Detective Lauren Cunningham has seen the ways artificial intelligence is making the world worse off: harmful hallucinations, dangerous commands, manipulative misinformation.
So when she was asked to test out a different kind of AI tool at the Oklahoma City Police Department, Cunningham was skeptical. Longeye, its founder said, was built for investigators like her — firewalled from the public, federally compliant and entirely reliant on her original detective work.
Soon, the tool was helping Cunningham and her colleagues tackle crimes. The 20 or so hours per week she would normally spend monitoring jail calls from murder suspects had been reduced to less than five. A detective was able to sift through thousands of pages of financial documents and identify patterns in a fraction of the usual time.
And a sex crimes investigator used Longeye to translate 10 suspect phone calls into English — finding a confession that turned a child rape case headed for trial into a likely plea agreement.
“It is fascinating to know I get to exist in a time where AI is helpful,” Cunningham said. “Everywhere you turn, AI is nothing but confusing and deceitful, and it tricks people into things that are really detrimental to them. … I didn’t know I was ever going to work in a place where AI understood that its sole job was to be a fact finder.”
The Oklahoma City Police Department is one of 35 law enforcement agencies across the country in the early stages of adopting Longeye, which its San Francisco-based creator markets as an ethical, uncompromising way for all parts of the criminal legal system — police, prosecutors, defense attorneys, corrections officers — to fast-track the pursuit of justice.
The tool exists in an ecosystem flooded with AI tech marketed to law enforcement: license plate readers, facial recognition software, ballistics analysis, crime report drafting, predictive policing. Many of those tools have been met with intense scrutiny from data privacy and police reform advocates, who argue that generative AI is prone to “hallucinate,” or draw faulty conclusions — mistakes that could weaponize a justice system that already has immense power to strip away a person’s liberty.
CEO and founder Guillaume Delépine said he built Longeye to avoid the constitutionally suspect aspects of “quick and dirty” AI platforms.
The tool operates in a “closed sandbox,” meaning it is entirely walled off from outside information that could compromise the analysis or provide faulty feedback — common criticisms of public-facing AI chatbots like ChatGPT and Claude. Longeye is largely designed to analyze documents, data, audio and video obtained by police through a warrant, unlike some facial recognition tools that have proved unreliable. It does not currently write police reports, since generative AI is known to introduce errors into official narratives.
“Dirty doesn’t work in the justice system,” Delépine said. “You have to build a much more deeply thinking machine.”
AI-powered law enforcement faces major obstacles. Defense lawyers routinely challenge it as unreliable. Jurors are often skeptical. And crucially, judges have not yet settled on clear rules for when prosecutors must disclose that AI was used to obtain a piece of evidence or reach a particular conclusion.
“Justice isn’t always served by the latest piece of technology,” said Aramis Donell Ayala, a former Florida state’s attorney and executive director of Fair and Just Prosecution, a prosecutors’ group that focuses on equity. “We aren’t just here to win. We are here to provide justice, and justice is rooted in truth. And truth is rooted in accuracy.”
Delépine contends that Longeye is doing just that. So far in 2026, he said, the program has streamlined approximately 34 years’ worth of detective work into just a few months, processing 25 million files across 35 law enforcement agencies at the local, state and federal level.
While several of those agencies touted success in interviews with The Washington Post, most cases involving the use of Longeye have yet to be adjudicated, meaning its true usefulness remains untested in a court of law.
While Delépine said Longeye was built with all parts of the system in mind, the focus for now is on helping police investigators. Contrary to movie scenes featuring police officers racing after suspected killers or cajoling tearful witnesses, these investigators are often buried in digital evidence, document dumps and overwhelming caseloads.
Each case file entered into Longeye is self-contained, pulling only from the information uploaded by police, to avoid contamination by outside inputs. That information can be turned into timelines, maps or spreadsheets and help identify key moments from hundreds of hours of witness interviews or phone calls.
Any analysis done by Longeye generates a citation and link to the original source. Longeye complies with the FBI’s data security and information privacy protocols and creates an audit trail, preserving a clear chain of custody that is critical for eventual use in court.
Unlike other AI tools designed for the justice system, Delépine said he hopes Longeye can be used by public defenders and nonprofits that work to exonerate people. The company offers a discount to public defender’s offices, which have smaller budgets and fewer resources than prosecutors or police.
Marc Caudel, a private criminal defense investigator who has worked with innocence organizations and federal public defenders for decades, said he believes AI, by taking on labor-intensive tasks, can help level the field between defense and prosecution. He has begun testing various AI tools, including Longeye, comparing it to employing an intern.
