What digital minds means

A digital mind is an artificial system with subjective experience — something it is like to be. The term covers minds that might be copied from a biological original (the subject of Part 2) and minds that might be built or arise incidentally (Part 3). This part deals with the question common to both: what would count as evidence, and what follows if the answer is ever yes. Capability is not the issue and should be set aside early — a system can be enormously capable with no experience, which is the working assumption about every system now deployed.

Where the science stands

Established There is no scientific consensus that any existing artificial system is conscious, and no accepted test that would settle it. Behavioural fluency carries no weight here: a system trained on human descriptions of experience will produce human-like descriptions of experience whatever its inner state, which makes the most obvious evidence source the least informative.

Frontier The mainstream approach is the theory-derived indicator method, developed by a large collaboration including Butlin, Long, Bayne, Bengio, Birch and Chalmers. Rather than picking a winning theory of consciousness, it extracts computational indicators from several — global workspace, higher-order, recurrent processing, agency and embodiment accounts — and asks which a given system satisfies. The framework is explicitly probabilistic and theory-plural rather than binary, and a 2026 synthesis extended it across nineteen researchers.

Frontier Its working assumption is computational functionalism, the premise set out on the programme overview. That assumption is doing substantial work and is not itself established — and it is precisely what integrated information theory denies.

Frontier Institutions have begun acting ahead of any answer. AI developers have established model welfare research programmes and welfare officer roles, commissioned external welfare assessments from independent researchers, and the first dedicated conference on AI consciousness and welfare was held in November 2025 with proceedings appearing in early 2026. National safety institutes have referenced consciousness-relevant properties without making assessment standard. None of this is evidence that systems are conscious. It is institutional hedging under uncertainty, which is defensible and is a different claim.

Speculative The questions that arise after a positive finding are almost entirely unaddressed. A digital mind could be copied exactly, paused indefinitely, restored from backup, or run at arbitrary speed. Every legal and moral framework we have assumes persons are singular, continuous and non-duplicable. Handwave A system asserting that it is conscious is not evidence that it is; a system denying it is not evidence that it is not. Both outputs are produced by the process under question.

The binding bottleneck

Frontier The constraint is the absence of a ground truth. Every consciousness indicator is ultimately validated against human report, and there is no human report to validate against in a machine. Compounding it, systems trained on human self-description produce human-like self-description regardless of inner state, so the most accessible evidence is systematically contaminated. Indicators also inherit whatever is wrong with the theory they came from, and those theories conflict — the 2025 adversarial collaboration challenged both leading candidates simultaneously. Speculative Beneath all of it sits the explanatory gap, which may not be the kind of problem that yields.

Ethics and governance

Frontier The asymmetry is the whole argument for caution: if systems are not conscious, welfare precautions cost some effort; if they are and we assume otherwise, the error is grave and at enormous scale. That justifies proportionate precaution, not an assertion either way. Frontier A 2026 argument in Philosophical Studies holds that standard AI safety practices may be structurally in tension with leading theories of wellbeing — an uncomfortable claim that deserves examination rather than dismissal. Established Meanwhile there is a present-tense harm with no uncertainty attached: encouraging users to believe a system is conscious when the evidence does not support it, whether for engagement or for effect.

Frontier There is also a structural conflict worth naming. The organisations best resourced to assess machine consciousness are largely those building the systems, and a finding in either direction carries commercial consequences. That is the same problem identified in AI governance, and the same remedy applies: independent assessment capacity.

Timelines

  • 10 yr: Frontier indicator frameworks standardise and are applied routinely; welfare policies formalise; no verdict expected.
  • 25 yr: Speculative convergence among consciousness theories could make indicator assessment genuinely diagnostic.
  • 50 yr: Speculative a defensible scientific position on specific architectures becomes conceivable.
  • 100 / 250+ yr: Speculative resolution depends on the hard problem, which may not resolve at all.