AI-Powered Police Sketches Raise Accuracy and Bias Concerns
A new tool that uses artificial intelligence to produce police sketches from witness descriptions is drawing sharp criticism from experts who say it could distort memories and deepen existing biases in law enforcement. The program, called “Forensic Sketch AI-rtist,” was built by independent developers using OpenAI’s DALL-E 2, a text-to-image model that generates realistic images from written prompts.
According to a report by Vice, the developers have not yet released the program. They told Vice in a joint email that they plan to approach police departments to obtain input data for testing. The tool’s interface presents a fill-in template asking for details such as gender, age, skin color, hair, eyebrows, eyes, nose, beard, and jaw shape. Once a user completes the form, they can select the number of images they want and hit “generate profile,” and the system produces sketches via DALL-E.
But the approach is fundamentally flawed, according to Jennifer Lynch, a lawyer with the Electronic Frontier Foundation (EFF), a digital rights group. “Research has shown that humans remember faces holistically, not feature-by-feature,” Lynch told Vice. She explained that a sketch process relying on individual feature descriptions, as this AI program does, can produce a face that is “strikingly different from the perpetrator’s.” More concerning, she added, is that once a witness sees a generated composite, it may “replace” what they originally remembered, especially if the image is uncannily lifelike.
Bias in the Machine
The concerns extend beyond accuracy to the potential for the AI to exacerbate racial and gender biases that already affect human-made sketches. OpenAI’s DALL-E 2 has been widely documented to produce stereotypical and sometimes overtly racist or misogynistic results. Sasha Luccioni, a researcher at Hugging Face, a platform for open-source AI projects, told Vice that such biases are a “compounding factor.” She noted that marginalized groups are often further marginalized by these technologies because of biases in the training datasets, lack of oversight, and the prevalence of racist and unfair representations of people of color on the internet.
The cumulative effect, experts say, is a chain of biases: the witness’s own recollections, the officers’ questioning, and now the AI’s inherent prejudices. The generated images, regardless of their accuracy, could reinforce the preconceptions of the people and systems involved, as well as the public who may see them.
This is not the first time AI has been used in police work. In a separate incident, cops uploaded an image of a suspect generated from DNA, but deleted it after mass criticism. The new tool, however, marks a more direct application of AI to a traditional investigative technique.
While the developers have not yet released the program, their intention to collaborate with police departments raises practical and ethical questions. Law enforcement agencies considering such tools will need to weigh the potential for inaccurate or biased sketches against the promise of faster, cheaper image generation.
