At Rise of AI 2023 in Berlin, Sandro Gaycken, founder of Monarch, argues that despite popular myth, Western defense and intelligence agencies are not major drivers of deep-tech innovation. He frames the problem as three interlocking barriers he calls the 3P problem: payment, procurement and political procrastination. Public salaries cannot compete with industry pay for hackers, public procurement bodies such as Germany's BAAINBw take up to 18 months just to review a submission and require dozens of sign-offs, and politicians avoid deciding on technically complex, high-risk topics because failure is punished while inaction is not. As a result, he says, military AI is mostly bought from the consumer electronics and advertising industry rather than built for defense, which imports biases unsuited to battlefield conditions. He describes a Pentagon simulation in which two commercial AIs both recommended total nuclear annihilation, and warns that military AI systems are rarely built under vetted, secure conditions and are therefore highly hackable. He closes by describing intelligence use of AI as mainly dependent on the analyst asking the right questions.
Key takeaways
- Gaycken frames Western defense-tech dysfunction as a 3P problem: payment, procurement and political procrastination.
- Public-sector hacker salaries lag industry pay by a factor of 10 to 20, driving talent away from agencies like the BND.
- Germany's federal defense procurement office takes about 18 months to first review a submission and needs up to 20 sign-offs.
- In a Pentagon simulation, two commercial AI systems both recommended total nuclear annihilation as the optimal response.
- Military AI is mostly bought from the consumer electronics and advertising industry, importing biases unsuited to battlefield conditions.
Detailed notes
The 3P problem: payment, procurement, political procrastination
Gaycken says the common belief that defense and intelligence agencies drive AI and deep-tech innovation is a myth, at least in the West. In an article he wrote for West Point, he described three linked obstacles: payment, procurement and political procrastination, which together block innovation in Western militaries and intelligence services.
The payment gap for hackers
Public salaries cannot attract top technical talent, he says. Agencies such as the BND try to hire hackers at roughly 50,000 to 60,000 euros a year, against industry pay of 300,000 to 500,000 for a good hacker and 1.2 to 1.5 million for a top one, a gap of a factor of 10 to 20. Those who join out of patriotism often find themselves mostly given compliance and legal training rather than technical work, which he says reflects a broader cultural problem around innovation in public service.
Public procurement: the BAAINBw and the US Donovan Group model
Gaycken calls public procurement in intelligence and defense a mess. At the German federal procurement office (which he refers to as "buy nbv"), it takes about one and a half years, despite 6,000 staff, before a submission is even reviewed, and at least 18 to 20 parties must sign off, each changing specifications, so the outcome often differs completely from what was requested. He cites a frigate that was painted with water-soluble ink as an example of the result. He says the problem extends across Europe, the UK and the US. The US did build a faster model over 18 years: a Special Operations-owned private company, the Donovan Group, can buy from the market in two weeks and, through framework contracts with vetted startups, draw on technology relatively quickly. A general expressing interest in a startup's technology at a conference, he adds, still only gives that startup an estimated 8 percent chance of a contract within 4 to 5 years.
Political procrastination
Even when procurement fails are recognized, politicians who try to fix it quickly discover the field's technical, non-technical and regulatory complexity, and face industry actors who lie or withhold information to serve their own interests. Because in Germany officials are punished for failure but not for inaction, Gaycken says decision-makers tend to retreat to safer portfolios such as agriculture, calling the avoidance pattern the "OGP tactic" (other guy's problem). Once money has been spent on a bad decision, he adds, officials tend to defend it rather than admit the mistake.
Where military AI actually comes from
Gaycken says AI used in defense mostly originates in consumer electronics and advertising, which brings different biases than defense engineers would build in, including from high-functioning autistic engineers he says are deliberately recruited into core AI design in Silicon Valley and Boston. He describes a Pentagon test in which two commercial AI systems, placed in a simulated thermonuclear scenario with incoming Russian missiles, both instantly recommended total annihilation as the mathematically preferable strategy, leading the Pentagon to reject them. Because fast ballistic missiles such as Russia's Satan II leave little time for human review, he says China, the US and Russia are all considering automating nuclear response decisions, but are now cautious about it given the risk of misunderstanding, bad interpretation or hacking. He adds that military AI is rarely developed under vetted, secure conditions and is therefore easy to compromise, a vulnerability he has written about for 15 years. In intelligence analysis, he says the choice of AI system matters less than having an analyst who asks the right questions.
Q&A highlights
Asked about Israel's defense-tech reputation, Gaycken distinguishes tactical uses, such as identifying a missile in a static scenario, where AI can add functionality, from dynamic, multi-domain battle situations, where AI tends to get overwhelmed and unusable. Asked about Ukraine, he says Western arms suppliers use the conflict as a field-testing ground, but most of what is deployed there is conventional advanced weaponry rather than deep tech, with Ukrainian hacking efforts limited mainly by a lack of human resources.
Quotes
AI mostly comes from consumer electronics and the advertising industry, that's a well-known fact, and the defense industry mostly buys its AI and its cyber security from consumer electronics and advertising.
Sandro Gaycken
In defense and intelligence you will be hacked, we do this all the time.
Sandro Gaycken
In Germany you're not punished for not doing anything, you're only punished for failure.
Sandro Gaycken
Questions answered in this talk
Why is military AI and deep-tech procurement so slow in Germany and Europe?
According to Sandro Gaycken at Rise of AI 2023, Germany's federal procurement office takes about 18 months just to review a submission and requires at least 18 to 20 parties to sign off, often changing the outcome far from what was originally requested; he says the same problem exists across Europe, the UK and the US.
Does military AI mostly come from defense research?
No. Sandro Gaycken says at Rise of AI 2023 that, contrary to popular belief, AI used in Western defense mostly comes from consumer electronics and advertising, and the defense industry largely buys it rather than developing it in-house.
Should AI be used to automate nuclear response decisions?
Sandro Gaycken says at Rise of AI 2023 that major powers are exploring automating nuclear responses because of the speed of modern ballistic missiles, but have become cautious after a Pentagon simulation in which two commercial AIs both recommended total annihilation, given the risks of misunderstanding, misinterpretation and hacking.