OpenAI and Anthropic at the UN: What the AI rules debate means for deepfake evidence

What happened at the UN on 23 September

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei addressed the UN Security Council at a meeting on AI and international security on 23 September 2026. Altman called for complementary national and international standards for powerful AI systems. Amodei warned that poorly managed AI could pose a risk to humanity and said no company or country could manage the challenge alone, according to Reuters.

Their comments were about the development and control of increasingly capable AI. The Council did not pass a new deepfake law at this meeting. No UN rule now requires a business to buy an AI detector, issue a certificate for every image, or use a particular watermark. That distinction is worth making before turning a prominent speech into a compliance claim.

What OpenAI proposed, in plain terms

Altman described shared ways to measure model capabilities, assess risks, judge safeguards and maintain meaningful human oversight. He also called for incident-reporting protocols and secure channels for governments, infrastructure operators and experts to exchange information about emerging threats. In a policy paper published two days earlier, OpenAI said governments should decide whether and how to use such technical standards in domestic law.

These proposals concern frontier AI development, particularly more autonomous systems. They are not a draft specification for identifying deepfakes in social posts. There is a useful operational lesson, however: if different teams cannot describe an incident in the same terms or compare the evidence behind their conclusions, they cannot reliably respond together.

• Read the existing EU deepfake labelling guide

Synthetic media is where the abstract debate becomes ordinary work

Imagine an insurer receiving a photograph of a damaged vehicle, a marketplace reviewing a seller video, or a finance team receiving an executive’s recorded instruction. The immediate question is rarely ‘is AI dangerous?’ It is ‘where did this file come from, what does it claim to show, and what should we check before acting?’

A real photograph can carry a false caption. A generated image can be clearly disclosed and harmless. A missing Content Credential does not prove a file is authentic or synthetic. A detector score may help prioritize review, but it does not identify the uploader, verify a person’s identity or establish fraud. Keep these questions separate in the case record.

• Check image and video detector scope
• Review Content Credentials and their limits

The EU already has a rule; the UN discussion did not change it

For teams serving Europe, Article 50 of the EU AI Act has applied since 2 August 2026. The European Commission explains separate duties for providers to mark certain AI-generated output in a machine-readable form and for deployers to disclose deepfakes to people who encounter them. The exact obligation depends on who provides the system, who publishes the media and how it is used.

The 23 September Council discussion did not amend Article 50 or create an equivalent law for the UK or the US. Nor does a detection report certify Article 50 compliance. If a company publishes synthetic media, it must assess its own disclosure duty; if it receives suspicious media from somebody else, it needs an investigation and evidence-retention process. Those are different jobs.

• See the full Article 50 explanation

A practical media review that another person can repeat

Before the file is re-encoded or the post disappears, save the best available copy and record the original URL, account, caption, timestamp and time zone. Write down the claim the media is supposed to support. Compare it with independent material from the same event, sender or transaction. Then inspect any available provenance and run the appropriate image or video check on the preserved file.

Log the file hash, the tool and version used, the date of the check, the result and any conflicting evidence. Put the analyst’s interpretation and the final human decision in separate fields. If the signal is inconclusive or the source cannot be established, say so. The report should let a colleague find the original material and challenge the conclusion without recreating the whole investigation.

• See a sample report
• Explore JSON and PDF reporting

• Save the earliest lawful copy, its URL, context and collection time.
• Record what the file is claimed to prove and check that claim independently.
• Preserve the file hash and any provenance information; note if either is unavailable.
• Treat detector output as a probabilistic signal and keep conflicting findings.
• Name the person who reviewed the case and record the reason for the decision.

What this means for businesses this week

No procurement deadline came out of the UN meeting. The near-term task is to make existing reviews intelligible: name the file, preserve its history, show where each finding came from and allow a person to stop a decision when the evidence is weak. For teams that receive many images, a batch workflow can help keep reports attached to the correct files; it cannot turn uncertain inputs into certain facts.

If international standards on advanced AI mature, incident reporting and shared definitions may become more consistent. That is a possibility, not a rule announced on 23 September. Today, the stronger response is a review trail proportionate to the decision at stake, with the applicable law checked in the relevant jurisdiction.

• Check a batch of images
• See how the evidence API works