Thread by @zoemsl

Alors les amis, The French Ministry of Armed Forces #AI task force has released a new report covering #AI in #defence & a roadmap toward more military adoption. You can read it en français here: defense.gouv.fr/salle-de-press…

Or read on for a few findings & preliminary thoughts.

FR wants its very own JAIC-like cell, the CCIAD, embedded within their Defence Innovation Agency (IAD). (Shout out: FR speakers should follow @KrajeckiMichael). The Coordination Cell for AI in Defence will facilitate efforts on coordination & strategy. p21 for this buried lede

Nice to see that France does not see AI as an end unto itself. Intro is pragmatic & focusses on the weaknesses of second-wave AI systems (although without using the #DARPA-esque vocab), including by noting that #AI does not change the nature of war.

There’s plenty on threats, capability development, interoperability/cooperation, relationships with the private sector… but hardly anything on arms control & strategic stability.

FR appears interested in using AI to achieve at least two specific effects are mentioned: speed & deception. Near term: AI necessary tool to defend against surprise & augment human decision-making by providing more planning options more quickly.

But France also worries about how these effects (& more) widen the attack surface against their own forces, especially in #asymmetricwarfare

Similar to US DIB ethics guidelines conversations, it’s not a surprise that the three ways FR sees AI providing better situational awareness/increasing safety will relate to the three #IHL/#LOAC principles: discrimination, proportionality, and necessity.

(I personally think it’s too soon to assert that AI will necessarily ameliorate the conduct of war along the lines of these principles (& prefer ICRC’s more nuanced take here.)

This report also provides useful context for future #AI middle powers (the “second circle”). One important takeaway: beyond large players like US or China, countries like France will have to find the appropriate balance between reliance on foreign data & own #strategicautonomy

Another lesson for middle powers is that cooperation will be necessary. Also, political note: #techsovereignty seems more national than European (p10); but #strategicautonomy may be more European than national (p27).

French mil #ML applications will rely mostly upon open-source systems. Also notes a few difficulties of which any aspiring #AI middle power should consider: structuring defence data is labour-intensive and compute (I’d add especially for small units) is expensive and inconsistent

Many of the specific applications of #AI seek to reinforce the human role in decision-making by increasing her accuracy, speed, options. Vulnerability of systems also makes human redundancy crucial (esp. for C2, I’d add).

Really glad to see FR provide such a nice sentence on allied #AI #interoperability – in the future it won’t only be about technical interop, but also about common norms & strategic importance to defend against attacks on unity/cohesion [of alliance]

FR notes the challenge AI poses to mil procurement cycles, but TBD whether the approach is useful. Yes it’s essential to consider AI integration into new equipment...

...but is it really useful to still think about lifecycles over the course of decades? Maybe more iterative processes are more important than squeezing AI lifecycles into planning.

Be still my heart – FR notes the necessity of auditability! But FR is not particularly transparent on this front so I don’t know how to assess the audit ‘culture’ into which AI will inevitably enter. I wish we would see more on auditability & accountability together.

Despite my concerns on how FR is framing #AI as a force for good re: IHL principles, I think it’s a great that these conversations are taking place while the FR ministerial committee on #AIethics is being stood up. That said, not clear what role #defence will play in committee.

Most exciting part: data governance! Similar to the DoD Cloud Strategy – but smart for FR to embed important parts from 2018 DGNUM ambitions here to drive home how interconnected cloud & AI are. Also: if anyone wants to send sources on ARTEMIS/POCEAD my way, je vous écoute ! Ciao