I spent the better part of a year citing a single number before I understood what it could and couldn't do. The number was 61 percent — the share of respondents, across a widely circulated multi-country survey, who said they opposed the development of Autonomous Weapons Systems. I used it the way most people in this debate use it: as the empirical floor under a legal argument. If a clear majority of people across many countries recoil at the idea of machines selecting and killing human targets, then surely that recoil is evidence of something the law already recognizes — a violation of the public conscience.
Then a colleague who studies survey methodology asked me a question I couldn't answer. She did not ask whether the public opposed these weapons. She asked: opposed compared to what, measured how, and would the same person say the same thing if you described the alternative honestly. I went looking for the answers and came back less certain than I started. This is the sixth lesson nobody teaches you about this debate, and it is not that the public is wrong. It is that the evidence we treat as load-bearing was never built to bear that load.
The myth a smart reader has actually heard
Here is the argument as its most careful proponents make it, reconstructed as charitably as I can.
The Martens Clause, woven through the Hague Conventions and the Geneva Protocols, says that in cases not covered by specific treaty law, civilians and combatants remain under the protection of "the principles of humanity" and "the dictates of the public conscience." Advocates read this as a live legal standard, not a rhetorical flourish. If a weapon offends the public conscience, the Clause supplies a basis to regard it as unlawful even before any treaty bans it.
From there the move is straightforward. We have polling. The polling shows durable, cross-national majorities opposed. Therefore the public conscience has spoken, the Clause is triggered, and the legal case for prohibition gets an evidentiary foundation that is harder to wave away than philosophy alone.
The argument is structurally sound. Each link follows from the last. The problem lives entirely in one link — the move from "polling shows majorities opposed" to "the public conscience has been authoritatively measured" — and that is the only link I want to examine.
What the Clause would actually need from a poll
If you are going to convert the Clause's "dictates of the public conscience" from a principle into an evidentiary claim, you are making a measurement claim. You are saying a survey instrument captured a stable, considered, morally-loaded judgment held widely enough and firmly enough to count as a dictate rather than a mood.
That is a demanding standard, and it should be. A legal warrant that can render a weapon unlawful in the absence of treaty law cannot rest on a reaction that flips when you change a sentence in the question. So the relevant test is not "do majorities express opposition." It is: would those majorities hold when the question is posed neutrally, when the realistic alternatives are described, and across the populations whose conscience the Clause invokes. On those terms, the current evidence does not get there. Not because it points the other way — because it is too unstable to point anywhere with the authority the legal claim requires.
Where the chain breaks
Start with framing. Opposition figures move, sometimes by twenty points or more, depending on whether a survey describes "killer robots" or "AI systems that reduce risk to a country's own soldiers." Both descriptions are arguably accurate. When a result swings that hard on diction, the instrument is measuring the wording at least as much as the conviction.
Then performance-matching. Many surveys ask respondents to evaluate autonomous weapons in the abstract, against no baseline. When researchers instead ask people to compare a hypothetical autonomous system to a human operator with an equivalent or worse error rate, stated opposition softens considerably. A judgment that depends on whether you mentioned the human alternative is not yet a settled judgment about the technology.
Then the splits. The cross-national majorities are real in aggregate but conceal wide variance — between countries, between people with and without military exposure, between those primed to think about their own forces and those primed to think about civilians abroad. A "dictate of the public conscience" that holds in one population and dissolves in another is doing something different from what the Clause's universalist language implies.
And then the question almost nobody asks of a number before citing it: who commissioned the survey, and what did they want it to show. A meaningful share of the most-cited polling in this space was funded by organizations whose stated mission is prohibition. That does not make the data fraudulent. It does mean the instrument design, the question order, and the publicized framing were produced inside the campaign they support. Opinion measured this way is, at minimum, not independent of the effort to mobilize it.
Why the opinion moves
The mechanism underneath all of this is mundane: most people do not arrive at a survey with a pre-formed, stable position on machine autonomy in targeting. They construct an answer on the spot, from the cues the question gives them. When the cues emphasize loss of human control, opposition rises. When they emphasize protecting soldiers or matching human fallibility, it falls. This is not public ignorance — it is what attitude formation looks like for any genuinely novel question that ordinary life never forced people to resolve.
That instability is fatal to the specific legal use, even though it is perfectly compatible with sincere moral discomfort. The discomfort can be real and the measurement still too soft to function as a dictate.
What I am not arguing
I want to narrow this as hard as I can, because the narrowing is the point.
I am not arguing that these weapons should be developed or deployed. I am not arguing that moral intuitions are irrelevant to law or policy. I am not arguing that the Martens Clause is empty, or that the philosophical case for prohibition fails — there are arguments from accountability, dignity, and meaningful human control that do not route through polling at all and are not touched by anything here. I am arguing one thing: the evidentiary pathway from existing public-opinion data to a triggered public-conscience standard is the weakest link in an otherwise sound argument, and treating it as solid weakens the cause it is meant to serve.
Before you cite the next survey
A short test I now run on any opinion study before it goes in a footnote:
- Framing: Does the question describe both the threat and the realistic alternative, or only one?
- Baseline: Are respondents comparing the system to a human operator with a stated error rate, or to nothing?
- Variance: Are cross-national and demographic splits reported, or only the headline aggregate?
- Provenance: Who funded it, who designed the questions, and is that the same body advocating the outcome?
- Stability: Did the study test whether the same respondents hold the view under re-framing?
If a number fails three of these, it can describe a mood. It cannot anchor a legal standard.
The honest takeaway
The opposition is probably sincere. The legal pathway built on top of it is not yet sound, and pretending otherwise hands critics an easy and deserved rebuttal. Build the prohibition case on the arguments that survive scrutiny, and treat the polling as what it is — a signal of unease, not a verdict.
Which leaves the question I cannot resolve and neither, yet, can the field: if conscience on a genuinely new question is constructed in the moment of being asked, is there any instrument that could measure it well enough to bind a state — or is "the dictates of the public conscience" a standard we can invoke but never actually verify?
