Twenty Analysts' Work, One Every 3.6 Seconds: FT Warns Military AI Targeting Is Scaling Errors at Machine Speed

The Financial Times has published one of the most consequential AI stories of the week — and it is not about valuations, chips or model releases. On September 17, the paper reported that the speed and scale of AI-assisted target generation in modern militaries has now decisively outpaced the ability of humans to verify what the machines are recommending. The core finding: twenty soldiers equipped with AI targeting tools can now handle a workload that during the 2003 invasion of Iraq required roughly 2,000 analysts, and military programs are pushing toward 1,000 tactical decisions per hour — one decision every 3.6 seconds.

That last number deserves a moment of reflection. A human analyst working a traditional target-development cycle measures work in hours or days: fuse intelligence from multiple sources, cross-check imagery, consult signals intercepts, verify that the structure in the crosshair is what the model says it is. At one decision per 3.6 seconds, there is no cross-check. There is barely a glance. The human in the loop becomes, functionally, a rubber stamp operating at the tempo of a GPU.

What the FT actually found

The FT’s reporting, summarized by AI Weekly’s trackers on September 17, makes three interlocking claims:

First, the productivity gain is real and enormous. The ratio the FT cites — 20 operators replacing 2,000 analysts from the 2003 Iraq invasion era — represents a 100-fold compression of the human capital required to run a targeting pipeline. For militaries facing recruitment shortfalls and demanding high-tempo operations, this is precisely why the technology is irresistible.

Second, verification has not scaled with it. The analysts who used to manually confirm each target are exactly the bottleneck AI removes. The FT’s warning is that errors — stale intelligence, misclassified structures, outdated maps, civilian facilities that were once military sites — now propagate at machine tempo instead of being caught by deliberate human review.

Third, the error-scaling dynamic is not hypothetical. It maps directly onto documented incidents. A preliminary Central Command assessment of a March 2026 strike, reported by the Arms Control Association, found that U.S. intelligence maps failed to show a school facility had long ago been converted from military to civilian use — and that it had been added to an AI-generated target list “without adequate human supervision.”

The Iran campaign: the proof case

The FT’s warning lands with particular force because of what the world watched during Operation Epic Fury, the U.S.-Israeli campaign against Iran launched on February 28, 2026. It was the first large-scale deployment of generative AI targeting systems against a sovereign state, and by every public account it was a tempo revolution: more than 1,000 targets struck in the first 24 hours, and — per a sworn declaration by the Pentagon’s Chief Digital and AI Officer — the Grok government model inside the Maven Smart System supported more than 2,000 munitions against 2,000 distinct targets in 96 hours.

Officials have consistently maintained that humans retain the final engagement decision, and that Maven “recommends, ranks and surfaces” targets rather than approving them. But critics, including lawmakers demanding greater oversight, argue this framing obscures the operational reality: when the system generates targets faster than any staff can meaningfully evaluate them, the distinction between recommendation and decision collapses. The OECD’s AI incident tracker formally logged the Iran campaign’s civilian-harm reports, citing algorithmic errors in AI-driven targeting that accelerated attacks ahead of verification.

Research published this year anticipated exactly this failure mode. An April 2026 study by the Institute for AI Policy and Strategy on AI decision-support systems concluded that such systems “increase the pace and scale of decision-making, producing recommendations that are difficult to independently verify under time pressure.” West Point’s Lieber Institute has gone further, describing an “illusion of precision” — the sense of confidence that machine-generated target packages project, even when the underlying data is stale or wrong.

Why error scaling is the defining problem

The economics of AI targeting are the same as the economics of AI everywhere: marginal costs fall toward zero while throughput rises by orders of magnitude. In a commercial context, an error that slips through at machine speed costs a refund or a bad answer. In a targeting context, the equivalent error destroys a building and the people in it.

Three structural factors make this uniquely dangerous:

Latency asymmetry. A target-generation system can surface a candidate in milliseconds. Confirming that candidate — checking the provenance of the imagery, the recency of the intelligence, the current use of the facility — still takes a human minutes to hours. The verification bottleneck is physical and cognitive, not computational.

Automation bias under tempo. Operators under operational pressure tend to accept machine recommendations, especially when the interface presents them with confident, well-formatted target packages. At one recommendation per 3.6 seconds, acceptance becomes the default and scrutiny the exception.

Error correlation. Because a single model with a single set of training data and a single intelligence picture generates the entire target stream, its mistakes are not independent. One stale dataset can systematically mislabel dozens or hundreds of structures — the exact pattern seen in the converted-school incident, where the failure was in the underlying map, not in any individual decision.

The governance gap

What makes the FT’s report politically pointed is its timing. The same week, the U.S. House voted 417-3 on AI infrastructure costs, the EU’s Ursula von der Leyen endorsed “pacing the frontier” in her State of the Union address, and a trio of tech CEOs was reported to have killed a formal AI oversight framework in White House lobbying. The regulatory conversation is consuming itself over frontier-lab governance — while the fastest, most consequential deployment of AI decision-making is happening inside classified military pipelines with far thinner public scrutiny.

The uncomfortable truth the FT surfaces is that military AI adoption is not waiting for the safety debate to resolve. Every incentive — personnel shortages, peer competition, demonstrated operational advantage — pushes toward more automation, faster loops and fewer humans in the decision path. The verification capacity that would make machine-tempo targeting safe has no equivalent of a scaling law working in its favor.

What to watch

The report sets up several concrete developments worth tracking. Lawmakers, already pressing for stricter oversight of Pentagon AI use after the Iran campaign, now have a authoritative journalistic account of the structural problem to cite. International humanitarian law scholars continue to argue over whether “meaningful human control” is achievable at these tempos at all, or whether the concept needs to be redefined around system-level audit rather than per-decision approval. And the Pentagon’s own after-action processes — including the Central Command assessments of civilian-harm incidents — will show whether error rates scale sublinearly with tempo, as proponents hope, or linearly and beyond, as the FT’s reporting implies.

The 2003-era targeting cycle was slow because humans were the constraint. The 2026-era cycle is fast because humans have been removed from the constraint. The FT’s contribution is to insist that the question — what did we lose when we deleted 1,980 analysts from the loop? — be asked out loud before the next campaign, not after it.