Not all AI-generated content carries the same disinformation risk. A labelled, sourced summary on a low-reach channel is very different from an unlabelled, mass-produced claim pushed to a large, trusting audience. The AI Disinformation Risk Analyzer scores the structural factors that turn AI output into a disinformation hazard and suggests concrete mitigations.
How it works
You describe the AI system or the specific piece of generated content, then characterise the audience and distribution channel. The tool scores five factors that research and regulators (including the EU Code of Practice on Disinformation) consistently identify: verifiability of the claims, source transparency, whether AI involvement is disclosed, the manipulation/scale potential, and the reach and trust of the audience and channel.
These combine into an overall risk profile — low, moderate, or high — with a short rationale and a prioritised list of mitigations for the factors that scored worst.
What each factor measures
Verifiability assesses whether the claims in the content can be checked against named, accessible sources. A news summary that cites a named study or official statement scores well; a confident-sounding claim about a specific event with no sourcing whatsoever scores poorly. Unverifiable claims are the raw material of disinformation because they cannot easily be corrected.
Source transparency looks at whether the origin of the content is clear. Does the reader know where the claim comes from? Is the generating system’s identity disclosed? Opaque sourcing combined with authoritative-sounding claims is a common pattern in high-risk content.
Disclosure of AI involvement is one of the single highest-leverage factors, which is why it is scored separately. Content that is clearly labelled as AI-generated — in a visible disclosure, not a buried footnote — cannot mislead the reader about its nature. Content that reads as human-authored when it is not exploits the reader’s implicit trust calibration.
Manipulation and scale potential considers whether the system can produce content at volume and whether that content is designed to influence beliefs or emotions rather than inform. A recommendation engine that personalises political news to individual beliefs at mass scale is high risk; a single summary paragraph is low risk.
Reach and audience trust accounts for where the content lands and how much the audience trusts the channel. High-reach, high-trust channels amplify disinformation more effectively than low-reach alternatives, and audiences with limited contextual knowledge about a topic are more susceptible to confident but wrong claims.
How to read the output
The five-factor profile produces an overall risk level, but the per-factor scores tell you where to focus. A content piece might score low on verifiability but high on disclosure (it is clearly labelled) — the risk is contained not because the claims are checkable but because the AI-generated nature is visible. A different piece might have perfect sourcing but no disclosure, scoring moderate overall because the deception risk is still present. The mitigations are prioritised to the factors that actually drove the risk score, not a generic list.
What to do with the results
The most common high-priority mitigations across most content types are:
- Add clear AI disclosure — a visible statement that the content was generated with AI, before the content itself, not in small print elsewhere.
- Attach provenance metadata — C2PA content credentials can be read by detection tools and allow the content’s origin to be verified programmatically.
- Add human review before publication on high-reach channels — a reviewer who fact-checks the claims and confirms accuracy before wide distribution.
- Link claims to specific, named sources — rather than stating facts without attribution.
Tips and notes
The single highest-leverage mitigation is almost always disclosure — labelling content as AI-generated and attaching provenance metadata (C2PA) — because it costs little and directly addresses the deception vector. The second is human review before high-reach publication. A low score is never a substitute for fact-checking the actual claims; it only tells you the structural risk is contained. Everything runs locally in your browser and nothing is uploaded.