AI Use Disclosure

How large language models are and are not used in producing CLCI Hub content, what the human-editorial-review pipeline looks like, and the rules that govern AI-assisted writing on this site.

What AI does for us

Summarisation of public material at scale. Long court judgments, government investigation reports, and book-length sources are summarised by LLMs as a first pass; a human editor then verifies every factual claim against the underlying source before it appears in a profile.
Drafting initial body text and structural scaffolding. Profile bodies, blog-post outlines, and glossary definitions are often drafted by an LLM based on a specific source list and editorial brief. Every draft is reviewed against its sources before publication.
Cross-referencing. Auto-derivation of related glossary terms, related blog posts, and related groups uses deterministic text matching rather than LLMs — but LLMs are sometimes used to propose candidate cross-links during editorial review.
Audit tooling. The duplicate-group audit, source audit, and content-quality audit scripts use deterministic regex and Jaccard-similarity heuristics, not LLM classification. Audit-output review (deciding what to act on) sometimes uses LLM assistance.

What AI does NOT do

No unreviewed AI output ships. Every claim on every profile is read by a human editor against its sources before publication. There is no “auto-publish” pipeline.
No AI-generated citations. LLMs are notorious for fabricating plausible-looking but non-existent sources (the “hallucination” problem). We do not allow LLM output to introduce sources we have not independently located in primary form. Where an LLM draft includes a citation we cannot verify, the citation is removed before publication.
No AI-generated URLs. Per the Source Policy, source strings either carry a verified URL or no URL at all. We do not fabricate URLs.
No AI-driven score changes. CLCI scores and confidence ratings are human-assigned per the methodology on the about page. LLMs are not used to score groups, to revise scores, or to assess BITE-axis evidence independently.
No user-facing AI tools. The site does not run any user-facing AI features (no chatbots, no AI Q&A, no AI-driven recommendations). This is a deliberate cost and editorial choice — we do not want ongoing inference costs, and we do not want users to receive AI-generated answers when they are consulting the site about a personally consequential situation.

How a typical profile gets built (LLM-assisted workflow)

A human editor identifies the target group and assembles a source list: court records, government reports, academic work, journalistic investigations, public ex-member testimony, organisational statements.
An LLM is briefed with the source list and the CLCI template structure, and drafts an initial profile (summary, body, red flags, timeline events, recovery resources, related groups).
The human editor reviews every claim in the draft against the underlying sources. Unverifiable claims are removed. Fabricated or hallucinated source citations are removed. Sensitive language is calibrated for compassionate, neutral framing per the editorial policy.
BITE-axis scores (0-10 each) and the signed modifier are human-assigned based on the verified evidence. The total CLCI score is the arithmetic sum, clamped to 0-40.
Confidence is assigned (High / Medium / Low) based on the source mix per the Source Policy.
The profile is committed to the Git repository. The build pipeline validates the data shape (every claim in the right field, every field within bounds) and rejects entries that fail validation.
Subsequent edits — correction requests, score updates, new sources — follow the same pipeline. The lastReviewed field is updated whenever a substantive factual review occurs.

Limitations of the AI-assisted workflow

LLM bias toward the source mix it sees. An LLM drafting a profile from primarily journalistic sources will echo those sources' framing. We mitigate this by mixing source types (court + academic + journalism + ex-member) and reviewing against the editorial-neutrality rules.
Date-of-knowledge limits. The LLM's training-data cutoff may pre-date recent court findings or news developments. The human editor checks the date of every claim and updates against the most recent public record.
Calibration drift. LLMs sometimes generate confidently-worded text about claims they have insufficient evidence for. The editorial review catches and removes overclaiming; readers should still treat confidence ratings as the primary indicator of evidentiary strength.
Stylistic homogenisation. LLM-drafted profiles can sound stylistically similar to each other. We accept this trade-off — consistent tone across 683 profiles is, on balance, a feature rather than a bug for a reference site.

What you can do about it

If you spot a factual error you suspect originated in unverified AI output (a fabricated source citation, a wrong date, an invented quote), please report it via the Corrections process. We log AI-originated errors specifically in our internal review notes so we can improve the editorial pipeline.

