A cold-case homicide investigation in California shows how quickly separate investigative technologies can be combined in ways their creators say were never intended. In 2017, detectives at the East Bay Regional Park District Police Department sent crime-scene genetic information to Parabon NanoLabs, a company that says it can use DNA to predict a person's appearance.
The case involved Maria Jane Weidhofer, who was found dead and sexually assaulted at Berkeley, California's Tilden Regional Park in 1990. Nearly 30 years later, police received a computer-generated face based on DNA evidence. Then, in 2020, a detective asked for that face to be searched with facial recognition technology.
How a DNA Sample Became a Face
Parabon NanoLabs processed the suspect's DNA through its proprietary machine learning model and returned what it calls a Snapshot Phenotype Report. The result was not a photograph. It was a 3D rendering based on genetic attributes found in the sample.
The company predicted that the suspect was male, with fair skin, brown eyes and hair, no freckles, and bushy eyebrows. A forensic artist working for the company added a nondescript, close-cropped haircut and a mustache. The mustache was not derived from the DNA sample; it was based on a witness description.
In 2017, the department published the predicted face to seek tips from the public. That decision was already controversial because the image was an estimate, not a direct record of a person. The later request to use it in facial recognition raised a different concern: whether an algorithmic guess should be treated like an actual image of a suspect.
The Facial Recognition Request
In 2020, a detective asked the Northern California Regional Intelligence Center for analytical support. The center is described as a fusion center that helps federal, state, and local police departments collaborate.
In the request, the detective wrote: “Using DNA found at the crime scene, Parabon Labs reconstructed a possible suspect’s facial features,” and added, “I have a photo of the possible suspect and would like to use facial recognition technology to identify a suspect/lead.”
The request was found in hacked police records published by the transparency collective Distributed Denial of Secrets. According to the source article, it appears to be the first known instance of a police department trying to use facial recognition on a face generated from crime-scene DNA.
It is unknown whether the Northern California Regional Intelligence Center ran the search. The NCRIC did not respond to WIRED's requests for comment about the result. Captain Terrence Cotcher of the East Bay Regional Park District PD would not comment on the identification request, citing an active homicide investigation. Mike Sena, the executive director of the NCRIC, told The Markup in 2021 that when the fusion center gets facial recognition requests, it will run a search.
Why Experts Object
For privacy advocates and facial recognition experts, the problem is not only one department's decision. The deeper issue is that law enforcement can combine tools in ways that have not been tested together and may not have clear oversight.
Jennifer Lynch, general counsel at the Electronic Frontier Foundation, told WIRED: “It’s really just junk science to consider something like this.” She argued that facial recognition is more likely to misidentify someone when the input is unreliable, such as a face generated by an algorithm.
Lynch also said: “There’s no real evidence that Parabon can accurately produce a face in the first place.” Her concern is that people could become suspects for crimes they did not commit because a speculative image is treated as a searchable identity.
That risk is central to the controversy. A DNA phenotype rendering can suggest broad traits, but facial recognition systems are designed to compare faces. Using one prediction system as the input for another identification system creates a chain of uncertainty.
What Parabon Says Its Technology Can and Cannot Do
Parabon NanoLabs was founded in 2008 and primarily provides forensic genetic genealogy services for law enforcement. That work involves comparing DNA data with profiles in genealogy databases to help locate potential suspects or victims.
In 2012, the company received a grant from the US Department of Defense's Defense Threat Reduction Agency to study DNA phenotyping. According to a 2020 article in Nature, the DOD was initially interested in technology that could re-create faces of people who made improvised explosive devices, using DNA left on bomb fragments.
Ellen Greytak, Parabon NanoLabs' director of bioinformatics, says the company uses machine learning to build predictive models “for each part of the face.” The models are trained on DNA data from more than 1,000 research volunteers and paired with 3D scans of their faces. Greytak says each scanned face has 21,000 phenotypes, meaning observable physical traits, that the models analyze.
Parabon says it can confidently predict hair color, eye color, skin color, freckling, and general face shape. But its methods have not been peer-reviewed, and scientists are skeptical about whether predicting face shape is feasible.
Paula Armentrout, Parabon NanoLabs vice president, told WIRED that although the company's methods are not public, it has presented its work at conferences and tested the technology on thousands of samples. She said the company posts “every single composite that is publicly disclosed by a customer, so people can draw their own conclusions about how well our technology works.”
The Terms of Service Gap
Parabon itself says using its face predictions for facial recognition is a bad idea. When the company launched its face-prediction service around 2015, its terms of service did not explicitly ban that use. Soon after, law enforcement clients began asking whether phenotype-generated faces could be run through facial recognition tools.
Greytak said: “We were surprised when we heard this.” She added: “It’s just not the intended purpose of the composite images.”
In 2016, Parabon added a clause prohibiting customers from using facial recognition on its Snapshot Phenotype Reports. But Armentrout told WIRED the company “does not have a way to ensure compliance” with its terms of service.
That gap matters because it leaves enforcement largely outside the technology provider's control. A tool presented as a way to generate investigative leads can become the starting point for a facial recognition search, even when the company says the rendering cannot support individual identification.
The Weidhofer case shows the stakes of that shift. A DNA-generated face may look concrete, but the source article makes clear it is a prediction shaped by models, assumptions, and in at least one detail, a witness description. Treating that image as if it were a real suspect photo can turn an uncertain lead into a search target.