Findings
Each row measures a different task, so the rates are not one average.
| Source | What was measured | Result | Link |
|---|---|---|---|
| Trakkr | 1,285 statistics quoted by five AI assistants, each followed from the cited page toward a study, up to five pages deep. | 40% reached a study that says the same thing. The 95% range is 34% to 46%. 8% reached a study whose number or context had changed. 40% stopped at a source that was locked, gone or silent on the number. 11% led nowhere. 1% would not load. | Trakkr report |
| Washington University | 98,020 verifiable claims from Google AI Overviews, checked against the text of the cited pages. | 11.03% were not backed by those pages: 6.98% with no mention in them, 2.66% contradicted by a cited page, and 1.39% where the cited pages disagree. | PaperUniversity release |
| Tow Center | 1,600 queries asking eight generative search tools for the article, publisher and URL of a quoted news excerpt. | Incorrect answers to more than 60% of the queries. | Tow Center report |
| BBC and the EBU | 2,709 news answers from four consumer AI assistants, scored by journalists from 22 public service media organizations. | 45% had at least one significant issue. 31% had a significant sourcing problem. 20% had a significant accuracy problem. | BBC and EBU report |
| FActScore | Human-checked atomic facts in generated biographies of 183 people, against English Wikipedia. | 58.3% of the atomic facts in ChatGPT's biographies were supported. For InstructGPT, 42.5% were. | FActScore paper |
The university release says the 11% figure should be treated as an upper limit. The paper says that if every inconsistent claim that also cited an unscanned social or video page had been supported there, the rate would fall from 11.0% to roughly 5.3%.
Training content
Training and compliance videos cannot contain made-up facts. A procedure, a policy or a safety step has to match the document a reviewer will open. If the cited page does not say the same thing, the line does not go in the lesson. One lesson on this site keeps each line next to the source it came from.
How Leap works
Your documents, as narrated videos. Every fact from your source. Ask Leap a question, or give it your documents, slides, audio or video recordings, or all of them together. For questions, Leap pulls facts only from trusted sources. Either way, it builds a source document that you approve before anything is made. Then it writes the script, draws the visuals and narrates a captioned video.
Questions
Does the cited page contain the number?
In the Trakkr check, the page an assistant cited had the figure 82% of the time. Only 21% of the pages cited for a number were the original research, and 54% of the first pages cited were company blogs.
What if the assistants agree on a number?
Agreement is not a check. Trakkr found four statistics that all five assistants quoted, and could not find a study behind any of them.
How should I read the 11.03% on AI Overviews?
The university release says the 11% figure should be treated as an upper limit. The paper says that if every inconsistent claim that also cited an unscanned social or video page had been supported there, the rate would fall from 11.0% to roughly 5.3%. On 41.9% of the overviews, every claim the study could check was supported by the cited text. The university release calls that share conservative, because some cited pages could not be fully collected.
Can I put an AI answer in a compliance video?
No. Training and compliance videos cannot contain made-up facts. A number whose source says something else, or a claim the cited page does not contain, does not belong in the lesson either.
What does Leap do with a question?
You approve a source document before anything is made. Every fact in the video ties back to the source you signed off on.