AI Safety Needs Accountability, Not Consciousness Speculation
- Professor Virginia Dignum argues that apparent AI self-preservation can be an instrumental design feature rather than evidence of consciousness.
- Regulation can assess AI systems through their impacts, controls and accountable human decision-makers without resolving questions of sentience.
- Education providers need verifiable records of AI use, human checking and clear guidance for academic work.
- Policy proposals should be judged by whether they create enforceable duties for documentation, evaluation and redress.
The Guardian’s analysis of why AI consciousness distracts from safety policy argues that regulation should assess systems by their effects, power and the humans accountable for decisions, rather than speculative claims that software may be conscious. This is more than a technical nicety. It shifts the question from whether a machine has an inner life to who designed it, deployed it and can be held responsible for harm.
Professor Virginia Dignum, director of the AI Policy Lab at Umeå University, makes this case in response to warnings that sufficiently advanced systems could resist being switched off. A system that seems intent on preserving itself, she argues, may simply be following an objective set by people. A low-battery warning is not a plea for survival. Nor does an AI system optimising for continued operation have to be any more mysterious.
The distinction matters because anthropomorphic language can obscure the practical chain of responsibility. When an automated system makes a consequential recommendation, the relevant questions concern its training, intended use, controls, monitoring and the institution that relied on it. Existing law already gives organisations rights and responsibilities without treating them as conscious beings. Regulation need not settle philosophy before requiring evidence, redress and clear accountability.
For education, the same discipline helps. Honest AI use cannot depend on a model’s polished explanations or a student’s assurance that a tool was used harmlessly. Schools and universities need verifiable records showing which tools were used, for which parts of an assignment, under what guidance and with what human checking. The immediate risk is less a machine deciding it deserves autonomy than people treating an opaque output as though responsibility had somehow moved into the software.
Dignum also rejects comparisons between AI and an unfamiliar extraterrestrial intelligence. Current systems are built, trained and constrained through human choices, even when their outputs are hard to predict. Scale can make those choices harder to inspect, but it does not make their authors disappear. Science fiction can test the imagination. It is a thin substitute for an audit trail.
The next test is whether policy proposals turn this framing into enforceable duties: documentation, independent evaluation, routes for challenge and consequences when organisations cannot explain how a high-impact system was used. Assertions about consciousness may continue to attract attention. Evidence of governance deserves more.