Cold cases that sat untouched for decades are getting second looks, because artificial intelligence can do in minutes what investigators could not do in years: read thousands of pages of degraded case files, age a suspect’s photograph across two decades, and build a family tree from a partial DNA profile. The results have been real. The risks are also real. This piece covers both.
The Problem AI Was Built to Solve
Cold cases do not go cold because investigators stop caring. They go cold because the data wins. A single major investigation can generate thousands of pages of handwritten notes, interview transcripts, lab reports, and surveillance logs, spread across multiple agencies, stored in filing systems that were never designed to talk to each other. When a detective retires, that institutional knowledge does not transfer. When a case changes hands, the new investigator often starts from scratch.
That is the gap AI is designed to fill. Not instinct, not interviewing, not the judgment call that comes from thirty years of homicide work. The part no human being can do at scale: reading everything, cross-referencing everything, and surfacing connections that exist in the record but are invisible to exhausted investigators working against time and resource constraints.
In the summer of 2024, the Indianapolis Metropolitan Police Department brought in AI researchers to work a 20-year-old unsolved murder. The AI tool processed 3,000 pages of the investigative file, including handwritten notes, to construct a timeline of events. That volume of material would take a human investigator weeks to read and months to synthesize.
The Indianapolis case is not unique. It is representative of what departments around the country are discovering: the bottleneck was never motivation, it was capacity. AI does not replace the detective. It gives the detective something to work from.
Three Tools Producing Documented Results
Facial Aging and Recognition
When suspects disappear, they age. Photographs do not. For decades, the standard response to a cold case with aging suspects was a sketch artist and a prayer. AI has replaced that with predictive aging software that generates statistically probable current-day likenesses from old photographs, then runs those likenesses against facial databases that include driver’s license records, social media, and law enforcement databases.
The Kerala Police in India used this approach to close a triple murder case that had sat unsolved for 19 years. In 2006, a woman and her 17-day-old twin daughters were murdered in their home. Two suspects, both army personnel, vanished. In 2023, the Technical Intelligence Wing applied AI to age available photographs of the suspects. One generated image produced a 90 percent similarity match with a wedding photograph found on Facebook, revealing a man living in Puducherry under an assumed name. Both suspects were arrested in January 2025, located through the web of connections radiating outward from that single Facebook match.
Genetic Genealogy
Genetic genealogy is the most powerful cold case tool produced in the last decade, and AI is what makes it scalable. The process works like this: forensic scientists extract a DNA profile from crime scene evidence, including severely degraded material that older testing methods could not process. That profile is uploaded to public ancestry databases. Partial familial matches emerge. Investigators build family trees outward from those matches until they arrive at individuals who match the geographic and demographic profile of the original suspect. The combination of DNA science and AI-accelerated tree-building is closing cases that were genuinely considered unsolvable.
In 2024, forensic genetic genealogy identified Paul Hutchinson as the likely killer of Danielle Bertoldi, who was murdered in Michigan in 1997. When investigators contacted him, he exhibited erratic behavior and took his own life within 24 hours. Subsequent DNA testing confirmed the match. Texas-based Othram Labs identified the remains of Charles Howard Wallace, who died in Arkansas in 1977, by analyzing genetic patterns against documented records, resolving a 48-year-old identity case. In the Austin yogurt shop murders of 1991, genetic genealogy produced leads in a case that had frustrated investigators for three decades.
The Golden State Killer case, resolved in 2018, remains the most well-known example of this methodology. Joseph James DeAngelo was identified after investigators uploaded crime scene DNA to a public genealogy database and traced familial matches through a constructed family tree to a man in Sacramento. He was arrested and ultimately tied to 13 murders and more than 50 rapes committed across California in the 1970s and 1980s.
