Summary

  • Kenneth Roth, the former executive director of Human Rights Watch, argued in a Guardian column that the operative threat from AI in warfare is target selection, not weapons employment, and that a September Geneva meeting of 128 governments produced a treaty-foundation report whose human-review clause was narrowed by legal teams sent by US President Donald Trump and Russian President Vladimir Putin.
  • The International Committee of the Red Cross has identified automation bias as a central concern in AI-assisted targeting, and Roth invoked this concept to argue that nominal human presence in the targeting chain reproduces the deference patterns the meaningful-control requirement was designed to prevent.
  • Israeli intelligence officers quoted by Roth described their role in the Israeli military’s Lavender and “Where’s Daddy?” AI targeting systems as a “rubber stamp,” with one officer reporting 20 seconds per target review and a willingness to accept an anticipated 20 civilian deaths per low-level Hamas fighter.
  • Roth’s claim that the documented review process amounts to war crimes rests on a contested interpretation of international humanitarian law and on Israeli-source accounts whose methodology — including the 20-to-one ratio and the approximate 10 percent non-combatant flagging rate — the column does not specify.
  • The Geneva process, on Roth’s account, has not resolved whether “meaningful human control” will be specified to require substantive human review or to allow formal compliance through recorded human decisions and system-design certifications.

A Guardian column by Kenneth Roth, the former executive director of Human Rights Watch, frames the September meeting of 128 governments in Geneva as the moment at which a contested specification — whether “meaningful human control” of AI-assisted targeting will require substantive human review or permit formal compliance through recorded human decisions — moved against the substantive reading. Roth’s argument centers on the Group of Governmental Experts (GGE) process under the Convention on Certain Conventional Weapons (CCW), and on operational accounts of the Israeli military’s Lavender and “Where’s Daddy?” systems drawn from Israeli intelligence sources reported by +972 Magazine and Local Call. The column, published October 2, 2026, indicts the Geneva outcome without engaging the trade-off a weaker report implies — present weakness in exchange for a possible binding treaty later — and leaves the inferential chain from documented “rubber stamp” review to criminal liability unargued.

The conceptual revision Roth performs

Roth states the revision explicitly. “The colloquial term ‘killer robots’ — the label for fully autonomous weapons — is deceptive,” he wrote, “because the more pressing problem is AI’s growing role in selecting targets for conventional weapons.” The revision is twofold: it relocates the contested category from weapons employment — pulling a trigger — to target selection — deciding whom to fire at — and it identifies the variable that distinguishes acceptable from unacceptable human involvement as not whether a human is present in the chain but whether that human exercises what Roth terms meaningful or effective human control.

The displacement has analytical consequences that run through the column. Advocates of a narrow prohibition on fully autonomous weapons, in Roth’s framing, benefit when attention is directed at fully autonomous systems; advocates of meaningful human control benefit when attention is directed at target selection. The “human in the loop” formulation that the Geneva drafters were working toward, Roth wrote, was the operative phrase, and what survives in his account of the September meeting is a formally compliant but substantively hollow version — a recorded human decision and a system design satisfying the framework’s formal requirements.

Roth reaches for two historical analogies to dramatize the stakes. The first is The Terminator — Arnold Schwarzenegger’s film — which he treats as the pop-culture frame that has crowded out the actual problem. The second is the Tahrir Square uprising that contributed to the ouster of Egyptian President Hosni Mubarak. Roth wrote that Mubarak’s fate was sealed when the Egyptian army refused to fire on protesters, but killer robots would have had no such qualms. “If Mubarak had been able to deploy them, he might have stayed in power,” he wrote. He then notes that the Geneva negotiators are addressing AI weapons used in war rather than those used for repression, a limitation he attributes to the Trump and Putin legal teams: “The Trump-Putin lawyers made sure that the agreed report referred only to war, not repression,” he wrote.

The September Geneva meeting and the reportedly weakened clause

The September meeting of 128 governments, as Roth reports it, produced a report that could serve as the foundation for a binding treaty. The contested provision, on Roth’s account, required humans to review military targets developed by AI before a strike; lawyers sent by Trump and Putin reportedly teamed up to water down the report, particularly weakening that provision. Whether the GGE will revisit the provision, or whether the September text proceeds, is not specified in the column.

