Ex-NJ lieutenant governor claims AI reviews of harassment report prove his innocence

New Jersey's lieutenant governor Dale Caldwell resigned on September 25 following an investigation that found he had sexually harassed a staffer and repeatedly violated ethics rules, according to The Verge. Since leaving office he has been giving interviews arguing that he was unfairly targeted. During one appearance on NJ PBS, he reached for a distinctly modern kind of character witness: generative AI.

The claim

As The Verge reports, Caldwell told host Rob Nelson that he ran the investigative report through several AI systems, saying he had "AIed" it "from multiple AI platforms." "I put it through AI," he said, and then asked each system how it would rule. In his telling, "there was no instance in any of the AIs that said there would be a finding of sexual harassment." He also referenced a figure of 59 — "it said 59 times," he told Nelson — though that sentence is transcribed by The Verge with a [sic], and the precise meaning of the number is hard to pin down.

Why chatbots are not investigators

The Verge's weekend editor, Terrence O'Brien, points to the obvious flaw in this defense: chatbots are known for sycophancy, often producing answers that match what the user seems to want. As The Verge observes, it is highly unlikely that Caldwell was hoping the models would validate the investigation's conclusions.

The deeper problem is mechanical rather than psychological. A language model has no access to the record behind a report — no witness interviews, no credibility judgments, no applicable standards of proof. It sees only the text pasted into the prompt and the way the question is framed, and it generates a fluent answer shaped by both. Ask "what would your findings be?" and you get a summary refracted through your own framing, not an independent adjudication. Because output changes with phrasing, someone shopping for a particular conclusion can keep re-asking, or hop between platforms, until an agreeable "verdict" shows up, and then present it as neutral machine judgment.

Why it matters

A statewide officeholder citing chatbot output as exculpatory evidence on public television is a small but telling marker of how quickly AI answers are being treated as authority. It collapses the difference between two very different things: a formal investigation, with process, testimony and accountability, and a generated summary that will often tell the user what they want to hear. As chatbots become everyday tools, "the AI agreed with me" is likely to surface more often in ethics disputes, legal fights and news cycles. The Verge's story is a useful reminder for journalists and audiences alike: a fluent machine answer is not a fact-check, especially when the person quoting it has every incentive to prefer the result.