We just watched a disgraced politician use ChatGPT as his character witness, and it's the perfect stress test for how regular people will misuse AI in high-stakes situations.
The Summary
- Former New Jersey Lt. Governor Dale Caldwell resigned after a sexual harassment investigation, then told PBS that he ran the report through "multiple AI platforms" 59 times and they all said he was innocent
- This isn't about AI accuracy; it's about AI authority bias hitting public discourse in the dumbest possible way
- When people treat LLMs as neutral arbiters instead of pattern-matching text generators, we get a preview of every upcoming courtroom, HR department, and family argument
The Signal
Caldwell's defense strategy is both absurd and predictable. He told NJ PBS host Rob Nelson that he "AIed" the ethics investigation report through multiple platforms, asking each one 59 times what its findings would be. The implication: if the AI disagrees with human investigators, maybe the humans got it wrong.
This is the exact failure mode researchers have been warning about. Not hallucinations. Not bias in training data. The real risk is people outsourcing judgment to systems that have no judgment, just statistical patterns from text scraped off the internet.
"When a disgraced politician uses AI as a character witness, he's not being stupid. He's being early."
What makes this story relevant beyond political theater:
- AI authority bias is now showing up in real-world accountability moments
- The average person has no mental model for what an LLM actually does versus what it appears to do
- "I asked AI and it agreed with me" is about to become the new "I did my own research"
Caldwell's gambit reveals something important about how AI enters culture. Most people don't interact with these tools as productivity enhancers or creative aids. They use them as rubber stamps for pre-existing beliefs. You can run the same prompt 59 times, tweaking until you get the answer you want, then cite "multiple AI platforms" as if you'd consulted a panel of experts.
This isn't a bug in the technology. It's a feature of human psychology meeting probabilistic text generation. LLMs are designed to be helpful, to give you an answer that sounds authoritative. They have no stake in the truth. They have no access to the investigation's actual evidence. They're just really good at sounding like they know things.
The business implication: any company building AI tools for high-stakes decisions (legal, medical, financial, HR) is walking into a minefield of misuse they can't prompt-engineer their way out of. The technology works exactly as designed. The humans are the problem.
The Implication
Watch for this pattern to spread. Not in ways that make headlines, but in performance reviews, divorce proceedings, academic appeals, and every other situation where someone wants a second opinion that confirms what they already believe. The companies shipping these tools can add disclaimers and guardrails, but they can't stop someone from asking the same question 59 times until the answer changes.
The real test for AI literacy isn't whether people can spot deepfakes or avoid phishing scams. It's whether they understand that an LLM's confident-sounding response to "did I do something wrong?" is not actually an answer to that question. It's just the most statistically likely next tokens.