New Jersey's Former Lt. Gov Used an AI Chatbot to Argue He's Innocent of Sexual Harassment

Dale Caldwell, New Jersey's former lieutenant governor who resigned on September 25th after an investigation found he had sexually harassed a staffer and repeatedly violated ethics rules, has been making media appearances to contest the findings — and in at least one interview cited the output of an AI chatbot as supporting evidence of his innocence.\n\n## Key takeaways\n\n- Dale Caldwell resigned as New Jersey's lieutenant governor on September 25th after an ethics investigation found he sexually harassed a staffer.\n- During a subsequent media interview, Caldwell cited AI chatbot output as evidence supporting his innocence.\n- AI language models generate plausible-sounding text by predicting likely word sequences — they do not investigate facts, access case files, or render legal judgments.\n- The incident illustrates a growing pattern: people treating AI-generated prose as an authoritative, neutral third-party verdict rather than a probabilistic text output.\n- Experts and journalists have repeatedly warned that AI chatbots confidently produce inaccurate or fabricated information, a phenomenon known as hallucination.\n\n## What an AI chatbot actually does — and doesn't do\n\nThe tactic is striking less for its audacity than for what it reveals about how AI-generated text is being perceived in the wild. Language models produce output by predicting statistically likely continuations of a prompt. They have no access to investigation records, witness testimony, or case files. When asked a leading question — "Wasn't Dale Caldwell treated unfairly?" — a chatbot will generate a fluent, confident-sounding answer shaped primarily by the framing of the question itself. That output is not a verdict. It is a mirror.\n\nThe Verge reports that Caldwell deployed this during an interview on NJ-area media, presenting the chatbot's response as though it carried independent investigative weight. It does not.\n\nThis matters beyond the political spectacle. The same quality that makes AI text generators useful for drafting image captions, story prompts, or character descriptions — their fluency, their apparent authority — is exactly what makes their output dangerous when mistaken for factual analysis. A model that can write a convincing villain's monologue can write an equally convincing exoneration. Neither reflects truth; both reflect craft.\n\n## The authority illusion and its aesthetic logic\n\nThere is something almost photographic about the phenomenon. Early portrait photography was treated as objective truth — the camera, unlike a painter, could not lie. Courts admitted daguerreotypes as evidence partly on that logic. AI-generated text is acquiring a similar unearned authority in public perception: it looks like neutral, considered prose, so it reads as neutral and considered.\n\nThe parallel is worth sitting with. Photographers spent decades dismantling the myth of the objective lens. The AI-text equivalent of that reckoning is only beginning. For creators who work with AI tools daily — building characters, writing prompts, generating image descriptions — the gap between "this sounds authoritative" and "this is accurate" is already well understood. Anthropic's Claude Sonnet 5.5 and similar models have become faster and cheaper, but fluency has always outpaced factual reliability, and that gap has not closed.\n\nResearchers studying AI writing tells have found that models like Opus 5.5 use statistically anomalous word choices that trained readers can spot — but most people are not trained readers of AI output, and a politician's audience during a cable interview certainly isn't being asked to run a statistical analysis.\n\n## The practical problem for anyone citing AI output\n\nFor AI-art creators, the lesson is narrower but real: the same models used to generate image prompts, write character backstories, or draft creative briefs will, if asked, produce confident-sounding claims about contested facts. That is not a bug being patched — it is structural. The models are optimized for coherence, not accuracy.\n\nAnyone using AI-generated text in a public-facing context — captions, artist statements, character lore — carries the editorial responsibility of verifying what the model produced. The chatbot is a drafting tool. It is not a witness.\n\nCaldwell's gambit is unlikely to survive scrutiny; investigators and journalists can read source documents that no chatbot has seen. But the episode marks a new threshold: AI-generated text being formally deployed as exculpatory evidence, in public, by a public official. Expect that threshold to be crossed again.