ORLANDO — Artificial intelligence may help officers write reports faster, but when those reports are challenged in court, the technology will not be sitting in the witness box. The officer will.
“The software does not testify. Your officer is the one who testifies,” Eric Atstupenas, general counsel to the Massachusetts Chiefs of Police Association, told police executives attending the 2026 International Association of Chiefs of Police conference in Orlando.
That distinction was at the heart of his presentation, “Smarter Tools, Harder Questions: AI and the Future of Police Report Writing,” which examined the legal, operational and leadership questions agencies should confront before introducing AI into one of policing’s most consequential administrative functions.
Atstupenas was careful to explain that he was neither advocating for nor against the technology. AI-assisted report writing could offer legitimate benefits, particularly for officers facing growing administrative demands. But those benefits must be weighed against questions about accuracy, discovery, officer credibility, data security and agency accountability.
His central message to chiefs was that adopting AI is not simply a technology purchase. It is a command decision with consequences that could extend years beyond the original report.
When the report becomes a credibility problem
Consider an officer testifying about an incident documented in an AI-assisted report. The defense attorney asks who wrote the narrative, what information the software added and why portions of the report differ from the body-worn camera recording. The officer may have reviewed and signed the final document. But can the officer explain how every statement got there?
Atstupenas presented three questions that agencies should be prepared to answer:
- Who wrote the report?
- Was it reviewed?
- And can the agency prove how it was created?
The distinction is particularly important in use-of-force cases, where the officer’s perceptions and the totality of the circumstances are central to evaluating the encounter.
“The language model was not there, only the officer was,” he said.
AI may summarize recorded audio or organize notes, but it cannot independently establish what an officer saw, understood or perceived during an encounter. Nor should it introduce explanations, observations or conclusions that the officer cannot personally support.
Atstupenas pointed to federal litigation involving immigration enforcement operations in Chicago, where U.S. District Judge Sara Ellis questioned the credibility of federal agents after evidence emerged that at least one agent had used ChatGPT in preparing a use-of-force report. The judge also identified discrepancies between official accounts and video evidence.
The case illustrates the potential scrutiny of AI-assisted narratives, although it did not establish a binding rule governing police report writing. The Seventh Circuit subsequently vacated the preliminary injunction in the litigation after the underlying controversy became moot.
Atstupenas also raised a less obvious risk: the tendency for officers to become increasingly trusting of technology that appears to work well.
An officer may carefully review the first few AI-generated reports, checking every detail against notes and recordings. But after dozens of seemingly accurate results, that scrutiny may begin to fade. Eventually, the officer may accept a report without reading it closely enough to catch an invented fact or an inaccurate description.
What begins as an honest mistake can become a far more serious credibility issue when an officer signs a statement without adequately reviewing it.
The consequences may extend beyond a single case. An officer whose credibility is compromised could face disclosure obligations or prosecutorial decisions that limit the officer’s ability to testify, potentially jeopardizing a law enforcement career.
For chiefs, the responsibility is therefore twofold: protect the integrity of the agency’s reports and ensure officers understand the personal consequences of adopting language they cannot defend.
Prosecutors may set the rules before courts do
Agencies waiting for appellate courts to establish definitive rules for AI-assisted police reports may find that prosecutors have already made decisions for them.
“The law, I anticipate, is not going to catch up before your prosecutors do,” Atstupenas said.
In September 2024, the King County Prosecuting Attorney’s Office in Washington informed law enforcement agencies that it would not accept police report narratives produced with AI assistance. The office cited concerns about fabricated details and errors that officers might overlook during review.
Connecticut prosecutors have also taken a restrictive approach. In 2026, Chief State’s Attorney Patrick Griffin confirmed that the state’s prosecutors had imposed a moratorium on AI programs used to draft, author or narrate criminal reports. The concern was not opposition to technology itself, but the potential consequences of inaccurate interpretations of body-worn camera recordings.
Meanwhile, states are beginning to establish statutory requirements.
California Penal Code Section 13663, effective January 1, 2026, requires agencies to establish policies for AI-generated police reports. Covered reports must identify the AI program used, disclose AI involvement and include an officer’s signed verification. The law also requires retention of the first AI-generated draft and an audit trail for the same period as the official report.
Utah Code Section 53-25-902 requires law enforcement agencies to maintain policies governing employee use of generative AI. Reports or other law enforcement records created wholly or partially using the technology must disclose that involvement and include certification that the author reviewed the record for accuracy.
These are not identical approaches, and requirements will vary by jurisdiction. But they demonstrate that disclosure, officer verification and agency accountability are already becoming formal expectations.
The practical implication for chiefs is immediate: An agency could invest in a report-writing platform only to discover that its prosecutor will not accept reports produced through that system.
Atstupenas urged chiefs to meet with prosecutors before committing to the technology, rather than waiting until an officer is testifying to reveal that AI helped prepare the report.
The objective is not simply compliance with today’s rules. It is ensuring that officers, prosecutors and courts understand how the technology was used.
The hidden workload behind faster reports
AI report-writing systems are frequently promoted as a way to reduce paperwork and return officers to patrol. Atstupenas did not dismiss that possibility, but he challenged agencies to measure whether the savings actually occur.
Research published in the Journal of Experimental Criminology in October 2024 provides a cautionary example. In a randomized controlled trial involving 85 officers and 755 police reports, researchers found that AI assistance did not produce a statistically significant reduction in report-writing time. Additional analyses covering more than 6,000 reports supported the finding.
The study did not establish that AI-assisted report writing can never save time. It demonstrated that the anticipated efficiency gains should not be assumed without measurement.
“Test it. Try it. Use it, see if it actually saves time,” Atstupenas said.