Large-Scale Document Analysis
The third category is less dramatic but arguably the most broadly applicable: AI trained to read degraded documents, recognize handwriting, connect names across thousands of records, and flag inconsistencies within case files. Tools like CrimeOwl, developed by Ash Ghaemi after his mother disappeared from a Wheat Ridge, Colorado motel in 2006, scan thousands of pages of police reports and evidence logs in minutes. The platform pulls exclusively from the case files it is fed, which the developer argues prevents the contamination that comes from training on publicly available true crime content. All 87 sworn officers in Wheat Ridge are now authorized to use AI tools daily in their investigative work.
The Lab is Clutch Justice’s free suite of Michigan court literacy tools: FOIA request generators, judicial report builders, decision trees, and the Michigan Courts Glossary. Built for people who need to understand the system and document what they find.
Explore The Lab ?What the Data Actually Shows
The numbers coexist in uncomfortable tension. The same category of technology, AI-powered image analysis, produced a breakthrough in a 19-year-old Indian murder case and contributed to at least fourteen wrongful arrests in the United States. The difference is not the technology. It is the deployment, the oversight, and whether the output is treated as a lead or as evidence.
The Innocence Project has documented what it calls “evidence laundering”: AI tools marketed as investigative lead generators are being introduced at trial as direct evidence of guilt, often without disclosure to the defense or scrutiny by the court. The pattern echoes the history of bite mark analysis and hair comparison evidence, both of which were presented with false confidence for decades before being discredited.
The Bias Problem Is Not a Bug
Facial recognition technology performs measurably worse on darker skin tones. This is not an edge case or a controversy. Multiple independent studies have documented it. The reason is structural: the training datasets used to build these algorithms underrepresent Black and brown faces, which produces algorithms that are less capable of accurately distinguishing between individuals from those groups. When that degraded performance is deployed in investigations, the risk of misidentification falls disproportionately on communities that already face the sharpest end of law enforcement discretion.
Research has found that disproportionate arrests of Black people by law enforcement agencies using facial recognition may result from the absence of Black faces in algorithm training data, institutional belief in the technology’s infallibility, and officers’ own pre-existing biases amplifying those technical failures. The technology does not produce neutral outputs. It amplifies the biases baked into its construction.
One ACLU client spent six months in jail after police relied on a facial recognition result that incorrectly identified her as a suspect. Detroit has been particularly scrutinized. The city’s former police chief acknowledged that using facial recognition alone as the basis for suspect identification would produce misidentifications the overwhelming majority of the time. The city’s own officers have nonetheless used it to generate lineup candidates, a practice that advocacy groups have challenged in court.
Tips for the Sleuth: What Actually Works, What Causes Harm
Amateur investigators, true crime communities, and family advocates for missing persons are increasingly using AI tools to surface information on cold cases. Some of that work has been genuinely valuable. Some of it has destroyed lives. The difference is almost entirely in methodology.
AI is most reliable when it is analyzing existing records: case files, property documents, court records, and news archives. Feed it structured data from primary sources. Do not ask it to speculate about suspects. AI hallucinations in an investigative context can produce confident-sounding fabrications that contaminate a case record and destroy reputations.
Genetic genealogy identifies a probable candidate within a family lineage. It does not identify a perpetrator. The family tree constructed from a partial DNA match may contain dozens of people who fit the geographic and demographic profile. Investigators and amateur sleuths alike must resist the pull to treat a DNA lead as a conclusion.
AI output is a starting point. It is a direction to look, not a finding. Any investigation, professional or amateur, that treats AI-generated identification as dispositive evidence is doing harm, not justice. The role of a sleuth using these tools is to surface material for trained investigators to evaluate, not to reach verdicts in public forums.
Public ancestry databases used for genetic genealogy contain DNA submitted voluntarily by millions of people who did not consent to have their genetic information used in criminal investigations. Uploading crime scene DNA to those databases without legal authorization raises serious ethical and legal questions that vary by jurisdiction. Platform terms of service are not uniform, and not all platforms permit law enforcement use without a warrant.
