Research

Language models are effective at interpretation, language and orchestration, and unreliable as instruments of analysis: asked what effect an intervention will have, they produce a plausible answer rather than a derived one. Formal causal methods have the opposite profile. They produce estimates that can be examined and defended, but they require carefully specified models and clean, well-structured inputs, which is where most applications of them stall.

This programme combines the two. The language model handles what it does well: interpreting the question, cleaning and normalising data, checking formats, assembling the inputs, and orchestrating the analysis. The causal model performs the analysis itself. The result is an estimate of the effect of a proposed intervention that is derived from an explicit model rather than generated, and that can therefore be explained, audited and compared against alternatives.

This work addresses two related questions: how much did a particular action actually contribute to a given outcome, and who should be expected to act, given what they were in a position to bring about? We develop formal measures for both — a measure of an action's causal contribution to an outcome, and a forward-looking measure of who is responsible to act, assessed before the outcome occurs rather than assigned as blame afterward. The work is currently under peer review.

The analysis takes an external, evaluative standpoint. It assesses what an agent did and what it could have done differently, rather than requiring access to the agent's own stated intentions — a distinction that matters for systems, human or automated, whose internal reasoning is not directly observable. The framework extends to situations involving several agents acting strategically, where responsibility depends not only on what an agent did in isolation but on what it would have cost the agent to act otherwise, on obligations and promises already in place, and on what the agent could reasonably expect other agents to do.

It also addresses situations in which an outcome results from the combined actions of many agents — a team, an institution, or a mixture of human and automated decision-makers — where no single contributor appears fully responsible under a simple causal account. This is a long-standing difficulty known as the problem of many hands. The framework offers a principled way to distribute responsibility across such a group, which is the sense in which the work is intended to contribute to responsible AI by design: allocating accountability as part of a system's specification, rather than reconstructing it after something has gone wrong. No RigoMind product currently implements the framework; it presently informs how we think about accountability in the systems we build.