“It’s supposed to be that the scales of justice are balanced, but it’s really not,” Caudel said. “I think that AI can really help cut through a lot of the mundane parts of having to go through discovery and find the little nuggets you need to find.”
The integration of AI technologies into American policing is forcing political leaders and lawmakers to weigh the ways it could bolster — or upend — a case once it hits the courtroom. The criminal legal system, built on constitutional protections and human testimony, has evolved with the introduction of new technologies: fingerprinting, DNA evidence, Breathalyzers, wiretapping and social media surveillance.
There is much the courts haven’t addressed when it comes to the use of artificial intelligence, including prosecutors’ disclosure obligations under the 1963 Supreme Court case Brady v. Maryland and a defendant’s rights under the Sixth Amendment to confront their accuser.
When police use technology to process evidence or draw incriminating conclusions, a human being must come to court to explain how that technology works. The same should apply to AI, Ayala said: Lawyers cannot cross-examine AI, so a human detective or investigator must be able to testify about its findings. If they cannot independently verify the leads or patterns that AI has spit out for use in court, she said, they risk perjuring themselves.
“Once you start thinking of all the ways new technology begins to create litigation, we have to recognize technology is moving faster than litigation can be heard,” Ayala said. “We have to make way for that.”
The Policing Project at the New York University School of Law has published model legislative language for state and local lawmakers, who are best positioned to create guardrails around AI usage by police. The model language, which mirrors recommendations from the Electronic Frontier Foundation, requires police agencies to keep a public-facing inventory of the AI tools they are using; disclose in reports whether AI was used and how; and face civil action if they fail to follow disclosure regulations.
But few states have adopted such policies so far, advocates said.
In Utah and California, state law requires police departments to disclose when generative AI was used to help write police reports. The push for the laws arose from issues related to Axon’s Draft One software, an AI tool that allows police officers to turn audio from body cameras into a written police report.
Critics cite problems with Draft One, including inserted bias and inaccurate summarizing of the scene, both of which could undermine the integrity of a sworn affidavit. Axon says its product ensures human control of the process.
“Officers remain the author, responsible for reviewing, editing, and approving every word before anything is submitted,” Axon spokesperson Rasleen Krupp said. “That human step is foundational to how the product works.”
An investigative tool like Longeye would likely not be covered under existing state laws, said Kate Chatfield, executive director of the California Public Defenders Association, since it currently does not write reports.
In Washington’s King County — a tech hub where residents, including police, are unafraid of embracing new technologies — the district attorney’s office has banned local law enforcement agencies from using AI-generated police reports, concerned they could contain damaging errors.
“In the shadow of all these tech companies, we may be more aware of their pitfalls than others,” said Daniel J. Clark, the county’s chief deputy prosecutor. “Prosecutors are skeptical by nature. We need to make sure we can establish something beyond a reasonable doubt in court.”
That same tension is playing out in Oklahoma, where state lawmakers are considering several AI-related bills, including one that would require law enforcement agencies to disclose their AI usage.
Cunningham, the Oklahoma police detective, is already subject to a city policy requiring officers to disclose if they used generative AI to write a police report. A tool like Longeye, she said, would be considered “work product,” much like other organizational tools, and likely would not have to be disclosed.
She recently gave a presentation on Longeye to the State Bureau of Investigation, she said, touting the ways the smart use of AI tools could help investigators speed up the often-slow wheels of justice. The Oklahoma Department of Corrections is testing the tool, as well, said the prison system’s chief of operations, Jason Sparks.
“Law enforcement — corrections — needs a tool that is self-contained,” Sparks said. “The trash of the internet doesn’t need to be sucked in. It needs clean data.”
Correctional staffers are using the tool to sift through the nearly 50 million minutes of prison phone calls they monitor annually for potential illegal activity, Sparks said, and the data from the nearly 6,700 contraband cellphones they confiscate. One cellphone download analysis led investigators to an assault that was being planned on someone in custody by people on the outside — and Sparks said officers then prevented it.
“That would have gone unnoticed without this product for sure,” he said.
Whether people are supportive or skeptical of policing in America, Delépine said, nobody wants the legal process to be less efficient.
“When the justice system is there for you,” Delépine said, “it changes how you think for the rest of your life.”
The post AI could vastly streamline policing. Skeptics urge caution. appeared first on Washington Post.
