A year ago this month, during the high-level week of the UN General Assembly, Maria Ressa launched a declaration asking governments to agree on things artificial intelligence should never be permitted to do.
More than 200 people signed it, among them ten Nobel laureates, Geoffrey Hinton and Yoshua Bengio, a former Colombian president, a former Irish president and a former Italian prime minister. It asked for an international agreement by the end of 2026, with enforcement and an independent body to implement it.
There are fifteen weeks left in 2026. No such agreement exists.
What the letter did and did not specify
The declaration's method was deliberate: it asked governments to define the prohibitions rather than defining them itself. It offered candidates — lethal autonomous weapons, bioweapon design, mass surveillance, AI impersonation, generated material depicting child abuse.
That choice was defensible and it is also why the deadline was always going to be hard to meet. A call for others to draw the line cannot be enforced against anyone, and the signatories had no standing to negotiate.
Why prohibition is the hardest instrument to obtain
International law prohibits very little outright, and the prohibitions that exist share a feature: the states that accepted them concluded they were better off if everyone was bound.
Chemical weapons, biological weapons, nuclear testing above ground, blinding lasers — each ban followed either horrified experience or a judgement that the weapon's military value was low relative to the risk of proliferation. Neither condition holds for artificial intelligence, whose civil and military uses are the same research programme, and whose value to a state that defects is enormous.
The rejection was explicit
The United States did not let the question sit unanswered. Michael Kratsios said the administration "totally reject[s] all efforts by international bodies to assert centralized control and global governance of AI."
That sentence is the end of the road for the letter as drafted, because a red line that does not bind the country hosting most frontier development is a statement of preference. The European Union has its AI Act, which is regional and regulatory rather than prohibitive, and China has its own domestic rules. What does not exist is the thing the declaration asked for.
Where the binding rules are actually being written
Not in treaties. They are appearing in procurement rules, export controls, liability law and the terms on which compute is sold — instruments that bind because a company needs a licence, a chip or a customer.
That is less satisfying than a prohibition and considerably more effective at present. It is also where the commercial pressure is heading: the same industry now placing advertisers' agents inside its assistants is establishing facts on the ground faster than any negotiation can describe them.
Regulation is not prohibition, and the difference is the point
The obvious rejoinder is that Europe already regulates this. The EU's AI Act is real, extraterritorial in effect and expensive to comply with — and it is a different instrument entirely.
Regulation sets conditions: assess the risk, document the system, register it, accept audits. Prohibition says a thing may not be built at all, by anyone, with no compliance route available. The declaration asked for the second and the world has been producing the first.
That is not a failure of nerve so much as a reflection of what states can agree. Conditions can be negotiated down and traded off; prohibitions cannot, which is why they are rare and why they hold when they exist.
The implementation body was the serious part
The letter's most substantive request was the least quoted: an independent organisation to implement whatever was agreed, with enforcement.
Every prohibition that works has one. Chemical weapons have an inspectorate with a right of entry, nuclear safeguards have accountants who count material. Without that machinery, a red line is a sentence in a communiqué, and the states that intended to observe it would have done so anyway.
What to watch
Whether anything is tabled before the end of the year, and in what form. A UN General Assembly resolution would be an achievement of sorts and would still not bind anyone.
The more informative signal is narrower: whether any two major powers agree a single specific prohibition, on autonomous nuclear launch authority for instance. One narrow rule that binds is worth more than a comprehensive declaration that does not, and it is the only route by which the second ever becomes the first.
The launch of the Global Call for AI Red Lines on 22 September 2025 by Maria Ressa at the 80th UN General Assembly, its organisation by the French Center for AI Safety, The Future Society and the Center for Human-Compatible Artificial Intelligence, the figure of more than 200 initial signatories including ten Nobel laureates and the named individuals, the request for an international agreement by the end of 2026 with enforcement and an independent implementation organisation, the absence of any binding global red lines, and the quoted rejection by White House official Michael Kratsios are as described in the public record of the declaration. The count of fifteen weeks is our own arithmetic from 17 September to 31 December 2026. The analysis is our own.
Topics aigovernanceregulation
