Don't let AI answer for you

Dale Caldwell resigned as New Jersey's lieutenant governor in September after an investigation documented that he sexually harassed a staffer and violated ethics rules repeatedly. He has spent the weeks since trying to unseat those findings by doing something simpler: feeding the investigation into AI systems and presenting their output as evidence of his innocence. The systems agreed with him. The practical effect is corrosive — not because AI said it. Because AI can sound like it knows what it's talking about while saying almost anything.

An AI system can be prompted to sound exactly like a rigorous investigator while reaching the opposite conclusion. Once that template exists, the next executive accused of misconduct has a ready-made defense strategy. Caldwell did not invent deep dishonesty in reputation management. He just gave it a new suit.

What makes this weaponizable is how plausible it reads. An investigation relies on witness testimony, pattern-finding. Judgment calls about what testimony means — the same basic moves an AI can replicate in reverse. A human investigator concludes pattern X indicates guilt. An AI, fed the same materials and asked to find exoneration, can identify "alternative interpretations" of those same facts with enough specificity to sound like analysis rather than denial. It is the difference between "that's not what happened" and "here's why the evidence doesn't support that conclusion," and only one of them sounds like it came from someone who read the file.

Here's what changes: the next person accused of harassment can now point at Caldwell's precedent and demand their own AI analysis receive equal consideration. Why should one executive's AI-generated counter-narrative be dismissed while another's is entertained? The investigative report becomes just one interpretation among many, equally weighted against a denial that cost nothing to generate and required no accountability. The institutional finding doesn't disappear. It just gets outweighed by plausible-sounding noise.

An AI system can be prompted to sound exactly like a rigorous investigator while reaching the opposite conclusion, and once that template exists, the next executive accused of misconduct has a ready-made defense strategy.

This is how investigation becomes indistinguishable from spin at scale. It is not that AI lied. It is that AI made lying credible enough to demand a hearing. Once that move is available, institutional accountability becomes a game anyone with access to a language model can play. The distinction between documented misconduct and sophisticated denial collapses not because the findings were weak. Because the defense can now sound like it deserves to be called evidence.