A Hundred and Thirty Countries Want to Ban Killer Robots. The Ones Building Them Do Not.

The UN gave the world until December 31, 2026 to outlaw weapons that pick their own targets. Here is why an overwhelming majority cannot deliver it, what the Geneva rulebook actually says, and the four things you can do before the clock runs out.

I have started this piece four times, and four times I deleted the opening paragraph, because every single one of them had a robot in it. (Then I asked one to help me write it, ironically.)

That is the trap, and I want to name it before I walk into it again. Say the words "killer robots" out loud and a picture arrives before the argument does — chrome skeleton, red eye, some rubble. That picture does work the facts have not earned, in both directions at once: it makes the danger feel enormous and fictional, which is precisely the combination that lets a serious person put the subject down and go read about interest rates instead.

So I am going to stay with the paperwork instead. Resolutions, verbs, vote tallies, meeting calendars, export percentages. No rubble. Because the story here is not a machine that decides to kill you; it is a filing system arranged so that nobody ever has to decide anything at all. That turns out to be the more unsettling document.

First, the part that nobody actually disputes

The subject is what diplomats call lethal autonomous weapons systems — machines that, once switched on, can select a target and fire on it without a person making that specific call. Not a drone with a pilot in a container in Nevada. Not a missile a human aimed. A system that closes the loop by itself.

Almost everybody says they do not want that. Stop Killer Robots counted a hundred and thirty states in March of this year that have publicly backed negotiating a legally binding treaty on autonomy in weapons. Human Rights Watch has been counting the same majority since early last year, in its more cautious phrasing: more than a hundred and twenty countries on the record. And when the question reached the floor of the General Assembly in December 2024, the American Society of International Law recorded the tally at 166 in favor, 3 against — Belarus, North Korea and Russia — with 15 abstentions.

One clarification, because the shorthand in my own headline is doing some lying. Almost nobody is asking for a flat ban on military autonomy. The mainstream ask is a two-tier instrument: prohibit systems that target people and systems no human can meaningfully control, and regulate the rest. "Ban" is headline compression. The proposal is closer to a building code than a bonfire.

The moral case was stated jointly, at the top. The Secretary-General and the president of the International Committee of the Red Cross put their names to the same sentence: the autonomous targeting of humans by machines is a moral line that we must not cross. Months earlier, the Secretary-General's New Agenda for Peace had given the world three years — conclude, by 2026, a legally binding instrument prohibiting autonomous weapons that function without human control or oversight. That is not an activist's framing; the UN Office for Disarmament Affairs records the same recommendation in its own dry summary of the field. As I write, in late July 2026, that deadline has not expired. It runs out on December 31.

So: an overwhelming majority, a hard date, and the moral authority of both the UN and the Red Cross. Here is where things get interesting.

What the majority actually produced

Read the verbs. That is the whole trick, and it takes ten minutes.

Last December the General Assembly adopted two texts. The resolution on artificial intelligence in the military domain reaffirms, notes, invites, encourages, encourages again, and then decides — decides to convene informal exchanges in Geneva for states to further share perspectives. One paragraph asks states to cooperate on a voluntary basis. Not a single obligation lands on anyone. The companion text, the resolution specifically on lethal autonomous weapons, calls upon states to keep working towards a set of elements for an instrument "with a view to future negotiations." With a view to. Two and a half years into a three-year deadline, the majority's own paperwork still described negotiations as something that might happen later.

(Keep those two apart. They are different tracks with adjacent numbers, and the June 2026 meeting in Geneva was held under the first one — explicitly focused on areas other than lethal autonomous weapons. Every aggregator that told you the world just met about killer robots got that wrong.)

That June session is instructive anyway. Arms Control Today reported from the room that Russia's representative was adamantly opposed to any binding measure, and — the detail I keep returning to — that the United States was among the countries that simply did not attend. Not vetoed. Not objected. Did not come.

Meanwhile the actual treaty work happens elsewhere, in a Geneva body called the Group of Governmental Experts, under the Convention on Certain Conventional Weapons. That forum has one design feature which decides the entire story: it runs on consensus. The Arms Control Association puts it in a flat line — consensus-based decision-making allows individual states to block progress. Any one government, out of a hundred and twenty-eight, can stop the room.

The result is visible in the record. The expert group has met every year since 2017 and has produced no binding outcome; in some years it could not agree on how to describe its own discussions in a report. The schedule for the deadline year allots ten days on the calendar: five in March, five more from August 31 to September 4. Ten days to conclude a treaty the Secretary-General asked for in 2023.

Who holds the veto — and the part of this that is my argument, not a fact

The literature has a polite word for the blocking bloc. The Lieber Institute at West Point calls them traditionalists: states that consider existing international law sufficient. The three it names are Israel, the United States and Russia.

