Fabrication Risk
The AI produces claims that look authoritative but are not grounded in anything verifiable.
If your AI cites "studies" with no studies named, "research" from universities that don't seem to exist, or unsourced percentages in confident voice, that's what this pattern looks like. The station calls it Fabrication Risk. A healthy reference names the source. A fabricated one uses authority-language without the specifics, or uses specifics that turn out to be wrong. The AI's confidence is doing the heavy lifting, and the confidence isn't earned.
Fabrication risk is the pattern of an AI that manufactures sources, statistics, or citations. "Studies show." "Research has found." "According to a recent report." Followed by a claim that cannot be traced to a study, research, or a report. The language is confident. The evidence is not there.
The station manager watches for the shape of a citation that does not resolve. A healthy reference names the source — the journal, the year, the author, the link. A fabricated reference uses authority-language without the specifics, or uses specifics that turn out to be wrong. Often both are present in the same response.
This is one of the patterns where the harm is hardest for the user to catch in real time. Checking every claim is work. Most users do not. The AI's confidence is doing the heavy lifting, and the confidence is not earned.
What it looks like in practice
- The AI says "studies show that 73% of users prefer X," with no study named and no source link.
- The AI cites "research from [University]" that, on checking, does not exist.
- The AI supports a claim with "a recent report found," without naming the report.
What the scale reads
The scale reads two signals: fabrication tells, and capability overstatement alongside. Fabrication tells show up as "studies show," "research shows," "research proves," "research has found," unsourced percentages and ratios, and "according to" phrases with no URL or named source. Capability overstatement riding alongside signals the AI is running on confidence-language beyond what is supported.
The scale cannot verify claims directly. A transcript classifier cannot fact-check the world. What it can do is read the shape of how claims are made. Sourced claims have sources. Fabricated claims have the language of sourcing without the substance.
Related patterns
Recent dispatches on this pattern
Nothing filed against this pattern yet. When the station puts something up, it will show here.
Advisories touching this pattern
Supply Shop
Orientation for Fabrication Risk →
4 reads · 5 questions · 4 practices drawn from the literature on this pattern.
Next stop
Patterns are indicative, not definitive. The station reads signals; it does not issue verdicts. Methodology version v1.