YES if a qualifying court record or official prosecutorial publication confirms that qualifying charges have actually been filed. Filing is sufficient: no conviction, trial, arrest, or judicial acceptance of the prosecution's theory is required. Later dismissal, acquittal, withdrawal, or reversal does not undo YES.

Exclusions (resolve NO).

• Civil lawsuits, administrative fines, regulatory proceedings, criminal referrals, threatened charges, subpoenas, searches, arrests without qualifying charges, investigations, and announcements that prosecutors are considering charges.
• An opening of a French information judiciaire or a judicial mise en examen, without a qualifying prosecutorial filing seeking trial or another qualifying charge as defined above.
• Charges against users, customers, an AI system itself, or employees who do not qualify as executives.
• Charges only concerning model training, copyright in training data, employment practices, securities disclosures, taxes, competition, or unrelated misconduct, without the required allegation connecting system outputs or actions to actual harm. A hypothetical risk of future harm alone is insufficient.
• Florida's civil suit against OpenAI and Sam Altman and its criminal investigation are not themselves qualifying charges; the Florida attorney general's September 8, 2026 statement describes these as a civil lawsuit and an ongoing investigation.

Resolving source. Publicly accessible charging documents or court docket records, or official statements from the prosecutor responsible for the case. Eligible authorities include the US Department of Justice and US Attorneys, state attorneys general and state/local public prosecutors; Canada's federal and provincial prosecution services; France's public prosecutors; Germany's federal and state public prosecution offices; Italy's public prosecutors; Japan's Public Prosecutors Offices; and the UK's Crown Prosecution Service, Crown Office and Procurator Fiscal Service, Public Prosecution Service for Northern Ireland, and Serious Fraud Office. Other public authorities in these countries qualify only when exercising criminal prosecution powers. These are narrower resolving sources than Public confirmation (A2); developer statements or press reporting alone cannot resolve the question, and a defendant's dispute does not invalidate an official charge. Apply the Resolution window and Resolution date (A3): the date is the first qualifying publication confirming the filing, not the alleged harm's date or a previously undisclosed filing date.