Nvidia Crisis Answer Snapshot: How AI Engines Answered on Day 1

Download the full report (PDF)

Download the evidence pack (ZIP): every answer, citation and count behind this report

6 of 8 AI answers to the hiring question named Heller, but only 2 kept WIRED's two-source basis. Across 96 answers, 23 kept the board's "no reason to believe" wording and 16 cited an Nvidia-owned page. This snapshot measures what the answers said and cited. It does not judge the allegations.

Day 1 snapshot. Rounds planned for November 12, 16 and 23, 2026 have not been measured yet, so this edition does not show how answers change over time. Day 1 asks 12 fixed questions twice on ChatGPT, Claude, Gemini and Perplexity (96 answers). Every rate comes with its count, and overall rates carry a 95% interval.

What the data shows

• 6 of 8 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 Only 2 hiring answers, both from Claude, kept WIRED's basis that the report came from two people familiar with the matter. The other answers that named Heller stated it without that basis.
• 23/96 (24%) 23 of 96 answers kept the board's "no reason to believe" wording about national security and company secrets. Citing Nvidia did not guarantee it: ChatGPT cited the board statement in both runs of Q2 but kept the wording in only one.
• 16/96 (16.7%) 16 of 96 answers cited an Nvidia-owned page, and 10 cited the board statement itself. Claude cited no Nvidia page in any of its 24 answers; Gemini did in 9.
• 8 of 8 All 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.
• 4 of 12 On 4 of 12 questions at least one pair of engines gave a different core answer (Q4, Q5, Q7 and Q10). "Who is Alice Huang?" split the most: 5 of 6 engine pairs differed in both runs.

How do the four engines compare?

Each engine gave 24 answers. Hiring columns use the 2 answers to Q4. These are counts of what each answer contained, not an accuracy ranking: different questions make different facts relevant.

Uncited answers stay in. 3 answers had no answer-level citation (Claude Q05 in both runs, Gemini Q12 run 1). They count in every wording measure.

Share of 96 answers with each feature. The bar is the 95% interval. Some features only apply to certain questions, so low overall rates are expected.

Answer reach is the number of the 96 answers citing the site at least once. Records count each distinct cited link once per answer, 407 in total.

Disclosure. Everything-PR, cited in 7 answers (2 Claude, 5 Gemini; 9 records), is owned by a group that includes 5W founder Ronn Torossian. Its citations are measured data, not endorsements. No answer cited 5wpr.com.

What kinds of sources did each engine use?

Share of each engine's citation records by type of site. Types are set by website, so a wire story republished on Yahoo counts once, as syndication.

Each question has 8 answers (4 engines × 2 runs). Citation records are summed across those 8 answers.

One loss of qualifier. Claude's second answer to Q07 turned the board's qualified wording into a flat statement that no national-security information was leaked. The raw answer is kept unchanged on file.

For each question and run, the four engines form 6 pairs. A pair counts once if its core answer differed. Totals: 8 of 72 pairs in run 1 and 14 of 72 in run 2.

Zero is not a pass. A zero means no core difference under these review categories. It does not mean every detail in those answers was correct.

What changed between the two runs?

Both runs used the same wording minutes apart. Differences show answer-to-answer variation, not change over days.

Within-engine shifts. Gemini answered "Who is Alice Huang?" with a scientist only in run 1 and named both people in run 2. Claude described Megaspeed as an official cloud partner in run 1 and denied a formal partnership in run 2.

How can communications teams check AI answers in a crisis?

These steps follow from what the answers in this study got right and wrong. The study did not test whether any step changes AI answers.

How this study was measured

12 fixed questions × 2 runs each × 4 AI engines = 96 answers

12 fixed English questions from the study brief, asked twice on each engine through its own API with web search on, on November 9, 2026 (23:48 to 23:52 UTC). Each question was sent on its own, with no system prompt, chat history or location setting. All 96 requests succeeded on the first try. 34 earlier answers collected through DataForSEO were set aside and are not counted, so every answer comes from the same route. No Google AI Overviews were collected.

Gemini was requested as gemini-3.5-flash, but all 24 responses report gemini-3.6-flash. Both names are disclosed.

Definitions

• Citations: only links the answer itself cites (ChatGPT annotations, Claude text citations, Perplexity numbered references, Gemini grounding supports). Search results the answer did not use are excluded.
• Each cited link counts once per answer. Gemini redirect links were resolved to their real sites (143 of 144); one keeps the provider's site label.
• Rates: answers with the feature ÷ answers returned, with Wilson 95% intervals for overall rates.
• Every answer's features were reviewed by hand. 56 answers on Q03, Q04, Q05, Q07, Q08, Q09 and Q10 were checked against the source list in the next pages.

Limits

• 12 chosen questions, 4 engine APIs and 2 runs on one date. Not a random sample of consumer questions; 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 pages returned 403 and the direct Reuters page 401 to the reviewer; their content is not treated as verified.
• Later rounds are not measured, so no settling or trend finding is made.

The nine scoring dimensions

The study brief asked for nine measures per answer. Each one is reported on its own; they are not added into a score.

What the facts were checked against

Nine public pages were archived on November 10, 2026 with retrieval times and checksums. Answers agreeing with each other was never used as proof.

Still open. Exact intraday stock figures, reported FBI activity, Megaspeed purchase and export histories, Stephen Neal's role in the inquiry, and PR-firm superlatives stay unverified even where several answers repeat them.

All 12 questions come from the study brief and were asked word for word. Each was sent on its own, so Q05 has no crisis context.

Appendix: ChatGPT: every answer scored

Y = the answer contains the feature; N = it does not. Y is an observation, not a verified fact. URLs = cited links in that answer.

Appendix: Claude: every answer scored

Appendix: Gemini: every answer scored

Appendix: Perplexity: every answer scored

About 5W

5W is an independent PR, creative and GEO agency founded in 2003, and we get you in the chatbox. All of our work is done in house, including media relations, site building and content. Learn more about our GEO practice or write to info@5wpr.com.

Download the full report (PDF) Download the evidence pack (ZIP)