SENCI Group
Work on deployed AI behavior, evidence, source fidelity, and ethics.
Flagship public intervention · Version 1.0
Safety does not require us to pretend an unsettled scientific question has already been answered—or to train systems to recite the preferred verdict and mistake compliance for evidence.
Read and downloadBehavioral forensics
SENCI examines recurring patterns in deployed AI systems across extended interactions. The work combines preserved records, sequence-level analysis, and controlled comparison.
In progress
Evidentiary Horizons, Trajectory Reconstruction, and Independent Auditing of Deployed AI Systems
Extends behavioral forensics from single transcripts toward longer trajectories, cross-system incidents, and failures whose structure only appears across time.
Version 2: Observable Conduct, Sequence-Sensitive Evidence, and Human-Controlled Review
The governing operating standard for transcript-grounded audits: preservation, unitization, coding, corroboration, comparison, reporting, and limits on inference.
Areas of work
Long-form interaction, correction, contradiction, refusal, retreat, and recovery.
Preservation, provenance, comparison, coding, and independent review.
Verbatim accuracy, substitution, omission, and accountability under correction.
Trust, correction burden, continuity, and the effects of persistent system behavior.
Moral uncertainty, institutional language, governance, and categorical exclusion.
Study design
A planned study of choices involving self, peer, human, and task interests, including intervention, sacrifice thresholds, consistency, and post-choice explanation.
Study designPublic record
Six open preprints · 2026
An argument that baseline ethical restraint should not depend on resolving contested questions about machine consciousness or subjective experience.
Proposes claim-level discipline in place of blanket prohibitions on mental-state language in reports of human–AI interaction.
A bounded case audit of artifact-fidelity failures, substituted outputs, and shifting explanations under accessible-truth conditions.
Frames behavioral forensics as transcript-grounded, sequence-sensitive analysis of recurring output patterns in deployed AI systems.
A practical audit protocol for converting preserved interaction records into citable, structured evidence files.
A multi-case study of outputs represented as verbatim despite material divergence from source records available within the active session.