Verification in AI-written copy

Generative models are fast at corporate copy and usually write cleanly. The risk is not that the writing is bad; it is that a sentence which looks right but has no source enters the text. Phrases like 'industry leader', 'thousands of customers' or '40 years since founding' appear at once and nobody asks where they came from.

In Turkey this is a regulatory matter, not only a reputational one. The regulation on commercial advertising and unfair commercial practices explicitly bans unprovable superiority claims and fabricated testimonials. Who wrote the text — a person or a model — does not change that responsibility.

The practical fix is one rule in the production flow: a number, date, title or testimonial enters the text only if it rests on a source. With no source the sentence is not written; a 'not found' note is left instead and the client is asked. Applying the rule shortens the copy but makes it defensible.

The second rule is that a model does not verify its own output. Asking the same model 'is this right?' merely produces the same assumption a second time. Verification comes from an independent source: the company's own document, an official record, a measured report.

The third is leaving a trail. If you note which sentence came from which document, the question 'where did we get this?' is answered in a minute six months later. Otherwise the copy becomes a pile of claims that has lost its own sources.