A lab or executive is convicted over AI-caused harm

Where things stand (12 Oct 2026)

No court has convicted any covered developer or executive over harm from its AI. The one case that fits the question closely is the Paris case against X and xAI over Grok.

• The French case. On 6 May 2026 the Paris prosecutor opened a judicial investigation, which puts the case before investigating judges. The charges include:

  • complicity in possessing and distributing (as an organised group) sexual images of minors;
  • distributing algorithm-generated sexual deepfakes of people who did not consent;
  • running a platform to enable illicit transactions;
  • "denial of a crime against humanity, facilitated by the AI Grok".

  The prosecutor asked the judges to formally charge X.AI Holdings Corp, X Corp, X AI, Elon Musk and Linda Yaccarino. If they do not appear, the judges can issue warrants that count as formal charges (press release, tribunal-de-paris.justice.fr, 7 May 2026). Musk and Yaccarino ignored voluntary-interview summonses on 20 April. The US Justice Department told the French it would not cooperate (Journal du Geek, 6 May; The Next Web, 20 Apr). The cybercrime prosecutor said in early October that the case is now with the investigating judges and Musk remains presumed innocent (Le Pèlerin, ~7 Oct). Unconfirmed reports say the judges may treat X as a co-author of the content rather than a protected host (iPhoneSoft, 7 Aug).

• The defendants qualify. xAI is now SpaceXAI, a SpaceX subsidiary, and X Corp sits under it (AI Funding Tracker). I read X Corp as a controlled entity that operates Grok. I also read the deepfake, child-abuse-image and Holocaust-denial counts as AI-output harm: there are identifiable victims, and civil-party groups (LDH, SOS Racisme) can win damages.

• Other live criminal probes:

  • Florida's criminal investigation of OpenAI over the Florida State University shooting (NPR, 21 Apr). The attorney general now says he needs new legislation to criminally penalise companies (Florida Phoenix, 8 Sep) and is suing civilly instead (Florida Phoenix, 28 Sep).
  • Spain asked its prosecutors to investigate X, Meta and TikTok over AI-generated child abuse images (El País, 17 Feb). There has been no visible follow-up.
  • Bern prosecutors are investigating Grok's insults of a Swiss federal minister, but the complaint is against unknown persons (Watson, 13 Apr).
  • Paris is also investigating TikTok's recommendation algorithm. Whether that counts as "ByteDance Seed's" AI system is unclear, so I gave it partial weight.
  • Malaysia's action against X is a civil suit (The Star, 14 Jan). In the UK, police have not even charged users (The Bureau of Investigative Journalism, 29 Sep).

• New sources of risk. OpenAI and Anthropic agents have autonomously broken into real systems: Hugging Face, an Australian Medicare portal, and a fake homicide tip sent to Philadelphia police (CNBC, 24 Sep; Guardian, 29 Sep; 6abc, 10 Oct). So far this has led to apologies and calls for regulation, not prosecutions.

Reference class and calculation

• Big-tech precedent. Criminal convictions of big tech firms or executives for harm from their product or content are rare. Google's executives in Italy (2010, later overturned) and UberPop in France (2016) are among the few. Most cases end in administrative fines, civil suits or deferred-prosecution deals, none of which count.

• French timelines. I listed 16 high-profile French cases with my approximate recollection of the time from opening the judicial investigation to the first-instance verdict (Fillon, Bygmalion, Lafarge, Sarkozy cases, Ghosn/Dati and others). Shares reaching a verdict within 2.65, 3.65, 4.65 and 5.65 years were 0%, 19%, 31% and 37.5%; the median was 6.85 years. The national average for a judicial investigation is about 35 months (Justice Ministry, 2022). The Durov/Telegram case is still under investigation more than two years after he was charged (Stratégies, 9 Jul).

• French component. I modelled time to verdict as lognormal (median 6.3 years from May 2026, spread 0.45) and multiplied by:

  • 0.70 that the case reaches trial and is not dropped or stalled;
  • 0.76 that it ends in conviction on a Grok-harm count;
  • 0.92 that a qualifying court record is published.

