By This Hour Technology Desk
Dale Caldwell, identified in a report as New Jersey’s former lieutenant governor, has invoked the output of artificial-intelligence platforms while publicly disputing an investigation that, the report said, found that he sexually harassed a staff member and repeatedly broke ethics rules.
Caldwell resigned on September 25, according to the account. In a subsequent NJ PBS interview, he said he had put the investigation report through several AI platforms and asked what their findings would be. He said none produced a sexual-harassment finding. The claim puts a familiar technology in an unusually consequential role: not merely as a tool for summarizing or discussing a document, but as part of a public argument against an investigation’s conclusion.
The available account does not establish which AI systems Caldwell used, the exact material supplied to them, the prompts he entered, or the full responses they generated. Nor does it show whether the systems were asked to replicate an investigation, assess particular passages, summarize allegations, or answer a more general question. Those gaps sharply limit what can be inferred from the assertion that AI did not reach the same conclusion as the investigation.
An AI response is not the investigation it is asked to assess
The distinction between the reported investigation and Caldwell’s description of his AI exercise is central. An investigation reaches findings through whatever process, evidence and rules govern that inquiry. A chatbot or other AI platform responds to the information presented to it and to the question framed by its user. Even where both are presented with a document bearing the same title, they may not be conducting remotely equivalent tasks.
Caldwell’s account, as reported, turns on what he says the platforms would have found after receiving the report. That is not the same as demonstrating that the platforms reviewed the complete record behind it, considered evidence outside the report, tested competing accounts, or applied the ethics rules at issue. Nothing in the available material indicates that they did. It also does not indicate that the tools were designed or instructed to make a formal determination on a workplace-harassment allegation.
The apparent disagreement therefore cannot, on the information available, be treated as a clear conflict between two comparable decision-making processes. It is more narrowly a claim that AI outputs, generated from an undisclosed setup, did not predict the conclusion Caldwell says the investigation reached. The difference may be significant to his public defense, but it does not by itself reveal an error in the investigation or validate the AI responses.
That limitation matters because a polished answer from an AI system can appear authoritative even when the question put to it is incomplete, imprecise or framed toward a preferred result. The source material characterizes AI chatbots as prone to agreeing with users or telling them what they want to hear. In this instance, the reported facts do not allow an outside reader to determine whether such behavior affected the responses, whether the outputs were consistent across systems, or whether Caldwell presented them in full.
Resignation followed reported findings, while Caldwell rejects them
The chronology described in the report is straightforward, though key underlying records are not included in the material available here. The investigation reportedly found sexual harassment of a staffer and repeated ethics violations. Caldwell then resigned on September 25. He has since appeared in the media to challenge the outcome and has said publicly that he is being unfairly targeted.
His reported use of AI belongs to that defense. Rather than describing the platforms as a source of new evidence, Caldwell said he submitted the investigation report itself to multiple systems. He said that after asking what their findings would be, no instance produced a sexual-harassment finding. The assertion appears intended to cast doubt on the report’s conclusion by suggesting another form of review did not arrive at it.
But the available account contains no copies of the report, no transcripts or screenshots of the AI interactions, and no details about the number of platforms beyond Caldwell’s description of using multiple ones. There is also no account of whether all systems received identical text, whether they were queried at the same time, whether follow-up prompts were used, or whether responses that differed from Caldwell’s account were excluded. Without that information, readers cannot evaluate the reliability, consistency or relevance of the claimed exercise.
Equally, the material supplied does not set out the investigation’s methodology or explain the evidence on which its reported findings rested. It does not include the investigator’s reasoning, the terms of the applicable ethics rules, Caldwell’s submissions to the inquiry, or any institutional response to his later AI-based argument. The public dispute is therefore clear in outline but thinly documented in the available source material.
The public case illustrates a high-stakes use of generative tools
AI systems are increasingly capable of producing fluent assessments of material that users paste or upload. Caldwell’s reported argument demonstrates how easily that capability can migrate from routine drafting or analysis into disputes over responsibility and reputation. The issue is not simply whether a system can generate an answer. It is whether the answer is being assigned weight that the process behind it can support.
In a matter involving an investigation’s conclusions, the wording of the request can be decisive. A request to identify information in a report differs from a request to decide what finding should have been made. A system that is asked to forecast an outcome may be responding to a condensed text rather than recreating the underlying inquiry. The available account does not say how Caldwell phrased his questions beyond his description that he asked what the AI systems’ findings would be.
That uncertainty makes it difficult to understand the reported “no instance” result. It could refer to direct answers to a single question, to multiple interactions with several services, or to another approach entirely. The report provides no basis for deciding among those possibilities. It also gives no indication that AI tools had access to anything beyond what Caldwell supplied. A conclusion generated under those conditions cannot be assumed to reflect all information considered in the original investigation.
There is a broader communications consequence as well. A public official or former official can use a system’s confident language to make a defense seem externally validated, even if the user selected the material and controlled the prompt. That does not prove an AI answer is wrong. It does mean that the origin, inputs and task have to be visible before the answer can be weighed alongside a formal finding.
Caldwell’s reported comments do not resolve those questions. They establish his position: he believes the AI outputs support his contention that the sexual-harassment finding should not have been reached. The account also establishes that he disputes the investigation and says he was targeted unfairly. What remains unestablished is whether the outputs were a meaningful test of the investigation at all.
Missing records leave the central claims untested
The case involves two assertions that should not be collapsed into one. The first is the reported conclusion of the investigation, followed by Caldwell’s resignation. The second is Caldwell’s own account of what several AI platforms told him after he provided the report. His stated disagreement with the first does not make the second independently verifiable, and the claimed AI responses do not on their own disprove the first.
For readers trying to assess the dispute, the missing documentation is consequential. The full investigation report would be necessary to understand the stated findings and their basis. A complete record of the AI exchanges would be necessary to assess Caldwell’s account of the platforms’ responses. Information about the systems, prompts, supplied text and any edits or follow-up requests would help establish what the tools were actually asked to do.
None of those materials is contained in the supplied report. There is no indication here that the AI platforms themselves have commented, that their outputs have been independently reviewed, or that the investigation’s authors have addressed Caldwell’s characterization of the exercise. There is likewise no information in the available material on any further proceeding after the resignation.
The most defensible reading is therefore limited. A report says Caldwell resigned after an investigation found sexual harassment and repeated ethics violations. Caldwell denies the fairness of that outcome and says AI platforms, when given the report, did not predict such a finding. The available evidence does not permit a conclusion about the validity of the investigation, the accuracy of Caldwell’s AI account, or the evidentiary value of the systems’ responses.
This report has not been independently corroborated. It is based on a single source-bound account and Caldwell’s reported public statements, and the underlying investigative materials and AI exchanges were not available in the supplied record.
For further context on this subject, see Fed watchdog finds no criminal violations in headquarters renovation overruns.
Reporting notes
What is confirmed: Caldwell reportedly disputes the investigation and says several AI platforms did not predict a sexual-harassment finding.
Why this matters: The dispute raises limits on treating AI responses as validation of, or a substitute for, a formal investigation.
What remains unclear: The investigation’s underlying evidence, the full report, and the AI exchanges have not been provided. This report is based on one source and has not been independently corroborated.
