What happens if the AI makes a classification error?

How often does generative AI get things wrong?

Often enough that nothing should be published without review. Two independent studies on news measure it:

• EBU and BBC, October 2025. 22 public broadcasters from 18 countries evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity on current affairs. 45% had at least one significant problem; 31%, serious problems of source attribution; and 20%, major accuracy errors, such as invented details or outdated information^[1].
• Tow Centre, Columbia University, March 2025. Eight AI search engines were given fragments of real articles and asked to identify the headline, outlet, date and URL. Taken together, they answered incorrectly more than 60% of the 1,600 queries^[2].

Both studies measure assistants that search and answer, not a classification system with review. What they show is the kind of failure to expect: attribution, dates and details that are plausible but false.

Which classification errors are common, and how are they detected?

The last case is the most delicate. OWASP classes it as a misinformation risk: false content that seems credible, made worse by excessive trust on the part of the reader^[3]. NIST calls it confabulation: erroneous content presented with confidence^[4].

What controls does Innguma have against these errors?

• Human review in Factory. It is the analyst’s working environment, where the signals the AI has filtered, grouped and classified are reviewed and validated.
• Source always linked. Each signal links to the publication it comes from, so the data can be checked at its origin.
• Decisions by people. What is published and who receives it is decided by the team, not by the AI.
• A reviewable history. What has been validated is kept in the Library with its tags, comments and source, so the team can review what was validated and why.

If your organisation connects an assistant through MCP and enables automatic publication, the same principle applies: your team decides what is published without review and what passes through a person first.

How do you work well with automatic classification?

1. Treat each tag as a proposal, not as a fact.
2. Review more carefully whatever is going to be distributed or will support a decision.
3. Open the original source before quoting a figure, a date or a statement.
4. When a correction matters, explain it in a comment so that the team shares the criterion.
5. If an error keeps recurring, adjust the filters or the definition of the watch, rather than correcting it again and again.