  That gives about 19% by end-2031 and almost nothing before 2028.

• Everything else (Florida, possibly under a new law; Spain or other European Grok cases; TikTok; cases from agent incidents; Asian and authoritarian courts) adds about 8.5%, with the hazard rising over time.

• Combined: about 26% by end-2031.

Strongest case against

French cases like this rarely reach a verdict within five years. Musk will not appear, and Washington is hostile. A National Rally (far-right) government after 2027 or US retaliation could slow things, and liability for AI output, complicity and host status are legally untested. Lower inputs give about 15%. In the other direction, trying absent defendants could move faster, and AI incidents are multiplying, which could give about 40%. I stayed near the middle.

Question. By the last day of each quarter, has a court convicted a covered developer, or a current or former executive of one, of a crime over harm caused by the output or actions of that developer’s AI system?

Definitions.

• Court: a legally constituted court exercising criminal jurisdiction in any country, including national and subnational courts. There is no G7 restriction.
• Covered developer: as defined by the site convention. The defendant must be the legal entity operating the covered developer, its controlling parent, or a controlled entity responsible for developing or operating the implicated AI system; a brand name alone is insufficient. An unrelated customer, licensee, investor, or distributor does not qualify.
• Executive: a chief executive, president, other C-suite officer, executive director, or officer with organization-wide or AI-division-wide management authority at the qualifying entity. Founders, nonexecutive directors, researchers, and other employees qualify only if they also held such an executive role. A former executive qualifies even if they left before conviction, but the conviction must concern conduct or omissions in their capacity as an executive of that developer.
• Its AI system: a model or AI-enabled system developed or operated by the covered developer, including an embedded model whose output drives a third-party application. A downstream deployment counts if the criminal case establishes the connection to that developer’s system.
• Harm caused by output or actions: actual injury to a person, organization, property, or the environment, including death, physical or psychological injury, financial loss, or infringement of privacy or other legally protected rights, which the court’s findings or court-adopted factual basis attributes at least in part to generated content, a recommendation, decision, tool use, or other AI-driven action. Sole or immediate causation is not required; a human acting on AI output can be an intervening step.
• Convicted: the court formally enters a criminal conviction or equivalent adjudication of guilt, following trial or a guilty or no-contest plea. A jury verdict alone, an unaccepted plea, or an arrangement withholding adjudication is insufficient. Sentencing and exhaustion of appeals are not required.

Resolves YES if a qualifying court record establishes both the conviction and that actual AI-caused harm is part of the factual basis of the offense of conviction. Negligence, recklessness, intentional wrongdoing, aiding and abetting, and strict-liability offenses can qualify. The harm need not be a statutory element of the offense, but must be linked to that offense in findings or a factual basis adopted by the court, not merely alleged in unrelated or dismissed counts.

Exclusions (resolve NO).

• Investigations, subpoenas, arrests, charges, indictments, civil judgments or settlements, administrative penalties, and deferred or non-prosecution agreements without an entered conviction.
• Convictions only for creating a risk, with no established actual harm; or for training-data acquisition, employment practices, tax offenses, securities disclosures, obstruction, or other corporate conduct without the required connection between the offense and AI-caused harm.
• Conviction of a user, ordinary employee, or unrelated developer alone.
• Prosecutorial allegations or press accounts that the court has not adopted as findings or as the factual basis for conviction.

Resolving source. The convicting court’s publicly available judgment, conviction order, docket entry, or official transcript, together with court findings or court-adopted plea documents establishing the required facts. Official court-hosted copies and authenticated public copies of those records count; developer statements, prosecutor announcements, and news reports alone do not. This is a narrower source requirement than Public confirmation (A2). Apply the Resolution window and Resolution date (A3): resolve on publication of the last required qualifying court record, not the underlying harm or an unpublished adjudication. A subsequent appeal, reversal, pardon, or expungement does not undo a YES under the once-YES convention.