The narrowing can be defended on the ground, articulated in academic literature on autonomous weapons, that the war-and-repression contexts require different legal doctrines and that an instrument suited to neither serves neither. States with internal AI-repression capabilities benefit from the repression-exclusion; states that deploy AI-assisted targeting in armed conflict benefit from a human-review clause drafted to allow formal compliance. The opposing reading of the question Roth frames — that formal human presence combined with deference to the machine’s outputs reproduces the conditions the ICRC has termed “automation bias” — is the position Roth adopts. He does not address the procedural reading on its own terms.

The trade-off a weaker report implies sits outside Roth’s framing. The column indicts the Geneva outcome without engaging whether a treaty that fails to specify the structure of meaningful human control may nonetheless constrain practice through the ratification and implementation process, whether domestic legislation by states that develop or acquire AI targeting systems may move faster but produce a patchwork, or whether the exercise of jurisdiction by international or hybrid courts depends on state referrals or prosecutor standing. Each addresses a different aspect of the problem; the column does not assess these alternatives.

The Lavender and “Where’s Daddy?” accounts

Roth turned to ongoing conflicts to illustrate what he described as the present reality of AI in targeting. He wrote that Russian and Ukrainian troops are experimenting with fully autonomous weapons in eastern Ukraine to overcome jamming of drone control signals, and that these deployments are likely to expand. On Gaza, he described the Lavender system, which the Israeli military deployed after the October 7, 2023 Hamas attack to identify patterns in phone communications equated with Hamas operatives. A separate automated system called “Where’s Daddy?” tracked those individuals to their homes, where they could be more easily targeted. “There was no pretense that Hamas was using its operatives’ homes for military purposes,” Roth wrote. “It was simply easier for the AI system to find suspected Hamas fighters where they slept. The families were the collateral damage.”

Roth acknowledged that the targeting process formally included human review, but described that review as nominal. He quoted an Israeli intelligence officer as saying: “I would invest 20 seconds for each target at this stage, and do dozens of them every day. I had zero added-value as a human, apart from being a stamp of approval. It saved a lot of time.” The human judgment exercised was, in Roth’s words, perfunctory — a “rubber stamp” — and from computer to killing the result was, he wrote, that “an algorithm played God.”

The sourcing for these operational characterizations is restricted to a limited set of named Israeli officers’ statements quoted in the column, drawing on +972 Magazine and Local Call. Official responses are not engaged. Three elements of Roth’s argument carry analytical weight but rest on a more limited evidentiary base than his categorical claims require: the officer accounts, the 20-to-one ratio, and the joint characterization of the US and Russian negotiators as narrowing the report for shared reasons. Roth does not distinguish whether the US and Russian delegations objected to different provisions for distinct reasons.

The proportionality question and the 20-to-one ratio

Roth wrote that the Israeli military permitted an anticipated 20 civilian deaths for the killing of a suspected low-level Hamas fighter. He cited the international humanitarian law principle of proportionality, which prohibits attacks expected to cause civilian casualties “excessive in relation to the concrete and direct military advantage anticipated.” According to Roth, virtually no legal expert considers Israel’s 20-to-one ratio lawful, though he does not name those experts. “These are war crimes,” he wrote. “They reflect an appalling devaluing of Palestinian civilian life.”

The 20-to-one figure’s methodology — who calculated it, against what baseline, and over what operational period — is unspecified. The column does not address whether a different ratio or different algorithmic thresholds would meet the proportionality standard, and does not engage with the academic literature on the IHL-lawfulness of AI-assisted targeting as a category. Roth also noted that the Israeli military conceded that approximately 10 percent of the human targets flagged for assassination were not members of Hamas’s military wing. That figure strengthens the conceptual argument that the system’s design tolerates a non-trivial collateral rate by construction, but the column does not integrate it into the proportionality analysis.

Roth’s treatment of deference to AI outputs as constitutive of war crimes depends on an interpretation of international humanitarian law that is contested. Some legal commentary distinguishes targeting errors and institutional tolerance of civilian harm from the intent requirements for the most serious IHL offenses, leaving the inferential leap from “rubber stamp” review to criminal liability unargued. Roth does not engage that distinction.

He did quote Israeli President Isaac Herzog as having said, in the aftermath of October 7: “It is an entire nation out there that is responsible. It’s not true, this rhetoric about civilians were not aware, not involved – it’s absolutely not true. They could’ve risen up, fought against that evil regime.” Roth infers from that statement that “the same contempt … contributed to the Israeli military’s willingness to tolerate civilian deaths.” The inference links a public political statement to operational targeting tolerances, leaving the causal chain between them unspecified.

