Using (and Abusing) Polling: The Martens Clause and Autonomous Weapons
Editors’ note: This post is based on the author’s recent article, “Abhorrence and Autonomy: On the Place of Public Attitudes in the Debate on Militarised Artificial Intelligence” published in Volume 39 of Philosophy & Technology.
Since debates around Autonomous Weapon Systems (AWS) picked up steam in the early 2010s, the Martens Clause has been invoked by ban advocates. The Clause, codified in the Additional Protocol I, Article 1(2), puts limits on the use of weapons absent treaty language.
In cases not covered by this Protocol or by other international agreements, civilians and combatants remain under the protection and authority of the principles of international law derived from established custom, from the principles of humanity and from the dictates of public conscience.
For advocates of banning AWS, the Martens Clause presents a clear path to prohibition or restriction on the use of lethal autonomy. Human Rights Watch (HRW) emphasises this in their 2012 and 2018 reports on AWS, and the Clause’s promise was seemingly behind the HRW and the Stop Killer Robots campaign efforts to commission a series of Ipsos polls in 2017, 2019, and 2021. Similar polling efforts have also resurfaced in 2026 following the U.S. Department of War’s standoff with Anthropic.
The Promise of the Clause, and the Absence of an Operational Standard
On an expansive (and not uncontested) interpretation, the Martens Clause provides a positive source of international law that is democratic and objective. By democratic, advocates mean that the opinion of the masses has sway over a field that, by and large, is the preserve of elite diplomats, politicians and lawyers. And by objective, such advocates infer that the opinion of said masses points towards natural law that it is not only lawful but morally correct.
The problem as others have noted, is that if advocates want to use such an expansive reading of the Martens Clause to ban AWS (or any other weapons system), the language doing the heavy lifting here, “the dictates of public conscience,” lacks a clear operational standard. And without such a standard to evaluate the polls advanced as evidence of bans, we have no way of distinguishing genuine instances of public abhorrence towards a weapon and mere advocate bias.
Therefore, if we are going to permit the Clause to bear meaningful weight in prohibition debates, whether decisive or merely suggestive, we need to establish how to tell whether any given measure of public conscience (namely polling) is authoritative or distorted. Drawing on my previous research, I argue that there are three standards we should consider: invariance; breadth; and independence.
Standard 1: Invariance
The first standard any evidence of public conscience must meet is invariance. By this, I mean that any measure of public attitude is only as authoritative as its robustness under perturbation. Or, more simply, if the contingent facts of a situation change, public opinion should not.
Invariance matters because there is a long documented history of framing effects shifting sample populations opinions on questions. Simply, the way you structure a question will shape the kinds of opinions that you elicit from your sample population. For instance, imagine someone asking the question: “Don’t you think the government spends too much on welfare?”
The phrase “too much” smuggles in a bias towards already seeing welfare as overfunded. And the phrase “don’t you think …” nudges the respondent towards agreement. Such framing pushes the dial towards one answer before the question is even contemplated. Compare a second, neutral, version of the question that doesn’t presuppose an answer.
“What’s your view on government welfare spending?”
The problem for activists seeking to use the Martens Clause to ban AWS is that the primary evidence, the Ipsos polling, commits exactly this blunder. See the question asked repeatedly between 2017-2021.
The United Nations is reviewing the strategic, legal and moral implications of autonomous weapons systems. These systems are capable of independently selecting targets and attacking those targets without human intervention; they are thus different than current day “drones’” where humans select and attack targets. How do you feel about the use of autonomous weapons in war?
Firstly, the question primes the audience by situating the weapons as being reviewed by the UN, presenting the class of technology as morally suspect. Secondly, the question provides a morally salient alternative, drones, providing only one morally relevant dimension (human intervention) for consideration, in which AWS are portrayed as deficient.
If people’s attitudes were robustly anti-AWS, questions asked without such framing and with perturbation along other morally relevant axes would still elicit a majority opinion against such use cases. However, as empirical evidence demonstrates, as soon as you vary along the axis of performance (the machine is more combat effective or accurate than the human), people’s opinions become quite plastic, and even a modest gap in machine excellence can flip majority opposition to majority approval for the employment of AWS.
My point here is that any evidence for a Martens Clause based ban requires abhorrence towards a system or tactic that retains at least a general majority absent unwarranted frame priming and across relevant perturbations of morally and legally relevant factors.
Standard 2: Breadth
The second standard required to activate the Clause is breadth: “humanity” in a broad sense must hold a common attitude towards said technology or tactic, not merely an average. As others have noted, such demographic flattening can erase genuine moral disagreement between groups into a single authoritative statement, concealing fundamental incompatibilities between the judgments held by different groups. If abhorrence is only narrowly held across some demographic segments, this is evidence of parochial sentiment, not public conscience.
This poses a real problem for AWS ban advocates: within Ipsos’s own polling there is huge variance between genders, age groups and nations, spanning from 60% approval in India to two-thirds opposition in Russia, Peru, Spain and Argentina. Elsewhere, Chinese respondents have shown double the positive attitude toward AWS compared to other nationalities.
My point here is that a poll that claims that 61 percent of the public support a ban on AWS, but which includes a 60/40 pro-AWS split in India and a 34/66 anti-AWS split in Spain is not a consensus; it is an average. And if the expansive Martens Clause reading is going to live up to its democratic and objective claims, it needs to be truly representative of the plurality of the opinions of the global public.
