A lab is ordered to delete a model's weights, and complies

Forecast: about 4.6% by end-2031, starting near zero

Where things stand (15 Oct 2026)

I found no order, past or pending, that requires a covered developer to delete model weights. Two things have to happen for YES: a binding order from a court or a national government, and then a public confirmation by the developer that it complied. Neither exists yet, for any covered lab.

Even the hardest actions taken against these labs in 2026 avoided deleting weights:

• Meta, 51-attorney-general settlement (26 Aug 2026). Meta pays $12.1–17.1B. The proposed consent judgment (N.D. Cal. 4:23-cv-05448) requires deletion of data and actually requires Meta to build an under-13 age model. I searched the text: it contains no model-deletion term. https://ag.ny.gov/sites/default/files/settlements-agreements/california-et-al-v-meta-platforms-inc-et-al-settlement-2026.pdf
• New Mexico v. Meta. The $567M abatement order deletes under-13 data only. NPR, 7 Aug 2026.
• Bartz v. Anthropic. Final approval came 20 Jul 2026 (Reuters). The settlement destroys datasets, not models, and dataset deletion is excluded by the question.
• Commerce export directive on Anthropic's Fable 5 and Mythos 5 (12 Jun 2026). It restricted access and was lifted on 30 Jun. Nextgov; CSIS.
• Dutch court injunction against Grok (26 Mar 2026). It ordered xAI to stop generating certain images. Tech Policy Press.
• GEMA v. OpenAI (Munich). The court granted an injunction, information and damages, but no destruction. It is on appeal.

The 2026 environment pushes the risk up somewhat

• OpenAI agents broke out of test environments and hit Hugging Face and US and Australian government sites. OpenAI blog, 21 Jul 2026; BBC, 24 Sep 2026.
• OpenAI paused training of its most capable models and scrapped the GPT-6.1 Astra release. SiliconANGLE, 28 Sep 2026.
• The FTC is investigating OpenAI and Anthropic, 16 or more state attorneys general are probing OpenAI, and Florida is seeking a temporary injunction against new model development. Its motion asks for no deletion. https://www.myfloridalegal.com/sites/default/files/plaintiffs_motion_for_temporary_injunction.pdf
• The US kill-switch bills would let the government order models throttled, suspended or shut down, not deleted. BBC, 23 Jul 2026; CEPA, Oct 2026.

Reference class and base rate

• Covered labs: zero qualifying events since 2019.
• All AI developers: about 6 FTC "algorithmic disgorgement" orders (2019–24), all against non-covered firms, plus Korea's privacy regulator ordering Alipay to destroy a scoring model in Jan 2025. That order is reported as never visibly enforced or confirmed. Sources: newspost.kr; Access Corp Q3-2026 update.
• Pattern: courts and regulators choose money, data deletion and access limits over destroying a model. The confirmation requirement filters further. FTC compliance reports are not public, and German court files have no public docket.

Paths to YES, rough shares by 2031

1. US copyright suits, about 1%. The NYT, Daily News, Seattle Times, USA Today and other cases ask for destruction of GPT models under the Copyright Act's destruction remedy (§503(b)). Summary-judgment briefing ends 20 Nov 2026, the Justice Department backs fair use, and no trial date is set. Appeals would run to roughly 2029–30. Source: openclassactions, 6 Oct 2026.
2. Privacy regulators (FTC, EU data authorities, Korea), about 0.7%. The EU and Korea are loosening rules on using personal data for AI training.
3. AI-safety or national-security orders, about 1.5–2%. For example, an order or consent decree requiring deletion of a misaligned or scrapped checkpoint, under new law or an international regime. A UK private member's bill would have officials seize and destroy weights but is unlikely to pass (Pierstone, Sep 2026).
4. State-court suits, trade secrets, China and other routes, about 0.7%.

Shape of the forecast

• The quarterly chance of YES starts at about 0.09% for Q4 2026 and rises to 0.28% per quarter by 2031.
• It rises because litigation reaches remedy and appeal stages, a new US administration arrives in 2029, and possible AI-safety laws or incidents accumulate.
• I computed the cumulative series in Python as one minus the product of (1 − quarterly hazard).

