Dale Caldwell, who resigned as New Jersey's lieutenant governor on September 25, 2026 after a state investigation found he sexually harassed a staffer and repeatedly violated ethics rules, went on PBS and argued his conduct did not amount to harassment, citing an AI chatbot's answer as support. The move pushes generati
Dale Caldwell stepped down as New Jersey's lieutenant governor on September 25, 2026, after a state investigation concluded that he sexually harassed a member of staff and repeatedly breached government ethics rules. In the weeks that followed, he began a round of media appearances aimed at clearing his name. In early October he sat for an interview on New Jersey's PBS news program, where the host pressed him on the investigation's findings. Caldwell redirected the conversation to artificial intelligence: he said he had entered an account of the events into an AI chatbot and asked whether such conduct constituted sexual harassment, and that the model answered no. He prese
Generative AI has been in mass use for nearly four years, and habits have shifted from looking things up to asking for judgments, which is the soil this episode grew in. In June 2023, a federal judge in New York fined two lawyers for citing fabricated cases produced by a chatbot, making model hallucination a matter of public record. Since then, law, medicine and mental health have all produced examples of users treating model output as professional advice. Political campaigns moved earlier still, drafting bill text, generating advertising and automating
Liu Gong argues that Caldwell's decision to invoke AI on camera was less a lapse in public relations judgment than a signal of where generative AI now sits in public life, and it exposes three real gaps.
From a technical standpoint, a large language model is in no position to determine whether sexual harassment occurred. It has no chain of evidence, no witnesses, no copy of the state ethics code, only whatever narrative the user types in. The research consensus is that conclusions shift substantially with phrasing, and that models exhibit sycophancy, tilting toward the answer a user appears to want. The only verified fact in Caldwell's account is that his question already contained its own answer.
On the business side, the episode hands AI vendors a problem with no ready solution. A sentence produced inside a chat window acquires public weight far beyond its reliability the moment a public official quotes it on television. Vendors currently rely on disclaimers and usage policies, but a disclaimer sits in an annex to a user agreement while the chyron reads that AI says it was not illegal. Those carry nothing close to equal weight.
For the industry, the real concern is that AI endorsement is becoming a rhetorical instrument. When a New York lawyer was sanctioned in 2023 for citing invented cases, the misuse came from a professional trying to save time. Now the user is a public figure whose motive is to locate a seemingly neutral third party. Liu Gong predicts clearer platform-level responses within twelve months, such as mandatory uncertainty notices in conversations touching law, medicine or personal accusations, or default refusals on questions about legality. These are patches applied after the fact. When someone prefers a chatbot to an investigative finding, the problem is no longer the model. It is why anyone would hand over the judgment in the first place.
- A 2023 Pew Research Center survey found that 58 percent of U.S. adults had heard of ChatGPT while only 14 percent had used it, meaning most viewers lack the experience to assess an AI claim quoted on television.
- OpenAI disclosed roughly 800 million weekly ChatGPT users in October 2025, several times the 100 million monthly users reported about two months after its November 2022 launch, widening the channel through which model output reaches public debate.
- In June 2023 a federal judge in New York fined two lawyers $5,000 each for citing fabricated c
For the industry, the immediate effect is that compliance and communications teams at AI vendors will be forced to answer a new question: where does responsibility sit when a single model reply influences a public event. Liu Gong expects tweaks to terms of service and product prompts within three months, without changes to model behaviour itself, and within a year a default redirect to professional advice in conversations about defamation, harassment or employment disputes.
For ordinary users, the near-term effect is perceptual: seeing a politician cite
Questions worth exploring:
- The precedent risk of public figures using AI to vouch for their own conduct
- Sycophancy and phrasing dependence as the technical mechanism behind the defence
- Responsibility and the regulatory gap once model output enters public discourse
This item comes from The Verge AI (published 2026-10-04 16:16:04). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.