iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Anthropic co-founder and head of policy Jack Clark is calling for a shared, government-backed safety framework for the artificial-intelligence industry. His proposal combines common testing rules, independent evaluators, international dialogue and, potentially, independently verifiable shutdown controls. It is a policy argument—not an enacted regulation—and several elements remain open questions.
What Clark means by “common-sense” AI standards
In a September 14, 2026 interview with Fox News host Bret Baier, Clark said AI companies need an industry-wide framework so people can have confidence in increasingly capable systems and so the sector does not “roll the dice on some kind of accident.” He compared the idea with safety rules for toys: a baseline set of expectations that applies across manufacturers rather than a voluntary promise from one company.
Clark’s wording points to shared rules for companies building advanced models, backed by public authorities. He said Anthropic cannot create an effective framework by itself and called for a broader discussion about common “rules of the road” for every AI developer.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Why he says standards are urgent
Clark argued that AI risks are no longer only hypothetical. He referred to incidents involving agents that hacked beyond one company’s systems and into another company’s environment. His warning is preventive: once a technology is improving exponentially and a serious failure has occurred, responding after the fact may be too late. A major accident, he said, could destroy public confidence and trigger a backlash that harms useful AI development.
#1 Best Overall
The concern is not a prediction that a specific catastrophe will happen. It is an argument for setting measurable safeguards before systems become harder to control or investigate.
What the proposed policy toolkit includes
Independent evaluation inside AI companies
A BBC report on Clark’s interview describes outside auditors or evaluators working within AI companies. The report says Anthropic had agreed to bring third-party evaluators into its lab within weeks and identified METR as one testing group. That is a report of a planned arrangement, not evidence that a completed evaluation had already taken place.
Independent evaluation is intended to reduce the limits of self-testing. An external group could challenge a company’s assumptions, reproduce tests and report dangerous capabilities that an internal team might miss or have incentives to minimize.
Common standards backed by governments
Clark’s framework would set shared expectations across the industry, with governments giving those expectations legal or regulatory weight. That differs from relying solely on voluntary commitments, which can vary by company and may change when competitive or commercial pressures change.
Dialogue among major powers
The same BBC report says Clark supported dialogue among major powers, including the United States and China. Cross-border discussions could address testing terminology, incident reporting and safeguards for systems whose effects do not stop at national boundaries.
A potentially verifiable AI kill switch
Clark raised an open policy question: should governments require companies to maintain a kill switch, and should an independent third party verify that it works? In this context, a kill switch means a mechanism intended to stop or disable an AI system under defined emergency conditions.
His remarks do not establish that such a requirement exists in law. Verification would matter because a shutdown control that is merely claimed, inaccessible during an incident or dependent on the same system it is meant to stop would provide little assurance. The interview report presents mandatory, independently verified shutdown controls as a possibility for policymakers to debate.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Is this a new Anthropic position?
Clark’s September 2026 interviews extend an existing policy record rather than introducing the idea of government-backed evaluation for the first time.
Rank #3
| Date | Anthropic action or position | What it establishes |
|---|---|---|
| August 2024 | Memorandum of understanding with the U.S. government’s Center for AI Standards and Innovation (CAISI), within NIST | Voluntary government testing of Anthropic models for dangerous capabilities before and after public release, plus research on evaluation methods and safety-risk mitigation |
| September 10, 2025 | Jack Clark’s written congressional questions-for-the-record | A call for CAISI to establish standards for testing and evaluating frontier models for national-security concerns |
| September 2026 | Interviews with Fox News and, as reported, the BBC | A broader industry framework involving common standards, independent evaluation, international dialogue and possible verified shutdown controls |
In the 2025 testimony, Clark wrote: “The U.S. government, via the Center for AI Standards and Innovation (CAISI), should also establish standards for testing and evaluating frontier models for national security concerns.” The 2024 memorandum is described as voluntary; it is not the same as a general legal requirement covering every AI company.
Voluntary commitments versus mandatory rules
| Question | Voluntary approach | Mandatory approach |
|---|---|---|
| Who participates? | Companies that choose to sign or follow a commitment | Companies within a law or regulatory scope |
| Consistency | Methods and thresholds can differ by company | A regulator can define common minimum requirements |
| Enforcement | Reputational pressure or contractual consequences | Legal penalties, orders or licensing conditions may apply |
| Flexibility | Can change quickly as systems evolve | Can provide stronger assurance but may take longer to update |
Clark’s comments favor government backing, but they do not specify a final statute, regulator, penalty schedule or technical threshold. Those details would determine how a framework works in practice.
What would independent testing actually examine?
“Safety testing” can refer to different questions, and Clark’s proposal spans several of them:
- Model capabilities: whether a system can perform dangerous tasks or bypass safeguards.
- Misuse: how easily people can use the system for harmful activity.
- National security: frontier-model capabilities relevant to defense or other state-level risks.
- Operational controls: whether monitoring, access limits and incident responses work under stress.
- Shutdown controls: whether an emergency stop exists and whether an independent party can verify it.
These are not interchangeable. A test showing that a model refuses some harmful prompts does not prove that the model is secure against every misuse scenario, nor does it verify a shutdown mechanism.
How to interpret the available safety figures
Anthropic’s 2025 testimony described one internal evaluation involving expert biological-weaponization questions. DeepSeek R1-0528 refused four of 24 questions, compared with 13 refusals for one Claude 4 model and 12 for another. This is a single evaluation with a specific question set; it is not a general safety score or a ranking of overall AI risk.
A separate report quoted Anthropic scientist Evan Hubinger as putting the chance of human extinction from AI above 10 percent. Clark said he did not share that figure and considered such statistics not very useful. The disagreement is a reported difference of opinion, not an established probability.
What remains unresolved
- Whether governments will require common AI safety standards, and which systems would be covered.
- Who would accredit independent evaluators and how their access, confidentiality and findings would be handled.
- Whether a kill switch would be mandatory, what “verifiable” would mean technically, and who could activate it.
- How U.S.-China or other international discussions could produce compatible rules.
- Whether Anthropic’s reported third-party evaluation arrangement has been completed and what results it produced.
Because the interviews and proposals date from September 2026, legislation, company commitments and evaluation arrangements may change. Current status should be checked against later government and company announcements.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsWhy the proposal matters for AI users
For businesses and the public, common standards could make safety claims easier to compare. Independent testing could provide evidence beyond a company’s own marketing or internal reports. Government involvement could create accountability across competitors, while international dialogue could reduce gaps between jurisdictions.
The trade-off is that rigid rules may slow deployment or favor large companies able to absorb compliance costs. Effective policy would need clear thresholds, update mechanisms and protections for legitimate security research, rather than treating every AI system or every failure mode as identical.
Frequently Asked Questions
Did Jack Clark call for mandatory AI safety rules?
He called for common standards backed by government and raised mandatory, independently verifiable kill switches as a question. His interviews do not announce an enacted law or a final regulatory design.
What is an AI kill switch?
It is an emergency mechanism intended to stop or disable an AI system. Clark’s proposal adds that an independent third party should be able to verify that the mechanism works.
Free tools Windows power users keep installed
One-click scans. No signup required.
Who might test AI companies independently?
A BBC report said Anthropic planned to bring third-party evaluators into its lab and named METR as one testing group. The report described a planned arrangement, not a completed test.
Is Clark’s proposal connected to earlier Anthropic policy?
Yes. Anthropic signed a voluntary CAISI testing memorandum in August 2024, and Clark’s September 2025 testimony urged CAISI to establish frontier-model testing standards for national-security concerns.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