Changes to this disclosure

We will update this page if our AI-use practice changes substantively. As of the date below, the policies above reflect current practice. Substantive changes will be noted in the annual Transparency Report.

Last updated: 2026-05-19. See also: Editorial Policy · Source Policy · Corrections · Transparency Report · Legal Disclaimer

What AI does for us

Summarisation of public material at scale. Long court judgments, government investigation reports, and book-length sources are summarised by LLMs as a first pass; a human editor then verifies every factual claim against the underlying source before it appears in a profile.

Drafting initial body text and structural scaffolding. Profile bodies, blog-post outlines, and glossary definitions are often drafted by an LLM based on a specific source list and editorial brief. Every draft is reviewed against its sources before publication.

Cross-referencing. Auto-derivation of related glossary terms, related blog posts, and related groups uses deterministic text matching rather than LLMs — but LLMs are sometimes used to propose candidate cross-links during editorial review.

Audit tooling. The duplicate-group audit, source audit, and content-quality audit scripts use deterministic regex and Jaccard-similarity heuristics, not LLM classification. Audit-output review (deciding what to act on) sometimes uses LLM assistance.

What AI does NOT do

No unreviewed AI output ships. Every claim on every profile is read by a human editor against its sources before publication. There is no “auto-publish” pipeline.

No AI-generated citations. LLMs are notorious for fabricating plausible-looking but non-existent sources (the “hallucination” problem). We do not allow LLM output to introduce sources we have not independently located in primary form. Where an LLM draft includes a citation we cannot verify, the citation is removed before publication.

No AI-generated URLs. Per the Source Policy, source strings either carry a verified URL or no URL at all. We do not fabricate URLs.

No AI-driven score changes. CLCI scores and confidence ratings are human-assigned per the methodology on the about page. LLMs are not used to score groups, to revise scores, or to assess BITE-axis evidence independently.

No user-facing AI tools. The site does not run any user-facing AI features (no chatbots, no AI Q&A, no AI-driven recommendations). This is a deliberate cost and editorial choice — we do not want ongoing inference costs, and we do not want users to receive AI-generated answers when they are consulting the site about a personally consequential situation.

How a typical profile gets built (LLM-assisted workflow)

A human editor identifies the target group and assembles a source list: court records, government reports, academic work, journalistic investigations, public ex-member testimony, organisational statements.

An LLM is briefed with the source list and the CLCI template structure, and drafts an initial profile (summary, body, red flags, timeline events, recovery resources, related groups).

The human editor reviews every claim in the draft against the underlying sources. Unverifiable claims are removed. Fabricated or hallucinated source citations are removed. Sensitive language is calibrated for compassionate, neutral framing per the editorial policy.

BITE-axis scores (0-10 each) and the signed modifier are human-assigned based on the verified evidence. The total CLCI score is the arithmetic sum, clamped to 0-40.

Confidence is assigned (High / Medium / Low) based on the source mix per the Source Policy.

The profile is committed to the Git repository. The build pipeline validates the data shape (every claim in the right field, every field within bounds) and rejects entries that fail validation.

Subsequent edits — correction requests, score updates, new sources — follow the same pipeline. The lastReviewed field is updated whenever a substantive factual review occurs.

Limitations of the AI-assisted workflow

LLM bias toward the source mix it sees. An LLM drafting a profile from primarily journalistic sources will echo those sources' framing. We mitigate this by mixing source types (court + academic + journalism + ex-member) and reviewing against the editorial-neutrality rules.

Date-of-knowledge limits. The LLM's training-data cutoff may pre-date recent court findings or news developments. The human editor checks the date of every claim and updates against the most recent public record.

Calibration drift. LLMs sometimes generate confidently-worded text about claims they have insufficient evidence for. The editorial review catches and removes overclaiming; readers should still treat confidence ratings as the primary indicator of evidentiary strength.

Stylistic homogenisation. LLM-drafted profiles can sound stylistically similar to each other. We accept this trade-off — consistent tone across 683 profiles is, on balance, a feature rather than a bug for a reference site.

Changes to this disclosure