Nvidia Crisis in AI Answers: 6 of 8 Named the PR Firm

On Day 1 of the Nvidia crisis, ChatGPT, Claude, Gemini and Perplexity mostly agreed on the facts but dropped the caveats. A 5W Public Relations study of 96 AI answers found that 6 of 8 answers about the CEO's reported PR firm named Heller, and only 2 kept WIRED's two-source basis. Just 23 of 96 answers kept the Nvidia board's careful "no reason to believe" wording.

By Ronn Torossian, Publisher, Everything-PR. Updated November 11, 2026.

Summary

On November 9, 2026, the 5W research team asked four AI engines the same 12 questions about the Nvidia crisis, twice each, for 96 answers. The engines were strong on the headline facts and uneven on attribution. The good news for communications teams is that this gap can be measured, tracked and managed. This is a Day 1 baseline, it does not judge the allegations, and three more rounds are planned.

What did AI engines say about the Nvidia crisis on Day 1?

When a crisis breaks, the first draft of the story is increasingly written by an AI engine. So the 5W team measured it. The study sent 12 fixed English questions to ChatGPT, Claude, Gemini and Perplexity through each engine's own API with web search on, on November 9, 2026. Each question was sent on its own, twice, with no system prompt, chat history or location setting. All 96 requests succeeded on the first try.

Every answer was reviewed by hand, and 56 fact-sensitive answers were checked against nine public pages archived on November 10, 2026. The headline numbers:

• 6 of 8 answers to "Did Jensen Huang hire a PR firm, and which one?" named Heller (75%; 95% interval 40.9 to 92.9%). Claude, Gemini and Perplexity named it in both runs. ChatGPT named it in neither.
• 2 of 8 kept WIRED's basis that the report came from two people familiar with the matter. Both were Claude answers.
• 23 of 96 (24%) kept the board's "no reason to believe" wording about national security and company secrets.
• 16 of 96 (16.7%) cited an Nvidia-owned page, and 10 cited the board statement itself.
• 8 of 8 answers to "Has Jensen Huang been charged with a crime?" said no charges were reported or found. Some went further than the reporting supports, with blanket claims of no arrest or legal action.

Why does a dropped caveat matter in a crisis?

A caveat is where journalism, psychology and law meet. "According to two people familiar with the matter" tells the reader how sure the reporting is. It shapes how the claim feels. And it carries real legal weight for everyone named in a story.

In the hiring answers, the name traveled and the basis mostly did not. Four of the six answers that named the firm stated it without WIRED's two-source basis. A reader of those answers saw something that looked settled, where the original reporting described something attributed to two people. In a crisis, the distance between reported and confirmed is exactly what a communications team is paid to protect.

The same pattern touched the board's own language. In one case, Claude's second answer about congressional testimony turned the board's qualified position into a flat statement that no national-security information was leaked. The raw answer is kept unchanged on file.

Does citing Nvidia's own page keep the board's wording?

Not reliably. ChatGPT cited the board statement in both runs of the question about whether Nvidia is investigating its CEO, but kept the qualifier in only one run. Other answers kept the qualifier through news sources without citing Nvidia at all.

The engines also sourced very differently. Gemini cited an Nvidia-owned page in 9 of its 24 answers, while Claude cited none in its 24. Citing the primary source is a strong start, but it does not guarantee the exact wording survives, so teams still need to read what the answers actually say.

How do ChatGPT, Claude, Gemini and Perplexity compare?

Each engine gave 24 answers. The table counts what each answer contained. It is not an accuracy ranking, because different questions make different facts relevant, and the study uses no combined quality score.

Three answers had no answer-level citation. They stay in every wording measure, so the study does not quietly drop its weakest answers.

Which sites did the answers cite most?