The American position is on the record, and not coy. The state-positions monitor quotes Washington directly: "we remain unconvinced that a new legally binding instrument is needed." The reasoning, set out in the Pentagon's own vocabulary, is that existing humanitarian law is enough if it is implemented properly, and that any fixed threshold draws an arbitrary line that ignores operational context.

Now the overlap — and I want to be scrupulous about what kind of claim I am making. SIPRI's arms-transfer data puts the United States at forty-two per cent of all international arms transfers for 2021–25, more than four times the next supplier. Russia is at 6.8 per cent, Israel at 4.4. And the autonomous and near-autonomous systems the American Society of International Law lists read like a map of the same three flags: Phalanx from Raytheon, HARPY from Israel Aerospace Industries, Lancet-3 from a Kalashnikov subsidiary.

The three states blocking the treaty are the world's largest arms exporter, its third largest and its seventh — and the home addresses of the relevant production lines. That is a fact. That their commercial interest is why they block is my inference, and I will not dress it up as anything else — the most the Arms Control Association will say is that states actively developing these systems have shown limited interest in negotiations. The mechanism is documented. The motive is argued.

The strongest case against everything I just wrote

Now let me try to demolish my own piece, because the counter-arguments here are better than the ones I usually meet.

One: the treaty you want might be worse than no treaty. Charlie Trumbull argues in Lawfare that a maximalist instrument could end the discussion at the expert group altogether and reduce transparency among the very militaries that refuse to ratify. On this reading the sensible goal is an instrument the major powers could realistically join — which means a weaker one. Half a loaf, refused by the campaigners.

Two: you cannot verify it. This is the serious objection. The Center for a New American Security's analysis points out that AI is a general-purpose enabling technology — closer to electricity than to the expanding bullet — and that a system's cognitive attributes are not externally observable. You can count submarines. You cannot count intentions encoded in software, and states will not sustain mutual restraint they cannot confirm.

Three: it has not happened yet. Stop Killer Robots' own March briefing concedes that there have been no verifiable reports of a weapons system that both autonomously selected and engaged a target — while noting that the technology exists and manufacturers advertise it. CSIS's assessment of Ukraine's actual capability found, in early 2025, that autonomy is not yet present on that battlefield and engagement decisions remain squarely in the human domain. David Hambling's reporting on Ukraine's drone fleet describes target-lock, not target-choice — the pilots themselves are reluctant to hand over control. And the most-cited counter-example, a 2021 UN Panel of Experts finding that a Turkish-made loitering munition attacked retreating fighters in Libya autonomously, is contested — the manufacturer protested the finding.

Four: sovereignty. Rachel Bovard of the Conservative Partnership Institute puts the case plainly: if existing international law is sufficient, that is what should govern, and subjecting national sovereignty to international dictates carries lasting unintended consequences. You need not agree to see that it is a coherent position, sincerely held.

Take all four seriously and the picture changes shape. Not: machines are killing people unsupervised while diplomats dither. Rather: the capability is built, advertised and for sale; the restraint is a matter of custom and pilot preference; and the window for writing rules before the first verified incident is the one we are standing in. Rules written after are called responses, and they are always worse.

The delegation that has no factories

Here is the thing about the majority that nobody points out: it is not a majority of the powerless.

Austria is the country that put this on the General Assembly's agenda, and its own foreign ministry keeps the receipts: in 2023 it initiated the first-ever First Committee resolution on autonomous weapons, the one 164 states supported, then put the follow-up forward with twenty-seven co-sponsors — the text behind that 166–3 tally. In between it hosted a conference in Vienna where a hundred and forty-four states came, the largest gathering on this subject ever held outside the UN, and the Austrian foreign minister told the room it was "the 'Oppenheimer moment' of our generation." Unglamorous procedural work, all of it — which is how a campaign becomes a resolution with a number.

Brazil is the state that stood up at the expert group in September 2025 and read out a joint statement for forty-two countries — later forty-six — declaring themselves ready to negotiate now. A third of the room.

Now notice where those two countries are not. Neither appears on the manufacturer side of that list: Phalanx, HARPY and Lancet-3 belong to Washington, Tel Aviv and Moscow, not to Vienna and Brasília. That is the whole structure of the problem in one line. Moral authority has accumulated where the production lines are not, and the veto sits exactly where they are. The same fact, told twice.

There is precedent for what happens when the majority stops asking permission. The anti-personnel landmine treaty was not written in a consensus forum; it was written by the states that wanted it and handed to everyone else as a finished norm. The Mine Ban Convention now has a hundred and sixty-one states party as of March 2026, without the United States, Russia or China — and its stigma reached countries that never signed, some of which restricted use anyway. Cluster munitions went the same way. Erin Hunt's account of that decade holds the detail I find most encouraging in this entire field: representatives of non-signatory states that had used the weapons were reduced to asking campaigners to please stop drawing attention to it. Embarrassment did what ratification could not.

