AI captions that don’t invent facts

A caption model is good at continuing patterns. Give it a track, a mood, and an instruction to sound intimate, and it can produce the shape of a familiar studio story: three weeks of late nights, one broken microphone, a friend who saved the chorus. The sentences may fit the genre even when none of those details came from the artist.

OpenAI defines hallucinations as plausible but false statements. Its research also argues that standard training and evaluation reward guessing over acknowledging uncertainty. In a caption workflow, the same pressure appears when the prompt demands specificity but the context contains no specific fact. The model fills the empty slot because a complete anecdote sounds more useful than a blank. That is an explanation of the failure mode, not an excuse to publish it.

Voice and facts do different jobs. Voice can say that the artist writes plainly, avoids emojis, likes short sentences, or never calls listeners a family. Facts say what happened: the song title, a lyric, the mood, release state, confirmed collaborators, and dated events. A strong system gives the model both, but never lets voice manufacture facts.

Start with an allowlist. A caption may use the artist identity, the finished song and lyrics, the artist’s description of its mood, confirmed release information, and events stored on the artist timeline. Everything else is unknown until the artist supplies it. More prose in the prompt is not a substitute for a cleaner fact base.

The truth rule should be operational, not aspirational. Tell the model which fields are authoritative, name the details it must not infer, and define what to do when evidence is missing. The safe fallback is to omit the detail, write around it without implying it happened, or ask the artist a direct question.

For music captions, block unsupported claims in these categories:

• durations, counts, deadlines, and quantities;
• instruments, software, microphones, studios, and production methods;
• cities, venues, travel, weather, and recording locations;
• collaborators, friends, fans, labels, and other named people;
• shows, premieres, milestones, setbacks, and behind-the-scenes events;
• release status or calls to pre-save unless the confirmed date permits them.

“Out now” and “pre-save” are facts about timing, not decorations. Studio only allows release language close to a confirmed release: up to 30 days before and 14 days after. Its Post Generator combines artist identity with the song, lyrics, mood, release state, and current timeline context. That keeps a delayed release from inheriting an old countdown and stops an evergreen draft from claiming that a song just arrived.

The timeline also gives the draft a real reason to exist. A confirmed show, collaboration, release, or milestone can shape the angle. If nothing current is recorded, the caption can still discuss a lyric, feeling, or listening moment without inventing news.

TikTok’s current Community Guidelines say the platform does not allow misinformation that could cause significant harm to individuals or society. That is a narrower test than editorial truth. A made-up studio duration may not meet that harm threshold, but it still tells the audience something untrue about the song. The useful standard is therefore not merely “will a platform remove this?” It is “can the artist point to the source for every concrete claim?”

Before export, underline every concrete noun, number, date, place, person, event, and production detail. Match each one to artist data or remove it. Then check the temporal claim against the release date and read the draft aloud for voice. This review is small because the truth rule has already reduced the search area.

Artifex Studio prepares drafts, and nothing is posted until the artist approves it. That last review is not a ceremonial human-in-the-loop step. It is where lived experience remains owned by the person who actually lived it.