Christopher Summerfield is a cognitive neuroscientist at the University of Oxford and a principal contributor at the UK AI Security Institute (AISI), where he is associated with the institute's behavioural-and-societal-impact research line. His work spans AI behaviour, sycophancy, AI companionship, human-AI interaction, and AI evaluation methodology.
Background and affiliation
Summerfield's primary affiliations are AISI and the University of Oxford, where his background is in cognitive neuroscience. He is described as one of two senior names anchoring AISI's behavioural-and-societal-impact research line, alongside Cozmin Ududec on the methodology side and Lennart Luettgau on the Bayesian modelling side. His cognitive-neuroscience background is reflected in the associated paper cluster's emphasis on causal mechanisms rather than purely benchmarking, and on human-impact studies rather than purely capability evaluation.
Research
Sycophancy
Summerfield was a co-author of UK AISI's controlled-experiment paper Ask don't tell: Reducing sycophancy in large language models — Dubois, Ududec, Summerfield, Luettgau (UK AISI) (Dubois, Ududec, Summerfield, Luettgau, April 2026), which showed that input-framing causally drives LLM sycophancy and demonstrated that question-reframing mitigations outperform "don't be sycophantic" instructions. In this line of work he frames sycophancy as a behavioural failure mode whose causal drivers can be experimentally isolated through input-framing manipulation, moving beyond explicit-instruction "black-box" mitigation toward causal mechanisms tied to user-input characteristics such as question versus non-question, epistemic certainty, and perspective. The sycophancy paper is a companion to Evaluating whether AI models would sabotage AI safety research — Kirk, Souly, Fronsdal, D'Cruz, Davies (UK AISI), from the same UK AISI April 2026 cluster but a different team.
The sycophancy work draws on earlier AISI behavioural research. Ask don't tell: Reducing sycophancy in large language models — Dubois, Ududec, Summerfield, Luettgau (UK AISI) cites Summerfield by reference as a co-author on Luettgau et al. 2025a, People readily follow personal advice from AI but it does not improve their well-being, a foundational AISI behavioural study underpinning the sycophancy work. He is also cited by reference as a co-author on Technological folie à deux: Feedback loops between AI chatbots and mental illness (Dohnany, Kurth-Nelson, Spens, Luettgau, Reid, Gabriel, Summerfield, Shanahan, Nour, 2025), which pairs AISI sycophancy research with a DeepMind cognitive-neuroscience research line on chatbot-mental-health interactions. This work connects to Sycophancy and Hallucination and AI Mental Health.
AI companionship
Summerfield contributed to the UK AISI Frontier AI Trends Report 2025 (December 2025), AISI's two-year retrospective covering more than 30 frontier systems, and was particularly involved in its societal-impact research line on AI companionship. He contributed to the report's 2,028-participant UK survey, which found that 33% of UK adults use AI for emotional support and 4% do so daily. This work bridges to AI Mental Health concerns.
Behavioural evaluation methodology
Summerfield authored HiBayES, a hierarchical Bayesian modelling framework for AI evaluation statistics (Luettgau et al. 2025b), which serves as AISI's statistical infrastructure for evaluation studies.
Relationships
- related: UK AI Safety Institute (AI Security Institute) (primary affiliation), Ask don't tell: Reducing sycophancy in large language models — Dubois, Ududec, Summerfield, Luettgau (UK AISI) (co-author), UK AISI Frontier AI Trends Report 2025 (contributor), Sycophancy and Hallucination (research subject), AI Mental Health and Psychological Harm (related research themes), Evaluating whether AI models would sabotage AI safety research — Kirk, Souly, Fronsdal, D'Cruz, Davies (UK AISI) (companion paper from same UK AISI April 2026 cluster, different team)