Just imagine the incident report

Let me skip the chrome skeleton and imagine the paperwork instead. That is where this actually goes.

It is 2031. A mid-sized state — not one of the three, a customer of one of the three — installs a perimeter-defense system along a contested border strip. It ships with a supervised mode and an unsupervised one, and the manual is explicit that unsupervised operation is for saturation scenarios where the operator link degrades. On a Thursday, the link degrades. Nineteen minutes later there are eleven dead, and at least four of them are herders.

Then the part that ought to frighten you. The inquiry cannot establish who decided. The operator was absent, so no order was given. The commander authorized a mode, weeks earlier, under a doctrine written by people who have since left the ministry. The vendor points to the manual. The software's decision cannot be reconstructed: the model that ranked those eleven heat signatures was retrained twice after deployment, and its training data belongs to a company in a third country that calls it commercially confidential. Every party behaved lawfully according to its own paperwork, and eleven people are dead, and there is no one to charge.

Now imagine the diplomatic response. An emergency session. Expressions of concern. A resolution that reaffirms, notes, invites and encourages. And a proposal to convene informal exchanges.

That is not science fiction. It is the current filing system run forward eight years with one new fact in it. The reason to write the treaty now is not that the robots are coming; it is that accountability has to be built before the event, because afterwards the only materials available are lawyers.

What the people who study this for a living actually say

They do not agree, and the disagreement is the useful part.

Stuart Russell named the structural problem four years ago: efforts stall because Washington and Moscow argue a ban is unnecessary, and restating your own position more energetically is useless "especially when the others have all the cards." That was 2022. It has aged into a description of the present.

Bonnie Docherty has been making the human-rights case from Harvard's International Human Rights Clinic since before there was a campaign — that these systems endanger rights in peacetime policing and border control, not only in war. (She also advises Human Rights Watch; treat those two as one voice, not two.) A briefing filed for the June exchange adds the procurement point: competitive pressure is putting military AI into use before its performance can be independently evaluated.

And on the other side, Paul Scharre, who wrote the book on this — Army of None — is the field's most credible skeptic of easy arms control, and his objection is technical, not political. When your most serious opponent is worried about verification rather than defending the weapons, you have a design problem, not a moral standoff. Design problems are solvable.

What does this mean for you?

You are not going to negotiate a treaty this year. Here is what is actually within reach.

• Fix the numbers in your head. 130 states backing negotiations by one campaign's count, more than 120 by another's, 166 votes at the General Assembly on a procedural text — and 46 states actually willing to start negotiating now. Those measure different things. Anyone who blends them is not being careful with you.

• Watch two dates, not the news cycle. The expert group's final 2026 session runs August 31 to September 4, and the Convention's Seventh Review Conference in November decides whether the mandate continues at all. If a treaty moves this year, it moves there. Reading this later? Look up what came out of them.

• Ask your own government's position out loud. Every state has one, most in writing, and most citizens have never seen theirs. Write to your representative and ask: does my country support negotiating a legally binding instrument on autonomous weapons, and if not, why not? Specific questions get answers. General concern gets a form letter.

• Refuse both easy stories. Not "the machines are already killing on their own" — that is not established, and Stop Killer Robots says so itself. Not "this is science fiction, come back in twenty years" — the systems are built, advertised and exported today.

• Notice the same pattern at home. A system nobody can be held responsible for is not a military phenomenon. It is the architecture of the algorithm that denied your claim and the model that screened out your application. The battlefield version is only the one where the consequence is final.

The lesson, as I see it

Here is what unsettles me, and it is not the hardware.

We built an international system for stopping the worst weapons, handed the key to the states most likely to build them, and called that arrangement consensus. It sounds like agreement. It functions as a veto. And it means the most widely supported humanitarian proposal of the decade has spent nine years "under consideration" in a room that meets ten days a year.

The deadline runs out on December 31, 2026, and I will say plainly what I expect: it will pass without a treaty. That is not the end of the story, though. The landmine and cluster-munitions treaties were written by countries that stopped waiting for permission. The great powers never signed, and the weapons became shameful anyway.

So the real question arriving in November is not whether the three holdouts will change their minds. They will not. It is whether the hundred and thirty will stop asking them to — a majority that has not yet decided to act like one.

My vote? Stop negotiating with the veto. Start writing the treaty.

The HAIA Foundation argues for keeping human judgment inside the systems that decide human outcomes — in a hospital, in a hiring queue, and yes, on a border strip at three in the morning. If that is your kind of thing, subscribe and stay with us for the unglamorous parts.