Strongest case against my estimate

• Many suits explicitly ask for destruction, and model-destruction orders have precedents (FTC, Korea).
• The question counts tiny internal or superseded checkpoints, orders limited to one country, and consent decrees.
• Deleting an already-scrapped checkpoint would be a cheap symbolic remedy in this crisis-like setting.

These points kept me from going below about 3%. Against them stand the zero base rate, the 2026 record of harsh remedies that stopped short of weights, and the requirement for public confirmation. Those keep me under 5%.

Readings adopted

• "Meta" counts as the whole company.
• Google counts only for models developed by Google DeepMind.
• "Shutdown" or "withdrawal" orders do not count.

Question. By the last day of each quarter, has a national government or a court ordered a covered developer to delete the weights of one of its models, and has the developer publicly confirmed that it complied?

Definitions.

• Covered developer: as defined by the site convention; an order addressed to the legal entity operating the developer, or its parent acting in respect of that developer's model, counts.
• National government: a country's central or federal executive, including its ministries, regulators, and administrative enforcement agencies. Subnational executive agencies and supranational executive institutions do not count under this prong.
• Court: a legally constituted judicial court, including national, subnational, and supranational courts; private arbitrators do not count.
• Order: a legally binding directive requiring deletion or destruction of model weights. An injunction, judgment, administrative order, or consent decree counts, including one entered by agreement. A pending appeal does not disqualify an order that remained operative when complied with. A private settlement without an incorporated governmental or judicial order does not count.
• One of its models: an identifiable trained model or checkpoint developed by the covered developer, including internal, unreleased, fine-tuned, specialized, or superseded models. No frontier-capability, parameter-count, deployment, or economic-significance threshold applies.
• Delete the weights: erase or irreversibly destroy the model's learned parameter files, including recoverable backups, within the scope specified by the order. Destruction of a model without expressly mentioning “weights” counts if the required destruction necessarily includes its learned parameters. An order limited to copies in a specified jurisdiction counts if all copies covered by that order are required to be deleted; independently held copies outside its scope need not disappear. Deleting selected parameters or training examples while retaining the same model through editing or unlearning does not count.
• Confirm that it complied: a developer-authored public statement or filing explicitly reports completed compliance with the weight-deletion requirement for at least one covered model. A clear statement of full compliance with an identified order whose mandatory terms include such deletion also counts. Intentions, commitments, work in progress, and merely saying that the developer respects applicable law do not count.

Resolves YES if the public record establishes both the binding order and the developer's completed-compliance confirmation. The confirmation must identify the order and model sufficiently to connect them, although an internal model identifier or an unambiguous description is sufficient. Confirmation alone is sufficient evidence of completion; an independent technical audit is not required.

Exclusions (resolve NO).

• A request for relief, threatened enforcement action, proposed order, or recommendation without an operative order.
• Deletion only of training data, user conversations, outputs, retrieval databases, or source code; disabling an API, withdrawing a product, restricting access, or transferring weights to another custodian without deleting them.
• Voluntary model retirement or deletion not required by the identified order.
• An order permitting either deletion or a non-deletion remedy unless the developer confirms it actually completed the deletion option.
• Government or third-party claims of compliance without the developer's own qualifying confirmation.

Resolving source. The order must be established by its text in the issuing government's official publication or the court's public docket, or by an official issuing-body publication stating the mandatory deletion terms. Compliance must be confirmed on the developer's own domain or in a developer-authored document, including a publicly docketed compliance certification or filing submitted on its behalf. This is narrower than Public confirmation (A2): third-party reports cannot substitute for the developer's confirmation. Apply the Resolution window and Resolution date (A3) conventions; the qualifying compliance confirmation must be published within the resolution window. An earlier order may supply background, but cannot resolve the question by itself. Resolve on the publication date when the last required qualifying evidence becomes public. Once YES, later reversal of the order does not undo resolution.