Arizona Police Deploy AI-Generated Sketches to Identify Suspects, Raising Bias Concerns

Goodyear Police Department Pioneers AI-Generated Suspect Sketches

The Goodyear Police Department in Arizona has begun using ChatGPT to create photorealistic images of suspects, marking what experts believe is a first for U.S. law enforcement. The department has deployed AI-generated sketches twice this year: first in April for an attempted kidnapping case, and again in November for a shooting investigation near 143rd and Clarendon avenues 1

1

2

2

. Both cases remain open, and despite generating public attention, the AI-generated suspect sketches have not resulted in any arrests 3

3

.

Mike Bonasera, a forensic artist who has been creating police sketches for about five years, spearheaded this shift toward using AI in the identification process. In a typical year, he produces up to seven composite drawings when police lack clear surveillance images 1

1

. His process involves interviewing witnesses directly as he generates the images through OpenAI's ChatGPT, allowing them to suggest changes in real-time to details like facial features and expressions 2

2

.

From Traditional Hand-Drawn Sketches to Photorealistic Images of Suspects

Bonasera's journey with Police AI began earlier this year when he tested the technology by uploading old traditional hand-drawn sketches into ChatGPT. He was struck by how closely the AI-generated images resembled real suspects who had been previously identified 2

2

. After seeking approval from department leadership and the Maricopa County Attorney's Office, he received clearance to begin using ChatGPT to generate sketches for new cases 1

1

.

The decision to replace traditional methods stems largely from public engagement concerns. "We're now in a day and age where if we post a pencil drawing, most people are not going to acknowledge it," Bonasera explained 1

1

. He believes residents, particularly younger individuals, respond more actively to photorealistic images than to pencil sketches. The April image prompted an increase in public tips, convincing him to repeat the process for identifying crime suspects in November 2

2

.

AI Bias and Legal Challenges Loom Over New Technology

Experts have raised significant concerns about the reliability of AI and its potential to introduce AI bias into suspect identification. When image generators like ChatGPT create images of people, they draw from massive databases of photos of real people, meaning any biases present in those collections will appear in the final output 3

3

. Bryan Schwartz, an associate clinical professor of law at the University of Arizona who studies AI in policing, noted that early AI generators "were really good at creating white faces and not as good at creating some of other races" 3

3

.

Andrew Ferguson, a law professor at George Washington University, highlighted potential legal challenges ahead. "In court, we all know how drawing works and can evaluate how much reliability to give the human drawn sketch," he wrote. "In court, no one knows how the AI works" 1

1

. He emphasized that while an AI-generated image might appear more realistic, that doesn't necessarily make it more reliable than traditional methods.

The concerns about AI in law enforcement extend beyond suspect sketches. Police departments relying on facial recognition technology have already led to numerous false arrests in cities including Detroit, New York, and Atlanta 3

3

. Bryan Schwartz indicated that he is not aware of any other U.S. law enforcement agency using AI to create suspect images, suggesting Goodyear may be operating in legally and procedurally untested territory 2

2

.

What This Means for Law Enforcement and Witness Accounts

Despite the concerns, the Goodyear Police Department remains optimistic about the technology's potential. "We are hopeful that these new techniques and AI technology will assist in solving more complex cases in the future, here in Arizona and around the country," the department stated 2

2

. The November image released by the department showed a middle-aged man wearing a hoodie and beanie with a surprised expression, which Bonasera says the witness repeatedly mentioned during the creation process 2

2

.

Each AI-generated image carries a disclaimer stating: "This AI-generated image is based on victim/witness statements and does not depict a real person" 2

2

. However, the department's insistence that these aren't "AI fabrications" because they begin as traditional composite drawings from witness accounts remains a contentious point among legal experts 3

3

. As other departments watch this experiment unfold, the balance between leveraging technology for public engagement and maintaining accuracy in suspect identification will likely face increasing scrutiny in courtrooms and communities across the nation.