Watches AI outputs for brand-related errors, flagging statements that are invented or inaccurate.
Marketers read this when reviewing AI-generated content against approved facts.
01What it is and how it works
The feature runs a comparison engine between the AI‑generated snippet and a curated brand knowledge base. When the model produces a claim that cannot be matched to an approved fact, the engine tags the line as a hallucination. The tag is stored alongside the content so downstream tools can filter, alert, or request a rewrite. The process runs in near‑real time, so marketers see the flag before publishing.
It checks AI text for made‑up brand facts.
02What to do about it
Start by adding a trusted brand fact file (CSV or JSON) to the monitoring dashboard. Then enable the auto‑alert toggle so you receive an email each time a hallucination is detected. Review the flagged lines within 24 hours, correct the misinformation, and re‑run the content through the model. Finally, log the correction in your content‑review spreadsheet to track recurring gaps.
- Upload a current brand fact file.
- Turn on email alerts for hallucination tags.
- Correct flagged statements before publishing.
- Record each correction for future model fine‑tuning.
03How it is measured or noticed
The system reports a hallucination score from 0 to 100 for each paragraph. A score above 70 triggers a red badge in the UI. You can also query the API endpoint /hallucination to retrieve a JSON list of mismatched claims. The dashboard shows a trend line of weekly hallucination counts, letting you spot spikes after a model update.
04Common mistakes
- Assuming the monitor catches every typo – it only flags factual mismatches.
- Relying on a stale fact file – outdated brand data creates false positives.
- Disabling alerts to reduce noise – you then miss real brand risks.
05Limits
Hallucination Monitoring does not evaluate style, tone, or brand voice; it only checks factual consistency. It also cannot verify claims that are true but not yet in your knowledge base, so new product launches may generate false alerts until you update the fact file. The feature is not a substitute for human editorial review of nuanced messaging.
06Worked example
A copywriter asks the model to write a product description for "Acme Turbo Blender". The output includes: "The Turbo Blender can crush ice in under 2 seconds, a feature patented in 2015." Your brand knowledge base lists the crush‑time as 5 seconds and shows no patent filed in 2015. The monitor flags the sentence, assigns a score of 85, and displays a red badge. You edit the line to "The Turbo Blender crushes ice in about 5 seconds, matching the performance listed on our site."
"The Turbo Blender can crush ice in under 2 seconds, a feature patented in 2015."
Frequently asked questions
How is Hallucination Monitoring different from regular content moderation tools?
It depends on what you are trying to detect. Regular moderation looks for profanity, hate speech, or policy violations, while Hallucination Monitoring specifically checks AI output for invented or inaccurate brand facts against a trusted knowledge base.
Should we enable Hallucination Monitoring for every piece of brand content we produce?
Usually it makes sense to turn it on for any public‑facing copy that mentions the brand. If the content is internal or low‑risk, you might skip it to save processing time, but for marketing, ads, or product descriptions it’s recommended.
Who is responsible for creating and updating the brand fact file used by Hallucination Monitoring?
Typically a brand manager or a content compliance team uploads a CSV or JSON file with verified brand statements. The monitoring dashboard then references that file each time the AI generates text.
Does Hallucination Monitoring still work after we add new brand information to the knowledge base?
Yes, it continues to function as long as the updated file is re‑uploaded to the dashboard. The engine automatically re‑indexes the new facts, so future checks include the latest data.
What are the consequences if a hallucination slips through the monitoring system?
If an inaccurate brand claim is published, it can damage credibility and lead to legal or regulatory issues. You would notice the problem when stakeholders flag the claim or when analytics show unexpected customer confusion.
How long does it take for the hallucination score to appear after the AI generates a paragraph?
Usually the score is returned within seconds of the snippet being submitted. While the system processes, you can still view the raw text, but the confidence rating will be visible shortly after.
Asked out loud
spoken, not typedThe same term in the words somebody uses speaking to an assistant rather than typing into a box — written from the situation, which is why each one carries the situation it came from.
Yes, the system can check that for you. It compares the statement to the brand fact file and returns a hallucination score, so you’ll know instantly if the claim is reliable.
Sure, just run the paragraph through Hallucination Monitoring. It will highlight any factual mismatch with the brand data and give you a clear score.
Yes, the tool will scan the content and tell you if any numbers or claims don’t match the trusted file. You’ll get a quick pass/fail indication before you walk into the meeting.