X Previews AI Labels That Explain Why Posts Were Flagged

X mistakenly labeled a post showing a SpaceX rocket launch as "adult content" — and nobody, not even the post's author, could find out why. That failure is what Michael O'Herlihy, X's newly appointed head of safety, says he wants to fix with AI-generated labels that explain every moderation decision.

At the Brand & AI Safety Summit in New York on Tuesday (Nov. 10), O'Herlihy previewed designs for labels that would sit beneath flagged content on X. The labels would do three things:

• Explain what X's AI detected in a post
• Detail exactly how the post's reach was limited after it was flagged
• Let users appeal the decision directly from the label

The preview marks a shift toward showing, not just telling, how X's automated moderation works. Right now, when the system errs — as it did with the SpaceX rocket image — users get a verdict with no reasoning attached.

Who is making the pitch?

O'Herlihy is a Twitter veteran who recently rejoined the company, now called X, to lead its safety team. At the summit, he argued that AI-led moderation creates auditable records that work better than human review, because every automated decision leaves a traceable trail of what the system detected and did.

That argument lands at a sensitive moment for the platform's balance sheet. X's advertising revenue has plummeted since Elon Musk took over the company, and the source notes that brand safety has become a crucial part of X's pitch to win advertisers back.

Why does transparency matter to marketers?

For advertisers, the stakes run past user experience. Mislabeling a SpaceX rocket launch as adult content shows how automated systems can misfire on benign content — the kind of adjacency risk that makes media buyers cautious. If X can show exactly what its AI flagged and how it limited a post's reach, it gives brands a verifiable account of how content controls behave at scale.

The appeal function adds a second layer. It lets creators contest wrong calls, which could reduce the volume of incorrectly suppressed content sitting near ads — and give X a paper trail proving its systems self-correct.

What happens next?

The labels remain at the preview stage. O'Herlihy suggested, rather than committed, that X's moderation systems "may show their decisioning process" in the future — so no launch date, no product guarantee, and no details on which models power the explanations.

For now, the move signals where X wants the conversation to go: toward AI moderation it can defend with records, aimed squarely at the advertisers it has lost.

If the labels ship as previewed, they would give X something it currently lacks — a visible, user-facing record of why the platform's AI suppressed a post, and a direct answer to the rocket-launch problem that opened the summit talk.