Two District Attorneys, One Rule: Verify Before You Trust the AI

Two district attorneys in two different states reached the same rule about artificial intelligence on their own, without comparing notes: treat whatever the tool produces as a lead, never as a finding, until a person checks it against the original record.

KSL reporter Eric Cabrera described both offices on October 23, 2026. Sim Gill, the district attorney for Salt Lake County, Utah, said his office uses AI tools to sort large volumes of digital evidence, including several body-camera recordings from one incident, and to organize legal research. He was specific about where the technology's role ends: "Certainly the Salt Lake County (District Attorney's Office) is not surrendering its judgment to technology." When an AI tool narrows a stack of evidence down to what looks relevant, Gill said he still reviews it himself: "So now I'm going to go put eyes on that myself." He also drew a line for police officers who draft reports with AI assistance: disclose it, and stand behind it in court. "I'm not going to put AI technology to testify in court," he said. "I'm going to put that officer."

Mike Holley, the district attorney for Montgomery County, Texas, gave the reason the check matters, in plain terms: "AI does hallucinate. It does consistently, and it does it well. It can lie and lie well." Holley described a case where his office used an AI tool to search subpoenaed Snapchat records for conversations in a child sex abuse case, work the tool finished in minutes instead of the "hours, maybe days" a person would have needed. His office did not hand those flagged messages to a judge on the tool's word. Prosecutors checked each one against the original Snapchat records first, during the punishment phase of the case. Holley was blunt about the technology in general: "If I could un-invent AI, I would. But it's here, and so we're going to navigate it." He ruled AI out of a narrower set of decisions too: whether to dismiss a charge, offer a plea deal, or divert a case from prosecution. The same offense, he said, comparing a teenager who steals a tie for a job interview with someone who steals one for the thrill of it, can call for very different outcomes depending on the circumstances. "I would not ever want AI to make those judgment calls," he said.

Neither office named the specific AI product it uses. What both named, independently, was the habit that makes the tool survivable in a courtroom: a human checks the output against the source before anyone relies on it.

That habit is the same one at the center of how a research process that uses AI has to work anywhere the stakes are real, not only in a courtroom. An AI system can summarize a mountain of material in minutes, find the needle in the haystack a person would take days to find, and still get a specific detail wrong with total confidence. The two capabilities are not in tension. They are both true of the same tool at the same time, which is exactly why Holley's office checked its Snapchat leads against the original records rather than taking the flagged messages at face value.

Foragentis's intelligence product, ForIntel, designs around that same risk from the first step. Every factual claim in a ForIntel report traces back to a source a reader can click and check. Before any research runs, the ForIntel team registers the acceptance criteria a finding must meet, so that standard stays fixed in advance instead of bending to fit whatever the AI returns. Every report also carries a published section naming what it could not confirm, rather than presenting an unverified lead as a settled fact. A person sits at three separate points in that process, reviewing what the AI produced before it becomes a line in a report a client reads.

Gill and Holley reached a version of that same rule from the opposite direction, inside two prosecutors' offices with no reason to think about market research or AI product design. A hallucination that reaches a courtroom costs someone their case, or their freedom. A hallucination that reaches a business report costs someone a bad decision made with confidence it did not earn. The size of the stakes differs. The fix two prosecutors arrived at on their own is the same fix any report drawing on AI needs: check it against the record before you trust it.