Verification Standards
Quick Reference ¶
| Tier | Name | Requirement | Published? |
|---|---|---|---|
| 1 | Juridical | Court documents OR government acknowledgment | ✓ Yes |
| 2 | Journalistic | 3+ independent sources OR primary document access | ✓ Yes |
| 3 | Preliminary | Single source or unverified | ✗ Tracked only |
Tier 1: Juridical Evidence ¶
The highest standard. These sources have been subjected to legal processes—sworn statements, adversarial challenge, or official governmental review.
Qualifying sources:
- Federal or state court filings (complaints, motions, rulings)
- Congressional or parliamentary testimony
- Regulatory body statements (FTC, state attorneys general, international DPAs)
- Law enforcement reports where AI is noted
- Coroner findings mentioning chatbot interaction
- Settlement agreements (when public)
What this means: If a case rests on Tier 1 evidence, someone has staked their legal or professional standing on its accuracy. Perjury, sanctions, and official misconduct attach to false statements at this level.
Tier 2: Journalistic Evidence ¶
The standard of responsible reporting. These sources have been subjected to editorial review, often legal vetting, and professional reputational stakes.
Qualifying configurations:
-
Three or more independent outlets reporting the same core facts
- Wire service reprints do not count as independent
- Must include at least one outlet with documented editorial standards
-
Primary source journalism from a single major outlet
- Reporter reviewed chat logs directly
- Reporter conducted original interviews with family members
- Documentary evidence (screenshots, records) obtained and verified
-
Platform acknowledgment
- Company confirms the incident occurred
- Even without details, confirmation establishes the event
What this means: Professional journalists and editors have determined the information meets publication standards. Outlets risk defamation liability and reputational damage for false reporting.
Tier 3: Preliminary Evidence ¶
Insufficient for publication but sufficient for monitoring.
Examples:
- Single news article without independent confirmation
- Social media accounts, however detailed
- Forum posts or community reports
- Secondhand accounts from non-family members
- International cases without English-language verification
What happens: Cases at Tier 3 are logged internally and monitored for escalation. If additional evidence emerges, they move to Tier 2 or 1. They are never published until that threshold is met.
Special Considerations ¶
Minors ¶
Cases involving victims under 18 receive additional scrutiny:
- Pseudonyms used unless family has publicly identified the child
- Age verified through multiple sources
- Educational or identifying details minimized
- Family wishes regarding publicity respected where possible
Active Litigation ¶
For cases in ongoing legal proceedings:
- We note that allegations are unproven
- We update as rulings occur
- We distinguish between "lawsuit filed" and "court found"
- Settlements are reported without implying admission
International Cases ¶
Non-US cases face additional verification challenges:
- Language barriers may limit source access
- Legal systems vary in public document availability
- We note when verification is limited by jurisdiction
- International cases may require longer verification timelines
Jurisdictional Verification Limits (Tier 2 sub-label) ¶
A case may meet the Tier 2 standard for journalistic verification — multiple independent outlets, often including a primary-source outlet — while still being limited by the inaccessibility of primary court documents to readers and reviewers of the database. We mark such cases with the jurisdictional-verification-limited sub-label.
When the sub-label applies:
- The case is verified across the required number of independent sources in at least one language we can read
- Primary court records (indictments, rulings, transcripts) exist but are not accessible in the language of publication, are not available via the courts' public portals, or both
- The verification rests substantively on professional journalism rather than on documents we have read directly
What the sub-label commits us to:
- A multilingual source sweep, with at least one published-language and one primary-language pass, where the primary-language pass is documented in the canonical record
- Explicit identification of the most authoritative single source (typically an indictment summary or court-records-based article) for the dates, charges, and substantive facts
- Re-verification when a primary document becomes accessible, with corrections to the record where prior secondary coverage diverges from the document
- An honest in-prose note that primary records are not accessible in the language of publication, so readers can weigh the source basis themselves
Worked example. The 2026-01-KR-001 Gangbuk-gu motel serial poisoning murders (Kim So-young, Seoul) is the first case in the database to carry this sub-label. The case is verified across 10+ English-language outlets and across major Korean broadcasters and dailies, but Korean criminal court records are not accessible to English-language verification. The Khan/경향신문 March 18, 2026 indictment summary serves as the most authoritative single source for the charged dates, substances, and ChatGPT exchanges. Two earlier "corrections" recorded in the canonical record on the basis of secondary aggregator coverage — the identification of the substance as zolpidem rather than benzodiazepines, and the placement of the series start in October 2025 rather than December 2025 — were both later reversed against the indictment. Cases at this sub-label tier are expected to undergo similar reversals when primary records reach the reviewer; the sub-label exists to mark that exposure honestly rather than to disguise it.
Platform Self-Reports ¶
When companies acknowledge incidents:
- Self-reports are weighted as Tier 2 minimum
- We note any discrepancies between company account and other sources
- Policy changes announced in response are documented
What Disqualifies a Case ¶
We do not publish cases where:
- AI interaction was incidental rather than central to events
- The only source is the platform's competitors or critics
- Family has requested non-publication and no legal record exists
- Evidence suggests misattribution or fabrication
- Harm was non-fatal and did not result in hospitalization
This is not a database of AI complaints. It is a record of verified deaths and severe harms where chatbot interaction was a documented factor.
Escalation Path ¶
Tier 3 (Monitoring)
↓
Additional source emerges → Cross-reference check
↓
Meets Tier 2 threshold? → Editorial review
↓
Publication with Tier 2 notation
↓
Court filing or government action? → Upgrade to Tier 1
Challenging a Verification ¶
If you believe a case has been incorrectly verified or should be reclassified:
- Open an issue on the project issue tracker with "Verification Challenge" label (or email contact@aimortality.org)
- Provide contrary evidence with sources
- We review within 7 days and respond publicly
- Outcome documented in issue thread and, if warranted, commit history
We have no interest in maintaining inaccurate records. Corrections strengthen the database.
The Standard We Hold Ourselves To ¶
Before publishing any case, we ask:
If the family of the deceased read this entry, would they recognize it as accurate and respectful?
If the platform's lawyers reviewed this entry, would they find it defensible as factual reporting?
If a researcher cited this entry, would it withstand peer review scrutiny?
Only when all three answers are yes does a case move from internal tracking to public database.
For the full epistemological framework and ethical commitments underlying these standards, see methodology.md.
Standards last revised: May 30, 2026
Rendered from the canonical markdown (single source): view the source file. Archived with the dataset at Zenodo. Questions or corrections: contact@aimortality.org.