The “dispassionate mechanism” defense and the ICRC automation-bias concept

Two of the Israeli officers quoted by Roth offered a defense of AI-assisted targeting that the column presents as a live position but does not develop. One officer suggested the practice might be justified because AI makes better decisions than humans do, because it is dispassionate while people’s judgment can be clouded, such as by the animosity toward all Palestinians that some Israelis felt after the Hamas attack. Another soldier expressed greater trust in the AI system than in fellow soldiers: “I have much more trust in a statistical mechanism than a soldier who lost a friend two days ago. Everyone there, including me, lost people on October 7. The machine did it coldly. And that made it easier.”

The narrowness of that defense — that the empirical claim against AI-assisted targeting depends on showing that human reviewers, given the same information, would make more discriminate decisions, and that the counterfactual to a 20-second human review is not a 20-minute review but a more erratic one — is not developed in the column. Roth does not compare the 20-to-one ratio to expected ratios under alternative targeting methods or under a counterfactual military response to October 7.

The ICRC, which has contributed to the Geneva process, has framed automation bias — the systematic human tendency to defer to algorithmic outputs — as a central concern in AI-assisted targeting, and has argued in its published work on new technologies in warfare that meaningful human control requires not just attestation but the standing, information, and disposition to evaluate a targeting recommendation. Roth invokes this concept and argues that a “human in the loop” requirement without specification of the structure of the loop will fail the test its proponents mean it to pass. “When inevitably flawed human beings make decisions, they can expect others to challenge them,” he wrote. “But AI makes it easier for us to defer to its judgments.”

The unresolved specification of meaningful human control

The argument’s reach depends on inferential transfer from one operational system — Lavender as described — to a categorical claim about AI in war. A target-selection system with calibrated uncertainty, integrated checks, and slower human review would not, on Roth’s own logic, fail the meaningful-control test. The categorical extension to “an algorithm play[ing] God” therefore requires either evidence that no calibrated system is being deployed in this class of operation, or a more ambitious claim that target selection by AI is incompatible with meaningful control in principle. The Arendt “banality of evil” frame and the closing invocation of “the mechanical application of an algorithm” gesture toward the second claim without arguing it directly.

Roth closes by invoking Hannah Arendt’s “banality of evil” to describe what he characterized as ordinary soldiers becoming complicit in extraordinary cruelty by deferring to what they call “the system.” The essence of that system, he wrote, remained people — the generals who ordered its creation, the military lawyers who excused the unjustifiable, the programmers who tweaked the algorithm, the intelligence officers who passed on the computer-generated targets, and the soldiers who launched the attack. He wrote that part of what made it easier to trust the system was the pseudo-scientific precision of AI, and that humans are familiar with the breezy, authoritative tone of today’s chatbots. When AI is used for targeting, Roth wrote, it brings a similar feigned omniscience to killing. “Even in war, people should never be snuffed out by the mechanical application of an algorithm,” he wrote.

The unresolved question the column surfaces is whether the phrase “meaningful human control,” as the Geneva process is currently drafting it, can be specified in a way that distinguishes the kind of review the column documents from the kind of review the phrase was meant to require. The Geneva record, on Roth’s account, does not appear to resolve this; the ICRC, the parties to the GGE, and the academic literature on autonomous weapons have articulated different positions, and the September meeting, according to Roth, left the question open. Whether the GGE revisits the provision, whether a binding treaty emerges on the September text or a successor, and whether the operational systems the column describes continue to be deployed under the same review structure are the questions the column does not resolve.

Additional considerations

Roth argued that meaningful human control of AI requires not just nominal endorsement of its decisions but intense questioning of its product, along with accountability for commanders when junior officers fall short. The column does not specify the institutional design that would produce such accountability — whether through a treaty inspection regime, domestic chain-of-command reforms, or external judicial review — and the gap between Roth’s prescription and the institutional mechanisms available to implement it is the analytical terrain the column does not develop. The framing also applies asymmetrically: the operational system Roth describes is the Israeli military’s, while the diplomatic contest he describes is between the United States and Russia. The column does not address whether the Israeli system’s design is connected to the US-Russian diplomatic posture, whether Israeli negotiators participated in the Geneva process in a way Roth does not specify, or whether the proportionality analysis Roth conducts on Gaza would apply, on his own logic, to the Russian and Ukrainian autonomous-weapons deployments he mentions in passing.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Conceptual Engineering
Asks not just what a concept means but what it should mean, and re-engineers it.
Dialectical Analysis
Holds thesis against antithesis and works toward a higher synthesis.
Red-Team Advocate
Argues the adversary’s case in full to expose what a plan underrates.