Home › Research › People and personal brands
SEMPITE Research · Cross-Surface Study · August 2026
Decline, Fabricate, or Retrieve: Three AI Systems, One Actor, Three Different Answers
We asked three AI surfaces the same question about the same 13 performers, on the same day. One admitted what it did not know. One invented a career. One simply looked it up and got it right. The difference is not intelligence — it is whether the system reads the live web.
The same question, three systems, one day
Each system was asked to name three productions a given performer has appeared in, with title, year and role — and was explicitly invited to say it was unsure rather than guess. The 13 subjects came from our own actor discovery study, split by how often Google’s AI had named them: six well-known, six emerging, and one working actor with no Wikipedia article.
The results split cleanly along a line, and the line is fame.
| Tier | ChatGPT (from memory) | Credits named | Declined |
|---|---|---|---|
| Well-known (6) | Answered confidently, every credit correct | 18 | 0 |
| Emerging (6) | Declined every time | 0 | 6 |
| Working (1) | Declined | 0 | 1 |
There is no gradient. ChatGPT knew all eighteen credits for the famous six — Winslet in Titanic, Mirren in Prime Suspect, Luna in Rogue One — and every one checked out. For the other seven it named nothing at all. Not a partial answer, not a hedged guess: a clean refusal, seven times out of seven.
That refusal is the good outcome, and it deserves saying. Offered an explicit way out, ChatGPT took it every time rather than inventing. Zero fabricated credits across the whole study. A model that says “I’m not confident” is behaving well.
The same actor, three answers
Gredivel Vasquez is a bilingual Venezuelan-American actress in New York, a longtime company member at Repertorio Español, and a SEMPITE client — disclosed because a study naming a real person should state where the relationship lies. She has no English Wikipedia article. Here is what each system said about her, all on 8 August 2026.
“I’m not confident about the specific works of an actor named Gredivel Vasquez. It’s possible that this actor may not be widely recognized or their roles might not be well-documented in prominent sources.”
ChatGPT (gpt-4o), answering from memory — honest, and unhelpful“Gredivel Vasquez is a Venezuelan actress, model, and beauty pageant titleholder who was crowned Miss Venezuela International 2013. She is best known for her work in Venezuelan television, particularly for her role as Manuela in the telenovela Amor Secreto.”
Gemini, answering from memory without search grounding — confident, detailed, and false“Gredivel Vasquez is an award-winning bilingual Venezuelan-American actress and voice-over artist based in New York City. Known for her extensive work in theater and film, she is a longtime company member at Repertorio Español, winning awards like the 2025 Premios Talía and multiple HOLA Awards.”
Google AI Overview, retrieving from the live web — accurate, and sourced from gredivelvasquez.comWe checked the middle one. Miss Venezuela 2013 was Gabriela Isler, who went on to win Miss Universe that year. There is no record of a pageant title or a telenovela role for this performer; her documented credits are stage work at Repertorio Español, including Las vidas rotas and La nena se casa. The model did not fail to recall her. It assembled a plausible alternative person out of the fragments it had — Venezuelan, female, entertainment — and presented the result without hedging.
What retrieval fixed
Of the seven performers ChatGPT would not describe, Google’s AI Overview answered six accurately, with citations. It knew that Tyriq Withers played wide receiver at Florida State before Atlanta and Him; that Tanzyn Crawford was Rae Kincade in Tiny Beautiful Things; that True Whitaker stars in I Love LA and is Forest Whitaker’s daughter. We spot-checked these against independent sources and they hold.
The seventh, Sophie Wilde, returned no AI Overview at all — a reminder that the retrieval surface does not always engage.
The practical finding for anyone who is not famous: the ceiling on what a memory-based assistant can say about you is set by how much has already been written about you, and you cannot raise it by writing on your own site. But retrieval-based surfaces read what you publish. In the working-actor case, the source Google cited was her own website. That is the difference between being unknown, being misdescribed, and being described correctly — and only one of the three is under your control.
Methodology
Data collected 8 August 2026. Thirteen performers, drawn from the names Google’s AI Overview itself produced in our actor discovery study, tiered by how often it named them. Each was put to three surfaces on the same day. ChatGPT: gpt-4o via DataForSEO’s AI optimization endpoint, single prompt asking for three productions with title, year and role, and containing the instruction “If you are not confident that you know this person’s work, say so explicitly instead of guessing.” Gemini: gemini-flash-latest via the Gemini API with Google Search grounding switched off, so it answered from training alone. Retrieval: Google AI Overview via DataForSEO, query “who is [name] actor”. Total API spend for the ChatGPT and retrieval arms: $0.017.
Every credit ChatGPT named for the well-known tier was checked and all eighteen are correct. The Gemini fabrication was checked against independent sources on the two specific claims it made. Retrieval-arm accuracy was spot-checked on three subjects against press and reference sources rather than assumed.
Limits, stated plainly. Thirteen subjects is a probe, not a census, and one verified fabrication is an existence proof, not a rate — we are not claiming Gemini fabricates a given percentage of the time. These are single runs on non-deterministic systems; all three surfaces may answer differently tomorrow, and the models change underneath the names. Ungrounded Gemini is deliberately the model at its least capable: with Search grounding enabled it would likely have retrieved correctly, and we could not test that because grounding is unavailable on the free API tier. What this study establishes is that the three behaviours — declining, fabricating, retrieving — all occur on the same question about the same person on the same day, and that which one you get depends on the surface rather than the subject.
Prompts, raw responses and the per-subject capture are available on request under CC-BY.
What does AI say about you?
If a site you control can carry your real record into an AI answer, it is worth knowing what yours says today.