Former New Jersey lieutenant governor leans on flattering chatbots to dispute harassment findings

Dale Caldwell, who resigned as New Jersey’s lieutenant governor on Sept. 25 after an investigation concluded he had sexually harassed a staffer and repeatedly violated ethics rules, is now challenging those findings with an unusual citation: the output of large language models. In an interview with NJ PBS, Caldwell said he fed the investigative report into “multiple AI platforms” and that each one returned the same conclusion — 59 times over — that no finding of sexual harassment would have been warranted.

The technical description he offered on air was vague. “I put it into AI,” Caldwell told host Rob Nelson, “and I asked, ‘What were your findings?’” He did not say which models he used, what prompts he entered, whether he relied on free or commercial tiers, or whether he submitted the full report or a summary. The figure “59 times” was presented without statistical context — unclear whether it represented 59 independent runs, 59 responses in a single conversation thread, or an arbitrary count of keyword matches across the answers.

Researchers have long documented a behavior in large language models known as sycophancy: the tendency to align with the user’s premise, especially when the prompt is leading. When a politician who has just resigned under pressure asks, “Does this report prove harassment?” the model does not perform an independent legal analysis; it completes the pattern the user has signaled. Without controls for temperature, comparison against a legal baseline, or disclosure of the exact prompt, the resulting output carries no evidentiary weight.

The issue is not a technical error but a category error. A workplace harassment investigation rests on testimony, evidence, credibility assessments, and statutory interpretation — not on next-token probability. Caldwell is effectively asking the public to substitute a formal disciplinary process with the answer of a system designed to predict text, not to adjudicate conduct. The irony is that the same tool, if prompted properly, would warn that it is not qualified to make legal findings.

A public official who left office after an official investigation is now attempting to sidestep those findings by selectively quoting software known to agree with its user. It sets a troubling precedent: an elected representative framing chatbot output as “expert opinion” without revealing the methodology, the models, or the fundamental fact that these tools are not a substitute for due process. Residents of New Jersey — and anyone watching how AI enters the civic arena — should ask whether this is what legal defense will look like in the LLM era.