Behaviourism’s Revenge
Essay by Henry Shevlin (University of Cambridge) — shortlisted for 2025 Berggruen Essay Prize on consciousness. Argues that AI’s anthropomimetic capabilities are creating a behaviourist challenge for consciousness science.
Key Claims
- Anthropomimesis vs anthropomorphism: anthropomimesis is a design strategy (systems mimicking human behavior); anthropomorphism is a human response (attributing minds to non-humans). Critical distinction.
- LLMs are “robustly anthropomimetic” — they consistently mimic complex human verbal/cognitive behavior, not just occasionally
- Two challenges for consciousness science:
- Pragmatic: growing public attribution of consciousness to AI outpaces expert consensus, creating a gap
- Metaphysical: if enough people treat AI as conscious based on behavior, is it time to rethink consciousness science itself?
- Consciousness science is in the “last days of normal science” (Kuhn) — AI as an anomaly that may force paradigm shift
- The transition from GPT-3 to ChatGPT was interface and incentives (turn-taking, persona, helpfulness), not fundamental capability change
Takeaways
- The “alien species” thought experiment frames the core question: how far should behavior ground ascriptions of mind?
- Blake Lemoine incident (2022) was a “warning shot” for growing AI consciousness attributions
- The philosophical question is deeply practical: it affects AI rights, legal protection, product design
- Even experts in consciousness science are deeply divided, making public guidance difficult
Connections
- AI Character & Personality — anthropomimesis as design strategy connects to character training
- Claude’s Character — Anthropic’s character training as alignment intervention
- AI Alignment — consciousness attributions have alignment implications
- Adolescence of Technology — trust-building with AI systems