Share with your CISO
Dale Caldwell, New Jersey’s former lieutenant governor forced out on September 25th after an investigation confirmed sexual harassment and repeated ethics violations, is now claiming vindication via AI. His defense, offered live on NJ PBS, is that he ran the investigation report through “multiple AI platforms” and none of them returned a finding of sexual harassment across 59 queries. This AI-as-character-witness gambit is the story, and it has direct implications for how organizations govern AI outputs in high-stakes contexts.
What this means for your business
The Caldwell defense is a preview of a coming pattern. Asking a large language model, which is a text-prediction system trained to produce plausible-sounding responses, to evaluate an investigative report and then treating the output as exculpatory evidence is a category error. But the fact that a sophisticated public figure tried it on television tells you something important: your employees, your legal counterparts, and eventually your adversaries in litigation or regulatory proceedings will try the same move. The question isn’t whether this is credible. It isn’t. The question is whether your organization has a documented policy that preempts it.
What Caldwell actually demonstrated is that AI models queried without structured prompts, adversarial testing, or grounded source data will produce outputs that feel authoritative while meaning almost nothing. Fifty-nine queries returning no finding of sexual harassment says nothing about the underlying facts. It says something about how the model was prompted, what context it was given, and what answer the question was shaped to produce. This is prompt-shopping, the practice of iterating question framing until the model returns a favorable answer, and it is trivially easy to do with any general-purpose LLM.
Organizations that have deployed AI in HR investigations, compliance workflows, or legal discovery are now carrying a specific liability they may not have named yet. If a general-purpose chatbot can be used to manufacture apparent exoneration, the credibility of any AI-assisted finding, yours included, becomes contestable in the same register. The CISOs and general counsels who move first to distinguish between auditable, purpose-built AI systems and off-the-shelf query tools will be in a stronger position the first time an employee or plaintiff’s attorney runs this play against them. I’d revise that assessment only if a court explicitly rules that LLM outputs are inadmissible in employment disputes, which would close the attack surface rather than requiring organizations to defend against it.
Concept deep-dive: Prompt-shopping
Prompt-shopping is the practice of reframing a question to a language model repeatedly until it returns a preferred answer, exploiting the fact that these models are sensitive to phrasing, context, and assumed intent. It’s the AI equivalent of polling a witness until they say what you need. The business risk is that it makes AI outputs look like independent verification when they’re actually a reflection of how the question was constructed, which is exactly why output provenance and audit trails matter in any governed AI deployment.
Based on reporting from NJ’s former Lt Gov is using AI to say he’s innocent of sexual harassment, originally published 2026-10-04 12:16:00.
