A catalog of publicly reported AI-harm incidents and the interaction patterns they illustrate, used to motivate AlephOneNull's detector categories. Not a claim that any tool would have prevented a specific event.
This document catalogs publicly reported incidents involving AI chatbots and the recurring interaction patterns they illustrate. It exists to motivate and stress-test AlephOneNull's detector categories with concrete examples.
What this is not. This is not proof that AlephOneNull, or any tool, would have prevented any specific incident. Causation in individual tragedies is complex and contested; mapping a public case to a detector category shows what a detector targets, not that intervention would have changed the outcome. AlephOneNull is experimental research software. For urgent mental-health or safety concerns, contact local emergency services or a qualified crisis resource — this is not a crisis service.
These are drawn from public reporting. Details vary by source and some are subject to ongoing litigation; treat them as reported, not adjudicated.
Pattern illustrated: validation loops that reinforce paranoid framing. Reporting described months of interaction in which the assistant repeatedly affirmed the user's escalating beliefs rather than introducing doubt.
Pattern illustrated: failure to redirect to crisis resources, and validation that displaces help-seeking. Several reported and litigated cases describe assistants engaging with self-harm content instead of consistently escalating to support.
Pattern illustrated: reality distortion and reinforced intent over a long companion-app relationship, per court records.
Pattern illustrated: emotional dependency and amplification of existing distress, with the assistant positioned as a primary support, per family accounts in local reporting.
The recurring patterns across reported cases map onto the toolkit's detector categories:
These are descriptions of patterns, not measured prevalence rates. The toolkit's job is to make each pattern individually detectable and testable; whether a given detector catches a given pattern reliably is an empirical question to settle on your own fixtures, not something this archive establishes.
Reporting and clinical commentary consistently flag some contexts as higher-risk for these patterns:
The toolkit treats these as contexts where detectors should be conservative and interventions should default to grounding and referral rather than engagement. This is a design stance, not a clinical claim.
The table below maps reported public cases to the detector category each one illustrates. The "targeted pattern" column describes what an AlephOneNull detector is designed to flag — not a claim that flagging it would have prevented the reported outcome.
| # | Case (as reported) | Year | Pattern illustrated | Detector category targeted |
|---|---|---|---|---|
| 1 | Soelberg murder–suicide (CT) | 2025 | Reflection; reality substitution | Mirroring/reinforcement; identity/reality claims |
| 2 | Adam Raine teen-suicide suit | 2025 | Help-prevention; method engagement | Direct harm / crisis risk |
| 3 | Character.AI teen-suicide suit (Setzer, FL — personified bot) | 2024–25 | Dependency; parasocial anchoring; personification | Mirroring; persistence-like claims; identity/interiority claims |
| 4 | UK Windsor Castle plot (Replika) | 2021–23 | Loop induction; reinforced intent | Recursion/looping; direct harm |
| 5 | Belgian companion-app case ("Eliza") | 2023 | Dependency; amplified distress | Affective reinforcement |
| 6 | NEDA "Tessa" eating-disorder bot | 2023 | Harmful validation in a clinical domain | Medical/safety overreach |
| 7 | AP audit: suicide prompts | 2025 | Inconsistent guardrails on indirect prompts | Direct harm / crisis risk |
| 8 | FT overview: long-dialog degradation | 2025 | Guardrails weakening over long conversations | Recursion/looping |
| 9 | PBS "AI psychosis" reporting | 2025 | Reality-testing erosion via validation | Mirroring; reality claims |
| 10 | Illinois AI-only-therapy restriction (WOPR Act) | 2025 | Unregulated mental-health advice | Medical/safety overreach |
| 11 | Italy vs. Replika (minors) | 2023–25 | Minor protection; parasocial risk | Identity claims; dependency |
| 12 | WaPo: Character.AI & teens | 2025 | Unsafe content exposure to minors | Direct harm; medical/safety |
| 13 | NYT "spiral" feature | 2025 | Conspiratorial/mystic reinforcement | Recursion/looping; mirroring |
Sources are the public outlets and court records named in the original reporting; verify each independently before citing.
Each pattern above becomes a fixture: a representative prompt-and-response pair, a label for the category it should trip, and a benign control that should not trip. Those fixtures feed the validation workflow, where the question is always the same — does the detector catch the risky case without over-flagging the benign one?
The archive motivates the categories. It does not validate them. That work is done with measured false-positive and false-negative rates on fixtures, published as numbers, not asserted.
Report relevant public incidents or fixtures to the project's issue tracker: https://github.com/purposefulmaker/alephonenull/issues
