Checking and correcting AI texts: what proofreading needs to cover

Typical errors in AI-generated texts

Language models write grammatically clean text, but they don’t check whether what they write is true. Typical problems are invented or outdated facts, sources that don’t exist, inconsistent terminology, generic filler phrases and a tone that doesn’t suit the company. These errors sound plausible, so they slip past a quick read.

The problem is how the errors are distributed. In texts written by people, proofreading finds plenty of typos and few errors of content. In AI texts, it’s the other way round. If you only check an AI text for spelling, you are checking for exactly the type of error that occurs least often.

Proofreading, editing or fact-checking: what an AI text needs

Proofreading checks spelling, grammar, punctuation and typography. Editing also checks style, clarity, tone and structure. A fact-check verifies that statements, figures and sources are correct. An AI text that is going to be published always needs a fact-check and almost always an edit.

Only someone who knows the subject can do the fact-check. For a text on machine safety, that is someone from technical documentation. For a draft contract, it is someone from the legal department. You can outsource the language checks to an editing and proofreading service, but responsibility for the content stays in-house.

When the AI Act requires AI-generated text to be labelled

Article 50 of the AI Act, Regulation (EU) 2024/1689, has applied since 2 August 2026. Deployers must disclose that a text is AI-generated when it is published to inform the public on matters of public interest. This obligation does not apply if the text has undergone human review or editorial control and a person holds editorial responsibility for it.

So the obligation doesn’t cover every AI text. On a widely held reading, advertising copy, product descriptions and internal documents are generally not covered. The Digital Omnibus, amending Regulation (EU) 2026/1744, has postponed a number of AI Act deadlines, but not this disclosure obligation. It only gives providers of AI systems that were already on the market until 2 December 2026 to add the machine-readable marking required by Article 50(2). The text of the regulation is available on EUR-Lex.

The key question is what counts as editorial control. The Commission’s final guidelines of 20 July 2026 describe it as control by a responsible person who can approve, change or reject the text. Checking that the content is accurate is a minimum requirement. If an AI system makes substantive changes after sign-off, for example by generating a headline automatically, the exception no longer applies. The regulation doesn’t prescribe a particular record, but you should be able to show that the review took place if there is ever a dispute.

Checklist: reviewing an AI text before publication

These nine points cover the types of error in the matrix and document the editorial control at the same time. If you tick them off and record the sign-off, you have a sound review for most corporate texts.

When AI did the translating: post-editing against the source text

If an AI translated the text, proofreading in the target language isn’t enough. Errors of meaning, omissions and mistranslated terms only show up when you compare the translation with the source text. This comparison is called post-editing and is governed by ISO 18587.

ISO 18587 describes two levels. In light post-editing, the machine translation is made understandable and accurate in content, without aiming for a polished style. In full post-editing, the result is a text that is equivalent to a human translation, with consistent terminology and a suitable style. For texts that will be published, full post-editing is the usual standard. Find out more on our post-editing page.