The thing everyone gets wrong about fact-checking a TikTok
A newsroom drops a viral TikTok into a transcript tool, then asks someone to Google the sentences. The transcript missed the claim that was only on screen. The sentence that said "this morning" gets checked against today, which is three weeks after the post. The verdict is confident and wrong.
The thing everyone gets wrong about fact-checking a TikTok is treating it like an article you can quote from audio.
Two clocks, and a frame the transcript never saw
Trust-and-safety and news desks already have a workflow. Download, transcribe, search, write a note. It fails on short video for reasons that are easy to see once you stop defending the workflow.
The first is the frame. A creator can put a number, a name, or a date in type and never say it. The spoken line is a reaction. If you only verify speech, you verified the reaction. The claim a viewer actually read was the overlay. Teams who already read on-screen TikTok text know this layer exists. Fact-checking that ignores it will clear videos that are false in the title card.
The second is the clock. "Today," "this morning," and "just happened" are relative to when the video was posted, not when your analyst pressed run. I have seen a true crash report marked unverifiable because the search was for this morning's news and the jet went down last month. The post date is part of the claim. Without it, web search answers a different question than the viewer heard.
The third is the citation. A model that says "false" with no URL is a vibe. A desk cannot ship a vibe. Each claim needs the wording, roughly when it appears, whether it was spoken or shown, a verdict, and the pages that support the rationale. Risk is separate from truth. A wrong celebrity nickname and a wrong medical dose are not the same note, even when both are false.
What triage is, and what it is not
This is assisted triage. It is not a newsroom standards desk, and it is not a legal finding. Verdicts are judgments grounded in pages retrieved at check time. A source can be wrong. A claim can be too thin to verify. Opinion should stay opinion instead of being forced into true or false.
What I want on the row is still concrete. A short summary, an overall risk, the claims split out, and links an editor can open. The post's author, caption, and time are context for those claims, not a substitute for them. If the video says "this morning," the rationale should say which morning.
Claims with no checkable fact should not be laundered into a verdict to make the dashboard look full. Unverifiable is a result.
Paste the video, keep the sources
How it runs:
- Paste public TikTok links.
- Speech, on-screen type, and the post date are read together.
- You get claims, verdicts, risk, and cited URLs.
I built the TikTok Fact Checker API for that triage queue. Run on Apify.
Who should use a queue, and who should not publish from it
This is for trust-and-safety teams and newsrooms who need a first pass on public videos before a person spends twenty minutes on each. It is for compliance folks who have to show what was claimed and which pages were consulted.
Do not publish or enforce from the row alone. If you need a human write-up of a comment thread instead of a video, that is closer to summarizing a Reddit thread. If you only need the overlay text and not a verdict, stay with text extraction.
Private or deleted videos will not check. Short clips with no factual claim should come back thin. A high-confidence verdict can still be wrong when the retrieved pages are wrong. Treat the output as a brief with footnotes. Input, pricing, and the limitation notice are on the Apify Store README for TikTok Fact Checker.
The transcript was a different story
The cost of the old way is an analyst who cleared the voiceover and missed the graphic, or who fact-checked last month's "this morning" against today's homepage.
Check the claim the viewer saw, on the day the viewer saw it. More writing lives on all posts.