Dilemma
Should we hold AI systems responsible for what they do?
When an autonomous AI takes an action that harms someone, the question of who or what is responsible — the developer, the operator, the model itself — turns on whether the model is the kind of thing that can be a responsible agent.
Context
Self-driving cars, content-moderation systems, autonomous trading bots, agentic LLMs deployed in software development and scientific research — AI systems are taking consequential actions in the world with increasingly less direct human supervision. The legal and regulatory frameworks are scrambling to assign responsibility: to manufacturers, to operators, sometimes to the model itself, sometimes nowhere. Behind the policy debate is an older question: is an AI the kind of thing that can be the origin of an action, or only the kind of thing through which other agents' choices ripple out?
Why it matters
If AI agents can be responsible in the strong sense, then they can be punished, restricted, granted standing — and the moral weight of building them is the moral weight of bringing responsible beings into existence. If they can't, then the responsibility for what they do falls fully on humans, and the tendency to talk about 'the AI's decision' is mystification. The question is the same one asked of addicts, of corporations, of children, of trained animals: what conditions make an actor an agent?
The coordinates that split the schools
The stances
An AI's behaviour is fully determined by training and input; 'responsibility' applies if at all to its makers.
13 schoolsOn this view, the AI's output is a function of its training data, its architecture, and the input it received. There is no extra fact about the AI that could ground its responsibility, because there is no extra fact about the AI that could have produced different behaviour given the same conditions. Responsibility runs back to whoever set up those conditions.
Where this stance leads ⓘ
- 1% Choice is structural illusion — every event is fixed by the prior state. on Do you really choose?
- 1% The addict's behaviour is the outcome of causes; 'responsibility' is a useful fiction, not a metaphysical fact. on Are addicts responsible for their addiction?
- 1% Truth is mind-independent, universal, accessible in principle to all. on Is truth universal, tradition-bound, situated, or constructed?
- 1% Yes — matter was created and is conserved as a real substance. on Is the world created from nothing?
- 1% Yes — the physical world is fully real, mind-independent, persisting. on Is the physical world fully real?
The AI can be a genuine agent within determined conditions — and therefore genuinely responsible.
14 schoolsOn this view, what makes a being responsible is not indeterminism but the kind of process the being is. An AI that deliberates, considers consequences, can be given reasons, and modifies its behaviour on reflection is doing what responsible agency is, even if its underlying process is fully determined. The responsibility ascription is functional, not metaphysical.
Where this stance leads ⓘ
- 1% Choice is real within a determined order — agency and determinism aren’t opposites. on Do you really choose?
- 1% The addict is genuinely responsible within a determined order. on Are addicts responsible for their addiction?
- 1% Truth is mind-independent, universal, accessible in principle to all. on Is truth universal, tradition-bound, situated, or constructed?
- 1% A person exists from conception — when a new being comes into existence. on When does a person begin?
- 1% Marriage has a given form — it’s a kind of thing we recognize, not make. on What is marriage?
An AI without a free will is not the kind of thing that can be responsible.
85 schoolsOn this view, real responsibility requires real choice — a moment at which the agent could have gone either way and was the genuine originator of what it did. Current AI systems aren't that; they are very sophisticated downstream processes. Whether some future AI could be is an open question, but the bar is the same one ordinary human agency has to clear.
Where this stance leads ⓘ
- 1% The future is open and you are a genuine origin of it. on Do you really choose?
- 1% The addict could have chosen otherwise — that's why recovery is real. on Are addicts responsible for their addiction?
- 1% Causation runs one way — the arrow of time is real and structural. on Could causation work backwards?
- 1% The asymmetry is real because time itself has a real direction. on Is the asymmetry between memory and anticipation a real feature of time, or just of us?
- 1% The arrow is real and structural; the asymmetry isn't an artifact of description. on Is the arrow of time a real feature of the cosmos, or only of how we describe it?
Neither AIs nor anyone else are the locus of free agency; the question is the wrong one.
10 schoolsOn this view, the same reasons that undermine ordinary claims of human agency apply with equal force to AI. The brain is a coin-flipping organ; the model is a function on inputs. Neither is the kind of thing that can be the source of action in the strong sense. Responsibility is a useful social practice; not a metaphysical fact about either humans or AIs.
Where this stance leads ⓘ
- 1% Even if the universe is undetermined, you are not the chooser. on Do you really choose?
- 1% Even if the universe is undetermined, the addict isn't the chooser. on Are addicts responsible for their addiction?
- 1% Subject to a real natural order we did not make. on What is our place in nature?
- 1% Nature includes its limits; colonisation is bounded by what the cosmos allows. on Should we colonize space?
- 1% Biology is what it is; we modify it within real biological constraints. on Is genetic engineering of food stewardship or domination?
Schools the coordinates don't place
These schools don't satisfy any stance's coordinate pattern strongly enough to be assigned — either because they decline to commit on the question (Confucianism is famously silent on what comes after; Pyrrhonian and pragmatist traditions suspend judgment), or because their attribute signature crosses categories in a way the five buckets don't capture.
Related Experiments
Experiments engaging the same dimensions as this dilemma — they\'re where the same questions get stress-tested in cleaner cases.
Related Films
Films engaging the same dimensions as this dilemma.
Related Historical Debates
Historical confrontations where parties argued out questions on these dimensions.