TL;DR

Whether AI works in British defence and intelligence will come down to people rather than procurement, argues a briefing paper out this week from the Turing Institute’s security research centre. Its two warnings are that practitioners lose sharpness when they lean on automation, and that they wave through machine output they should be interrogating. Four priorities are set out for organisations adopting the technology.

The argument

CETaS, the Turing’s Centre for Emerging Technology and Security, has published the opening instalment of a longer programme looking at intelligence work and who will be doing it in a decade. Rather than theorising, the authors go looking for evidence in two professions further down the same road.

Law supplies the sharper example. Document review, legal research and drafting are already heavily automated — and those were precisely the tasks through which junior lawyers used to acquire reasoning and judgement. Remove the grunt work and you remove the apprenticeship hidden inside it. Medicine shows a parallel risk around documentation and diagnosis, where the concern is clinicians gradually deferring to a recommendation they no longer independently evaluate.

Applied to intelligence work, the paper accepts that translation, summarisation and mapping data are fair game for automation. What cannot be handed over is weighing how reliable a source is, reading uncertainty properly, spotting what is missing from a picture, and exercising judgement against an adversary who is actively trying to mislead you.

Four priorities

The recommendations are unglamorous: train people properly rather than nominally; work out deliberately what ought to be handed over, instead of automating whatever proves easiest; build systems that make human oversight workable rather than performative; and keep measuring what the technology is doing to performance after deployment.

Ardi Janjeva, a senior researcher there, said AI holds enormous potential for national security and resilience “but only if we preserve the human judgement needed to challenge its outputs”.

Looking forward

This lands in a week that keeps making the same point from different directions. Up Learn’s survey found teachers spending on checking what AI produced whatever time it saved them, which is the automation-bias problem measured in a classroom. In the legal sector, guidance published today tackles firms that cannot show where client data actually went, and last week a firm warned that public AI tools may waive privilege.

The uncomfortable implication for any organisation: the savings arrive immediately, the deskilling arrives later, and only one of them shows up in a business case.