Verifying content produced by AI: method and levels of control

Analysis note

Asmaa EL FORKANI 4 min read

A text produced by an AI assistant can be fluent, structured and wrong. Verification is not a precaution of principle: it is a condition for using the content. Its intensity depends on that use, and its method rests on a few repeatable checks, from the dated fact to the legal reference.

ASKonsult – Strategy & Operations

Observation: fluent text and the need for verification

Language models produce the most probable text in view of the instruction and the documents provided. However, they do not verify what they assert. The OWASP security framework for applications based on these models ranks misinformation among the ten main risks. It describes four manifestations of it. These are factual inaccuracies, unsupported claims and apparent expertise that the model does not possess. The fourth consists of code suggestions that are insecure or call on non-existent libraries.

Excessive confidence on the part of the user aggravates the risk. Indeed, the user takes up content without verification because it appears well mastered. Legal references, figures and quotations are the most exposed elements. In fact, a model may invent a law number, a date or an article with the same assurance as an accurate fact.

Analysis: proportioning verification to use

Three levels are enough to cover most situations:

  • N1 – Proofreading: internal use without commitment (draft, rewording, outline); critical reading for coherence and tone.
  • N2 – Fact-checking: content serving as the basis for work or a decision. Verification of facts, figures, dates, names and references against identified sources.
  • N3 – Validation by a competent person: external dissemination, commitment of the organisation, decision concerning a person. Review by a qualified person authorised to modify the content.

The choice of level precedes use; it does not follow it. Thus, content produced for internal use changes level when it ends up being disseminated.

Analysis: five verification checks

  • Facts and figures: go back to the primary source (official text, report of the issuing institution, statistical database) and note its date. A secondary source is not sufficient for a figure reused as it stands.
  • Legal references: check that the text exists, its number and its date. Next, check its publication in the Official Bulletin or the Official Journal. Finally, check the article cited and its wording in force at the date of use.
  • Quotations: find the exact passage in the original document; if it remains untraceable, remove the quotation.
  • Reasoning: redo the calculations, then check that the conclusion follows from the premises and that the text states its assumptions.
  • Omissions: identify what is missing (exception, condition, time limit, opposing view). Indeed, an answer may be accurate in what it says and misleading in what it leaves out.

In addition, a confidentiality check completes this verification before any dissemination. The content must not reproduce personal data or internal information that may have appeared in the documents provided to the tool.

Analysis: disclosing the origin of the content

Transparency about the use of AI is becoming a requirement. The EU AI Act imposes a disclosure obligation on deployers that publish text generated or manipulated by AI. That obligation applies where the purpose is to inform the public on matters of public interest. However, an exception applies where the text has undergone a process of human review or editorial control. The exception also requires that a person hold editorial responsibility for it (Article 50(4)).

In the context of electoral operations, the CNDP has requested the flagging of any content produced by AI. Similarly, Anthropic’s Usage Policy applicable from 12 November 2026 will set requirements for recommendations in high-risk domains. In that case, it requires review by a qualified person and information of the persons concerned.

Implications: documenting verification

Verification leaves a documented record: sources consulted, date, person who verified, level adopted. This record protects the organisation, which remains responsible for the content it disseminates, whatever the tool that produced it. In addition, the instructions given to the tool can facilitate verification, for instance by asking for sources or flagging uncertainties. However, they never replace it.

Points of attention

  • Set the level of verification before any use of the content.
  • Check every legal reference in the official text, in its wording in force.
  • Remove any quotation or reference whose source remains untraceable.
  • Redo calculations and check dates.
  • Check for the absence of personal data or internal information before dissemination.
  • Keep a record of the sources consulted and of the person who verified.

References

See also: Guide – Good and bad practices for generative AI at work; Library – Prompt templates by function; Templates – Artificial intelligence use charter (company, association, public body). Information Systems & Digital Transformation practice.

General information document; it constitutes neither legal advice nor an opinion on a particular situation. References to legal texts should be checked against the version in force on the date of use.

Consulting firm not registered with any Bar; no representation before the courts and no act reserved to the regulated legal professions.