Standard 3: Independence
The third standard evidence must meet is independence: the public attitude being measured should not be an artifact of the campaign invoking it. This is a structural claim, rather than an accusation of malfeasance on behalf of activists lobbying for the prohibition of AWS. If public conscience grounds the ban and the pollster wants the ban, there are incentives to cultivate rather than detect opinion.
This is not merely a potential observation, but an observed historical fact in the way that activists have used campaigns to shape rather than merely observe public opinion. For instance, the Ottawa Treaty came about in part through the use of emotional framing and reputational shaming of States that retained landmines. And activists promoting the cluster munitions treaty, used similar tactics, including its unusual duty to “promote the norm.”
My point here is not that advocacy is wrong; rather, it is an observation that the majority of empirical data on attitudes towards AWS until 2023 were commissioned by and for advocates of prohibition. And just as we would be suspicious of polling emerging from Palantir, Anduril, or Raytheon showing support for military AI, we should be aware of the biases held by those who have already decided that lethal autonomy is wrong and must be banned.
Should We Give Up On Polling?
Activists might argue that I am attacking a strawman, that they had always acknowledged that polling needed interpretation, and that the emergence of expert opinion and State positions is in any event a better indicator of public conscience. This seems to be the approach that Stop Killer Robots has taken in 2024 and 2025, and is implied by HRW dropping any mention of polling from their 2025 report, despite polling having a prominent place in their 2018 report.
But a Martens Clause reading that treats public conscience as whatever experts and activists can mobilise and agree upon is one that abandons any pretence of democratic legitimacy. If the Clause is to be a meaningful positive source of international law, living up to the high ideals undergirding its expansive reading, it needs to be grounded in public opinion that is broadly held and robust: the average person on the street. And such characteristics can only be measured empirically: by means of polling.
An opponent might argue that granting this, my above arguments have already proven too much. If public opinion is plastic to framing, irreconcilable across cultures, and structurally deformed by partisan interest, why should we bother with the Martens Clause at all?
Such an argument is overly pessimistic. Although framing effects are real, research shows that people revert to underlying principles when exposed to competing frames. The difficulties totalitarian societies face in controlling public opinion indicate fears of wholesale opinion manipulation are overstated. And even if people disagree on some issues across demographic lines, this doesn’t preclude shared values that recur beneath surface diversity.
A Systematic Approach
So if polling is the only meaningful way to gauge public conscience, how should we go about measuring it? Pollsters need to take measurements neutrally across diverse sub-sample populations, include a variety of valid parametric perturbations, and index them against the closest accepted comparator.
The issue of measurements is simply a matter of applying good social science practices to ensure that we aren’t fooling ourselves: ensuring questions are asked either in the most neutral frame possible, or with a randomised mixture of competing frames helps us measure the underlying opinion rather than the frame itself. And using practices like sub-sampling helps us avoid presenting flattened averages of opinion as definitive rather than capturing meaningful disagreement across demographic groups.
With parametric perturbation, pollsters should hold the use case fixed and vary only the relevant features in question, using something like factorial surveys, a well-established method of eliciting a population’s beliefs about the world, judgment principles, and decision rules. Varying along meaningful axes (such as machine performance, military necessity, risk to own forces, and risk to civilians) will help expose such valuable data.
Finally, indexing the opinion against common military practice (for instance, by using something like conjoint analysis) helps ensure that any negative sentiment is due to the technology or tactic at play, rather than just negative sentiment to warfare in general. For instance, if you vary the scenario so that the only difference is whether a human pulls the trigger or a machine, and both versions evoke the same concern, your polling is detecting only a general anti-war stance. What is needed is evidence that AWS (or any other Martens Clause candidate) is uniquely shifting the needle.
As I have mentioned in my own research, such combined methods should produce an abhorrence salience map, which if designed properly, will show areas of clear opposition, clear support, and “grey zones” where public opinion is divided on the use of such technologies. Such evidence, if we take the Martens Clause seriously, should provide concrete support for targeted regulation, rather than the broad all-or-nothing approach many advocates currently embrace. Furthermore, such a method generalises to any future Martens claim about any emerging capability—neurotechnological implants, bio-engineered super soldiers, fully autonomous machine combatants—providing us a general means of operationalising the Clause.
Conclusion
The key point is that the Martens Clause remains intact; what fails is trying to find a shortcut to it. The standards I have outlined are an invitation, not an obstacle. But advocates wishing to invoke the Martens Clause need to accept the risk that properly conducted research might not return the hoped-for answer. If the public genuinely do not find AWS (or any other particular system) abhorrent, shouldn’t that count as evidence against a Martens Clause based prohibition? By raising the evidential bar, we should be able to distinguish when public conscience genuinely speaks from when it merely echoes.
***
Tim Dubber is a philosopher of military technology completing his PhD at the Australian National University. He has over a decade and a half experience in the Australian Department of Defence, mostly within the Australian Regular Army, which included deployments to Afghanistan and the Middle East. This post does not necessarily reflect the views or positions of the Australian Department of Defence, nor the Australian government.
The views expressed are those of the author, and do not necessarily reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense.
Articles of War is a forum for professionals to share opinions and cultivate ideas. Articles of War does not screen articles to fit a particular editorial agenda, nor endorse or advocate material that is published. Authorship does not indicate affiliation with Articles of War, the Lieber Institute, or the United States Military Academy West Point.
Photo credit: U.S. Air Force, Airman 1st Class William Rosado
