Early military AI and the drone strikes in Pakistan

The drone campaign in Pakistan was the first case in which automated data analysis and armed unmanned aircraft came together at scale outside a declared front line. It was not a war of autonomous robots — people made the decisions. But it was the first conflict in which the bulk processing of communications and surveillance video began to shape who ended up on a target list, which is why it sits at the root of every later argument about military AI.

This article reconstructs the documented events, the figures that exist and the legal and political controversy they produced. As elsewhere in this archive, the treatment is historical and journalistic; no operational procedures are described.

Where it started: 2001-2004

The armed drone entered combat in Afghanistan in October 2001, when missile-equipped Predators were first used in real operations. The decisive step came on 17 June 2004, with the strike in South Waziristan that killed the Pakistani tribal commander Nek Muhammad Wazir — the first documented US drone strike on the territory of Pakistan, a country the United States was not at war with.

Strikes were sporadic between 2004 and 2007. Escalation began in 2008 and peaked in 2010, with more than a hundred strikes in a single year according to tallies by The Bureau of Investigative Journalism and New America. Almost all of them fell within the Federally Administered Tribal Areas along the Afghan border.

How much automation was actually involved

Nothing existed that resembled a system choosing its own targets. What did exist was a growing reliance on three automated processes:

• Persistent surveillance: continuous hours of video over a single area, a volume that forced the development of automatic object and movement detection — the direct ancestor of later programmes such as Project Maven.
• Telephony metadata analysis: handset location and contact-network analysis, which made it possible to follow devices rather than only identified individuals.
• Statistical classification: the US National Security Agency's SKYNET programme, revealed in 2015 through documents leaked by Edward Snowden, applied machine learning to mobile metadata in Pakistan to flag possible couriers for armed organisations.

SKYNET drew serious technical criticism once it became public. Statisticians pointed out that with such a tiny proportion of genuine positives among millions of users, even a very low error rate yields thousands of false positives. There is no public record of its output being the sole basis for a strike, and the agency described it as an analytical tool. The episode remains the most cited illustration of the limits of automated classification in security work.

Signature strikes and the heart of the dispute

The distinction that shaped the debate separates strikes against identified individuals from so-called signature strikes: attacks authorised on the basis of observed behavioural patterns — gatherings, movements, visible weapons — without knowing who the people are. US officials quoted in the press acknowledged the practice, and it is where the legal objections concentrate: the principle of distinction between combatants and civilians is hard to satisfy when the identity and function of those attacked are unknown.

Two further practices were denounced by human rights organisations: repeat strikes on the same site, which can hit those coming to help the victims of the first, and strikes on funerals or community gatherings. Amnesty International documented specific cases in its October 2013 report «Will I Be Next?», and the 2012 Stanford–NYU study «Living Under Drones» described the psychological and social effects of constant overhead surveillance on the local population.

Disputed numbers

There is no official, verifiable casualty figure. The most cited independent tallies — The Bureau of Investigative Journalism, New America and Long War Journal — place the Pakistan campaign at roughly 400 to 430 strikes between 2004 and 2018, with death estimates ranging from about 2,500 to nearly 4,000 people and several hundred civilians, within wide bands. From executive order 13732 of July 2016 onwards, the US government published annual aggregate figures for civilian casualties outside areas of active hostilities; they were far lower than the independent counts. That gap was never resolved and turns largely on who is classified as a combatant.

The political and legal response

• Pakistan's parliament adopted national security guidelines in 2012 demanding an end to the strikes as a violation of sovereignty, even as some cooperation by parts of the Pakistani state was being documented.
• The Peshawar High Court ruled in May 2013 that the strikes were unlawful and contrary to international law, a decision with no practical effect on the campaign.
• UN special rapporteurs Ben Emmerson and Christof Heyns published reports in 2013 calling for transparency, investigation of incidents with civilian victims and clarification of the applicable legal framework.
• President Barack Obama defended the campaign's legality in his May 2013 National Defense University speech and announced tighter criteria, including a near-certainty standard on the absence of civilians.

Why this case still frames the argument

The Pakistan campaign set three precedents that remain in force. It normalised remote strikes on the territory of a state with which no war had been declared. It introduced decisions based on data patterns before any specific rules governed them. And it showed that without transparency about victims there can be no effective accountability — a problem that automated recommendation systems have only sharpened.