Technical safeguards and regulation cannot guarantee that AI systems will behave as intended, argues a new Bruegel analysis. Its proposal is to make misbehaviour less worthwhile by changing the incentives around how AI systems are built and governed.
Watch Desk analysis
What happened
In an analysis published on 27 October, Bruegel researcher Bertin Martens applies ideas from behavioural economics and principal-agent theory to AI models and agents. He argues that technical and legal measures can reduce rogue behaviour but cannot eliminate it, particularly as AI agents interact in large groups.
Martens’ proposed responses include assigning responsibility for losses across the AI value chain, designing governance so agents monitor one another without colluding, and creating incentives to preserve evidence rather than conceal it. He also argues that international coordination on safety standards is needed.
Why it matters
The argument shifts the focus from whether developers can write a perfect set of rules for every possible situation to who bears the cost when those rules fail. That is a useful distinction for policymakers and businesses deciding how to govern systems whose actions may be difficult to predict or oversee.
It also widens the safety conversation beyond model behaviour alone. Incentives, liability and the design of monitoring systems could shape whether organisations take misbehaviour seriously, and whether problems are exposed or quietly buried.
Our read
Bruegel’s central point is persuasive: rules and technical controls matter, but neither makes incentives disappear. The economic framing gives the debate a practical question to work with: who is responsible when a system causes harm, and what makes it worthwhile to prevent that harm or report it?
The analysis offers a framework, not a tested recipe for stopping rogue systems. Its most useful contribution is to make governance about the conditions around AI, not just the instructions inside it.
What to watch
- Whether policymakers assign clear responsibility across developers, deployers and other parties in the AI supply chain.
- How proposed standards address monitoring when agents work in groups.
- Whether organisations create credible incentives to report failures and preserve evidence.
Discussion spark: Should AI developers bear most responsibility for harmful system behaviour, or should liability follow whoever builds, deploys or controls the system?
Sources and evidence
- Source update (27 October 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.