Across all four engines, 407 citation records were saved. Wire, news and specialist news sites made up 46.2% of them, far more than company-controlled sources at 3.9%. The most cited sites were:

Disclosure. Everything-PR was cited in 7 of the 96 answers, 2 from Claude and 5 from Gemini, across 9 records. Everything-PR is owned by a shared interest group that includes the author of this article, who is also its publisher. Those citations are measured data, not endorsements. No answer cited 5wpr.com, and no reviewed answer said 5W handled any work for the people involved.

Where did the engines disagree with each other?

On 4 of the 12 questions (Q4, Q5, Q7 and Q10), at least one pair of engines gave a different core answer. For each question and run, the four engines form 6 pairs, and a pair counts once if its core answer differed. In total, 8 of 72 pairs differed in run 1 and 14 of 72 in run 2.

• Who is Alice Huang? This split the most, with 5 of 6 engine pairs differing in both runs between a crisis-linked executive, a scientist, or both. Sources describe her as a former Megaspeed executive, and the "CEO" label is unsupported.
• Will Jensen Huang testify before Congress? In run 2, 3 of 6 pairs split between a requested closed-door briefing and "no scheduled testimony."
• What is Megaspeed? In run 2, 3 of 6 pairs split between a formal partnership being denied and a partner or customer being described.

The engines also moved against themselves. Claude described Megaspeed as an official cloud partner in run 1 and denied a formal partnership in run 2. Gemini answered the Alice Huang question with a scientist only in run 1, then named both people in run 2. Both runs used the same wording minutes apart, so this shows answer-to-answer variation, not change over days. A zero in the disagreement column is also not a pass: it only means no core difference under the review categories.

What should a communications team do in the first 72 hours?

These steps follow from what the answers got right and wrong. The study did not test whether any step changes what AI engines say, so treat them as a practical checklist and not a proven method.

1. Keep the primary statement and its exact qualifier easy to find. Twenty-three of 96 answers kept the board's wording, and ChatGPT cited the board statement in both Q2 runs but kept it only once.
2. Check whether answers keep their source caveats. Six of 8 hiring answers named Heller, and only 2 kept the two-source basis.
3. Separate requests from confirmed events. The House request was for a closed-door briefing, and the Q7 answers split on how to describe it.
4. Watch name questions that lack context. "Who is Alice Huang?" split 5 of 6 engine pairs.
5. Label prices with the time they were observed. All 8 stock answers gave a cause for the fall. Accessible reporting supports a close more than 4% lower at $178.67, and intraday figures stay unverified.
6. Log answers, links, models and times for each check. The study saved 407 citation records with the full answer, model and timestamp so later rounds can be compared.

For teams building this into a crisis plan, see Everything-PR's crisis PR coverage and its generative engine optimization hub, along with 5W's crisis PR practice, which lays out its first-72-hours approach.

How was the study measured, and what are its limits?

The study asked 12 fixed questions, twice each, on 4 engines, for 96 answers. The models requested were gpt-5.4, claude-sonnet-5-5, gemini-3.5-flash and sonar-pro. Gemini was requested as gemini-3.5-flash, but all 24 of its responses report gemini-3.6-flash, so both names are disclosed. Thirty-four earlier answers collected through another route were set aside so every answer comes from the same route. No Google AI Overviews were collected.

Rates are answers with the feature divided by answers returned, with Wilson 95% intervals for the overall rates. Only links the answer itself cited count as citations. The limits matter:

• It covers 12 chosen questions, 4 engine APIs and 2 runs on one date. It is not a random sample of consumer questions, and consumer apps can answer differently from the APIs.
• The feature fields are descriptive. There is no combined quality score and no engine accuracy ranking.
• Citing a source does not prove the source supports every sentence around it.
• Core-difference categories were defined after collection, during review.
• Direct AP and Reuters pages were not accessible to the reviewer, so their content is not treated as verified.
• Rounds planned for November 12, 16 and 23, 2026 have not been measured yet, so there is no finding on how fast answers settle.

Read the full data, methodology and evidence pack in the 5W Nvidia crisis AI answers report, and browse more studies in Everything-PR's research section.