AI-generated summaries of case files can pull from inaccurate secondary sources, including decades of press coverage that may contain factual errors. Always trace a claim back to the primary document. If you cannot locate the primary document, the claim should not be repeated as fact.
What Comes Next in Investigation
The near-term trajectory of AI in cold case investigation runs along three tracks. First, multi-modal analysis: systems that can process audio from old interview recordings, digitize handwritten notes through OCR, analyze crime scene photographs, and synthesize all of it into a unified case summary. Second, expanded genetic database access: as commercial DNA databases grow, and as law enforcement agencies continue to negotiate access agreements with platforms like GEDmatch and FamilyTreeDNA, the pool of potential familial matches available to investigators will expand substantially. Third, AI-assisted geographic profiling: the use of case data, travel patterns, and offense locations to generate probability maps of where a suspect was likely based, narrowing investigative geography in a way that previously required years of manual analysis.
None of these advances is paired with a comprehensive federal framework governing how AI evidence may be used in criminal proceedings. Defense attorneys currently have limited discovery rights regarding the AI tools used to generate the leads that produced their client’s arrest. That gap is beginning to produce constitutional challenges, and courts are just starting to articulate standards for what disclosure is required when AI played a role in an investigation.
The accountability question for investigative journalists, legal advocates, and court-watchers is not whether AI can solve cold cases. It can, and it has. The question is whether the oversight infrastructure necessary to prevent AI from manufacturing wrongful convictions will develop at anything close to the pace of the technology itself.
AI in cold case investigation operates exactly as AI operates in every other domain where the legal system has been slow to respond: as a force multiplier for whatever institutional biases already exist. Where investigations are rigorous and oversight is present, the technology produces results. Where investigations are slipshod, under-resourced, or aimed at producing confirmations rather than finding truth, the technology accelerates the damage.
AI is used across several investigative functions: processing thousands of pages of case files, aging suspect photographs for facial comparison, running forensic genetic genealogy to build suspect family trees from degraded DNA, and identifying geographic crime clusters. These tools augment human detective work. They do not replace it.
Reliability varies significantly. Facial recognition performs less accurately for people of color, and AI-generated identifications have contributed to at least fourteen documented wrongful arrests in the United States. The technology functions as a lead generator, not a final identifier.
Genetic genealogy involves extracting a DNA profile from crime scene evidence and uploading it to public ancestry databases to find partial familial matches. Investigators then build family trees from those matches to identify a likely suspect. AI accelerates the matching and tree-building process, allowing investigators to work from severely degraded biological material that older testing methods could not process.
The primary risks are algorithmic bias, hallucination, and evidence laundering. Biased training data produces skewed identifications that disproportionately harm Black and brown suspects. AI hallucinations introduce fabricated details. And when AI output is presented at trial as direct evidence rather than a lead, it bypasses proper evidentiary scrutiny entirely.
Bluebook: Williams, Rita. The Algorithm and the Archive: How AI Is Cracking Cold Cases, Clutch Justice (May 20, 2026), https://clutchjustice.com/2026/05/20/ai-cold-cases-investigation/.
APA 7: Williams, R. (2026, May 20). The algorithm and the archive: How AI is cracking cold cases. Clutch Justice. https://clutchjustice.com/2026/05/20/ai-cold-cases-investigation/
MLA 9: Williams, Rita. “The Algorithm and the Archive: How AI Is Cracking Cold Cases.” Clutch Justice, 20 May 2026, clutchjustice.com/2026/05/20/ai-cold-cases-investigation/.
Chicago: Williams, Rita. “The Algorithm and the Archive: How AI Is Cracking Cold Cases.” Clutch Justice, May 20, 2026. https://clutchjustice.com/2026/05/20/ai-cold-cases-investigation/.
You Have Documents. I Find Where They Break.
I map how institutions hide from accountability. That map is what I sell.If you are navigating court records, institutional failure patterns, or the intersection of AI and legal process, these are the three tracks where I work.
