NJ Lt. Governor Uses AI to Deny Harassment Claim

NJ Lt. Governor Dale Caldwell Cites AI to Deny Sexual Harassment Findings

On September 25, 2026, Dale Caldwell was removed from his position as New Jersey's lieutenant governor after a formal investigation concluded he had sexually harassed a staffer and repeatedly violated state ethics rules. The findings were the result of an official process — not a tweet, not a rumor, not a political opponent's accusation. Yet in the weeks that followed, Caldwell embarked on a media tour to challenge the record. During one such appearance on PBS, he offered a defense that stopped observers cold: he had consulted an artificial intelligence system, and it told him he had not committed sexual harassment.

The Dale Caldwell AI sexual harassment episode has since drawn attention far beyond New Jersey's borders, becoming a case study in the growing and genuinely alarming trend of public figures treating AI-generated output as a form of exculpatory evidence. The episode raises sharp questions about technological literacy, institutional accountability, and what happens when a powerful person conflates a chatbot's response with an investigative finding.

Caldwell was forced to step down following what official accounts describe as a finding that he had both harassed a member of his staff and engaged in repeated violations of ethics standards governing state officials. Those are serious, distinct categories of misconduct. Sexual harassment findings and ethics violations in New Jersey state government do not emerge from casual review. They require documentation, witness accounts, and formal adjudication through recognized institutional channels. That background matters enormously for understanding why Caldwell's AI defense is not simply unusual — it is a category error.

Why Using AI as a Character Witness Is Problematic

Generative AI systems — the large language models that power chatbots like ChatGPT, Gemini, and Claude — are trained on vast corpora of text from the internet, books, and other written materials. They are designed to produce fluent, contextually plausible responses. They are not designed to adjudicate facts about specific individuals, and they have no access to confidential investigation files, witness testimony, or non-public records.

Research from Stanford's Human-Centered Artificial Intelligence institute and work conducted by MIT researchers have documented what the field calls "hallucination" — the tendency of large language models to generate confident-sounding statements that are factually incorrect or entirely fabricated. Studies have found that LLMs produce inaccurate claims about real people and real events with notable frequency, particularly when queried about individuals whose documented histories are complex, contested, or not extensively represented in training data. One analysis examining AI output about named public figures found error rates high enough to disqualify LLM responses from use in any context where factual accuracy is consequential.

The reason is structural. A language model does not "know" whether Dale Caldwell committed sexual harassment in the same way a court record knows. It generates a response based on statistical patterns in text it has seen. If the training data available about a person contains more exculpatory framing than accusatory framing — or simply contains very little — the model will generate output that reflects that distribution, regardless of what actually occurred. An AI saying a person did not commit misconduct is not a finding. It is a probabilistic pattern-match dressed in declarative grammar.

There is also no transparency mechanism. When a New Jersey ethics investigation produces a finding, there is a record: who was interviewed, what documents were reviewed, what standards were applied. When an AI chatbot produces an answer, there is essentially none of that. The model cannot be cross-examined. Its training data cannot be subpoenaed. Its reasoning — to the extent that word even applies — is opaque. Treating these two categories as equivalent undermines the foundations of evidence-based accountability.

A Pattern of Politicians and Public Figures Misusing AI

Caldwell is not the first person in a position of public trust to misapply AI output as though it were authoritative evidence. The pattern has been developing for several years across jurisdictions and institutional settings.

Lawyers in American federal courts have submitted AI-generated legal briefs citing cases that did not exist. A New York attorney was sanctioned by a federal judge in 2023 after ChatGPT fabricated multiple case citations that were submitted in court filings. Politicians have used AI-generated images and audio to challenge the authenticity of real recordings. Others have circulated AI-generated text claiming to summarize findings of investigations — findings that differed from the actual documented conclusions.

The common thread is the exploitation of a gap in public understanding. Many people remain uncertain about what AI can and cannot do. When an official-sounding response emerges from a sophisticated-seeming tool, the psychological weight assigned to that response can be disproportionate to its evidential value. Public figures who know this — or who themselves misunderstand AI's limitations — can exploit the confusion, whether deliberately or not.

What distinguishes Caldwell's case is the brazenness of using this tactic on a public media platform, in direct response to findings from an official institutional investigation. It treats the audience as unlikely to distinguish between the epistemic weight of a state ethics inquiry and the output of a chatbot.

The Broader Implications for AI Credibility in Public Discourse

When AI outputs are deployed to challenge verified institutional findings, the damage runs in two directions simultaneously. First, it degrades the perceived legitimacy of genuine investigative processes. If audiences begin to believe that AI responses carry comparable weight to formal findings, public trust in accountability mechanisms erodes. Second, it damages legitimate AI applications by associating the technology with bad-faith argumentation.

Researchers who study digital literacy and AI governance have been sounding this alarm for years. The concern is not merely academic. As AI systems become more fluent and more accessible, the surface area for this kind of misuse expands. The barrier to generating a plausible-sounding AI denial of any allegation is essentially zero. Anyone with access to a consumer chatbot can query it in ways designed to elicit exculpatory-sounding responses. This is not evidence. It is prompt engineering.

Experts in the field emphasize a foundational point that gets lost in public debate: large language models are not truth machines. They are text-prediction systems. The fluency of the output creates an illusion of authority that the underlying architecture does not support. When a politician presents an AI response as though it settles a question of fact about their own conduct, they are misrepresenting what the technology does — whether or not they understand that themselves.

Ethics, Accountability, and the Limits of Technology

The formal process that resulted in Dale Caldwell's resignation exists precisely because accountability requires more than a person's denial. New Jersey's ethics oversight framework, like similar structures across American state governments, is designed to evaluate evidence independently of the subject's own account. Officials who face misconduct findings are rarely enthusiastic about those findings. The purpose of a formal investigation is to produce a record that stands independent of the subject's preferred narrative.

When Caldwell cited AI on PBS, he did not present new witnesses. He did not produce documents the investigation had overlooked. He did not cite legal authority that the findings misapplied. He cited a text-generation system's response to an unverified prompt. These are not equivalent acts.

The episode should prompt news organizations, public audiences, and institutional actors to establish clear public norms: AI-generated text is not evidence. It cannot corroborate or refute investigative findings. Its deployment as though it could is a form of informational pollution that benefits no one except the person seeking to cloud a documented record.

Technology expands what is possible. It does not alter what is true. Official investigations do not become less valid because a chatbot produces a sentence that contradicts them. Dale Caldwell's resignation followed from findings made through a process with legal standing, institutional accountability, and a documented evidentiary record. The AI response he cited on PBS had none of those things. That distinction is not a technicality. It is the entire point.

Source: The Verge