Forensics, particularly police forensics, has been grappling with human biases for a long time. The advent of Artificial Intelligence (AI) has only complicated the issue further, with experts warning of its potential to worsen existing racial and gender biases in initial witness descriptions. The recent creation of the Forensic Sketch AI-rtist program by two developers, Artur Fortunato and Filipe Reynaud, is a case in point.
Developed as part of a hackathon in December 2022, the program uses OpenAI’s DALL-E 2 image generation model to create highly realistic police sketches of suspects based on user inputs. The developers claim that the program is aimed at reducing the time it takes to create a suspect sketch, which they state takes “around two to three hours”. However, despite their intentions, AI ethicists and researchers have raised concerns about the use of generative AI in police forensics.
Jennifer Lynch, the Surveillance Litigation Director of the Electronic Frontier Foundation, told Motherboard that traditional forensic sketches already suffer from human biases and memory frailties, which AI can’t solve. In fact, Lynch believes that AI-based programs like Forensic Sketch AI-rtist will likely make these problems worse.
The program requests information from users in two ways: through a template that asks for gender, skin color, eyebrows, nose, beard, age, hair, eyes, and jaw descriptions or through an open description feature, where users can type in any description they have of the suspect. After the user inputs the information, the “generate profile” button sends the descriptions to DALL-E 2, producing an AI-generated portrait.
But this method of sketch generation is problematic. As Lynch explains, human memory recalls faces holistically, not feature by feature. A sketch process that relies on individual feature descriptions can result in a face that is significantly different from the actual suspect. Furthermore, the AI-generated image may replace the witness’ hazy memory of the actual suspect, particularly since AI-generated images often look more “real” than hand-drawn sketches.
The situation is particularly concerning for Black and Latino individuals, who are more likely to be stopped, searched, and suspected of a crime without cause. If AI-generated forensic sketches are released to the public, they could reinforce stereotypes and racial biases and mislead investigations by directing attention to individuals who resemble the sketch instead of the actual perpetrator. In fact, mistaken eyewitness identifications have contributed to 69% of wrongful convictions later overturned by DNA evidence in the US. Additionally, false or misleading forensics, including police sketches, have contributed to nearly 25% of all wrongful convictions across the country.
Adding AI to the already unreliable process of witness descriptions only exacerbates the issue. Research Scientist Sasha Luccioni of Hugging Face pointed out that DALL-E 2 is known to contain biases, such as displaying mostly white men when asked to generate an image of a CEO. Although efforts are being made to mitigate bias in AI output, the source of these biases remains unclear, making it difficult to take the right measures to correct them. According to Luccioni, marginalized groups are most impacted by these biases, as they are already marginalized by existing biases in datasets, lack of oversight, and racist and unfair representations of people of color on the internet. She describes the process as a feedback loop in which AI models contain, produce, and perpetuate bias as the images they generate continue to be used.
Despite these warnings, Fortunato and Reynaud stand by their program, stating that it runs with the assumption that police descriptions are trustworthy, and that “police officers should be the ones responsible for ensuring that a fair and honest sketch is shared”. The developers argue that any inconsistencies in
the AI-generated sketches can be attributed to human error in the input descriptions, not the program itself.
However, this argument fails to acknowledge the systemic and implicit biases that exist in society, particularly within law enforcement. These biases can lead to inaccurate descriptions and reinforce harmful stereotypes. It’s important to note that AI is only as unbiased as the data it’s trained on, and when it comes to police forensics, this data is often skewed by racial and gender biases.
It’s crucial that developers and law enforcement agencies take into account the potential negative consequences of AI-based forensic tools, especially since these tools have the potential to harm marginalized communities. Until AI systems can guarantee unbiased results, it’s essential to approach their use in police forensics with caution and skepticism. Additionally, alternative approaches to forensic sketch generation, such as using professional artists or 3D modeling software, should be considered.