Reviewing, correcting and verifying AI-generated material can offset time saved during drafting. And report creation is only one part of the workflow. An AI-assisted report may generate drafts, prompts, audit logs and other records that could become relevant in criminal discovery, civil litigation or public records requests. Material differences between an initial draft and a final report could be particularly consequential if they contain information favorable to a defendant.
That raises practical questions agencies need to answer before deployment: Can the system preserve original drafts? Can the agency retrieve prompts and audit logs? Can the vendor comply with litigation holds? Can records be exported in a usable format years after the report was written?
California’s requirement to preserve the first AI-generated draft demonstrates that these are no longer hypothetical concerns.
A system that reduces an officer’s report-writing time but creates substantial new work for prosecutors, records personnel and legal advisors may deliver a different return on investment than the agency anticipated.
Atstupenas’ recommendation was to test the complete workflow rather than rely solely on vendor demonstrations or officers’ perceptions of time savings.
AI is not an IT decision
Although the presentation covered significant legal ground, Atstupenas reserved some of his strongest language for police executives who delegate AI procurement and governance to their technology departments.
“It is not an IT issue. It is a you issue,” he told the audience.
His concern was that chiefs may surrender control over consequential operational decisions because they are uncomfortable with the technology or uncertain what questions to ask vendors.
IT personnel have an essential role in assessing security, integration and technical capabilities. But those considerations are only part of the decision. Legal advisors must evaluate contracts and disclosure obligations. Prosecutors need to address evidentiary concerns. Records personnel must determine whether the agency can retain and produce required information. Patrol officers should help assess whether the product actually improves their work.
Ultimately, however, the chief remains accountable for the decision. “No vendor, committee, or lawyer can take your seat,” Atstupenas said.
He encouraged executives to challenge vendors about data ownership, subcontractor access, retention settings, system prompts, audit capabilities and automatic software updates. “If we want responsible software, we have to ask for responsible vendors,” he said.
One concern is whether software settings actually reflect agency policy. Atstupenas illustrated the risk with an example involving automated license plate reader technology, in which data-sharing settings allowed access beyond what agency leaders had intended.
The lesson extends to AI: A department may have a policy restricting how information is used or shared, but default settings, automatic updates or new features could undermine those restrictions without command staff realizing it.
Chiefs should understand what settings are enabled, who can access agency data and whether vendors can introduce new capabilities without advance approval.
Another important consideration is what happens when the agency wants to end the relationship. Unlike a conventional equipment purchase, an AI contract may leave the agency dependent on a vendor to retrieve records, preserve evidence or remove sensitive information from its systems. Atstupenas compared the relationship to a marriage, with termination requiring careful consideration of how the parties will separate their assets and obligations. Agencies should therefore negotiate data retrieval, deletion, preservation and exit provisions before signing.
“If it’s not in the contract, it’s not a promise.”
A pilot is still a deployment — and policy is only the beginning
The same scrutiny should apply when agencies are merely testing AI report-writing technology.
Atstupenas warned that pilot programs can become a way to introduce systems without the review, policies and training that would accompany a permanent deployment. But if officers are using the technology to create reports for actual incidents, those reports carry the same legal and evidentiary consequences regardless of the program’s experimental status.
“Every pilot report is a report,” he said.
That means agencies should establish approved uses, retention requirements, supervisory review and officer training before testing systems on live cases.
He recommended limiting the scope of pilots, defining their duration and identifying how the agency will evaluate results. Agencies should also consider whether certain incident types, particularly arrests and use-of-force cases, warrant additional restrictions or review.
But even a carefully written policy accomplishes little if nobody checks whether the technology is being used as intended.
Atstupenas emphasized three essential elements of responsible implementation: policy, supervision and training. A policy that sits on a shelf without meaningful oversight is no protection against misuse or unintended consequences.
During the first 90 days, supervisors should review sample reports against available source material, including body-worn camera recordings where appropriate, to identify discrepancies and evaluate officer verification practices.
The agency should then revisit its technology, policies, contracts and settings regularly as the software and legal environment change.
Most importantly, someone must be specifically responsible for that oversight. Atstupenas warned against assigning accountability vaguely to a department or committee without identifying who will actually perform the work.
The AI officers may already be using
Even agencies that have not purchased AI report-writing software may already have an AI governance problem.
Atstupenas described the growing risk of “shadow AI,” in which officers independently use consumer chatbots such as ChatGPT, Claude or Gemini to assist with reports or other police work.
Unlike systems specifically evaluated and authorized by an agency, these tools may introduce uncertainty about data handling, security, records retention and the origin of report language.
Entering sensitive case information into an unauthorized platform can create additional risks involving criminal justice information and agency security requirements.
But Atstupenas cautioned that simply prohibiting the technology is not necessarily an effective long-term strategy. “If you don’t give your officers tools to succeed, they’re going to find tools themselves.”
Officers facing substantial paperwork demands may naturally look for ways to work more efficiently. Chiefs need to understand what employees are already using, establish clear boundaries and evaluate whether approved alternatives can meet legitimate operational needs. The starting point is an inventory of existing AI use, followed by deliberate decisions about what the agency will authorize and how those systems will be governed.
For chiefs, the immediate priorities are straightforward: determine what AI tools officers are already using, meet with prosecutors, establish policy before pilots or purchases, preserve required records and prohibit sensitive case information from being entered into unauthorized public chatbots.
Those steps do not require an agency to embrace or reject AI. They require leadership to make a deliberate decision and create a record demonstrating how that decision was made.
The technology will continue to improve, and the legal framework governing its use will evolve. But the fundamental responsibilities of officers and police executives will remain.
As Atstupenas reminded the audience: “Intent does not survive review, only the record does.”
