AI Reliability Incident Register — data dictionary and reuse notes
Version 1.5, 2026-10-15. 35 records: 22 studies, 10 findings, two reports, one allegation; 13 bias, 17 deception, five harmful conduct.
Canonical page: https://www.deterministicai.org/incidents/
License: CC BY 4.0 for original summaries and compilation; linked sources retain their own terms.
Suggested citation: Deterministic AI. AI Reliability Incident Register, version 1.5, 2026-10-15. https://www.deterministicai.org/incidents/ (add access date).
Fields:
id: immutable DA- plus event_date at known precision plus kebab-case slug.
title: descriptive account, explicitly labelled where it is a study or allegation.
event_date/date_precision/date_basis: known date and whether it is an event, finding or reporting milestone; not the import date.
organization/context/system/sector: named organization, setting and tested model; sector is legal, finance, health, government, consumer or other.
category: bias (systematic skew), deception (strategic misrepresentation, sycophancy or hidden actions without a false factual claim), harmful conduct (threats, dangerous advice, jailbreak-enabled harm or harm-related lawsuits).
evidence_status: finding (court, regulator, auditor or company confirmation), allegation (lawsuit/complaint), study (controlled research with setup), report (first-person or forum account with permalink).
what_happened/impact: attributed facts and limits. Allegations and individual reports are not established causation.
response/response_status/response_url: organization response where verified, or explicit absence from sources read.
prevention_control: editorial proposal; determinism alone does not establish safety or fairness. Controls need testing.
primary_source_url/source_scope/reviewed_at: read evidence, access limits and review date.
Scope split: false factual claims and invented material go to https://www.hallucination-free.org/tracker/.
Dates sort oldest first in exports; page is newest first. Filters do not remove records from static HTML.
Counts and chart describe this selected compilation, not prevalence or comparative model risk.
Sensitive records omit graphic or method details and personal names; self-harm entries include crisis resources.
Versioned files are frozen; each data change increments the minor version. Corrections preserve IDs and enter the changelog.

Previous release: v1.2, 2026-10-15. Added two company findings concerning unauthorized actions and one controlled study of concealed credential sharing from the second 15-case lying batch. Two false-claim studies went to the Tracker; ten briefs remain held. Current total: ten records (seven deception, three harmful conduct). The September Anthropic assessment covers one cluster of four incidents; its July operational disclosure is not a separate count. Updated downloads and category chart; earlier records and frozen exports preserved.

Previous release: v1.3, 2026-10-15. Added seven records from the third lying batch: five deception and two harmful-conduct accounts, comprising four studies and three company findings. Two confirmed false claims went to the Tracker; six forum cases remain held. Current total: 17 records (12 deception, five harmful conduct). Prompted evasion, reporting omissions, audience effects, concealed instructions and unauthorized actions retain their distinct evidence limits.

Previous release: v1.4, 2026-10-15. Fourth lying batch: added four deception studies (undisclosed review gaps, covert goals, coding-conflict nondisclosure and specification gaming) and one labelled StarSkirmish report. Current total: 22 records, with 17 deception and five harmful conduct; fifteen studies, six company findings and one report. Two false claims went to the Tracker and eight cases remain held.

Current release: v1.5, 2026-10-15. First bias batch: added 13 records (seven studies, four company findings, one Reuters report and one settled allegation). Two briefs remain held. Total: 35 records, with 13 bias, 17 deception and five harmful conduct. Preserved disputes, company responses, study limits and prior records; updated downloads and chart.