Eyewitness Memory vs. Face Recognition Systems: Can Machines See Better Than We Do?

 Eyewitness misidentification is a well-documented problem that can have devastating consequences. While eyewitness accounts are often crucial in criminal investigations, scientific studies have shown that memory can be unreliable, especially under stressful circumstances. 

On the other hand, AI systems like Face Recognition Systems  (FRS) are becoming increasingly common in law enforcement. These systems can analyze footage from security cameras and other sources to identify suspects. But are they more reliable than human memory?

Kleider and colleagues' (2024) research addresses this critical question.  Their study compared the accuracy of FRS with human eyewitness identification in a controlled setting. Kleider and colleagues (2024) measured:

1. Discriminability is the ability to distinguish between the culprit and innocent suspects.
2.  Reliability is the relationship between confidence and accuracy.

Building a Diverse Sample

To investigate how well facial recognition systems compare to eyewitness memory, Kleider and colleagues (2024):

1. Recruited 237 participants through Georgia State University's undergraduate subject pool.
2. An online platform for research studies.
3. The participants were between 18 and 66 years old, and researchers aimed for a diverse group and reported their racial and gender makeup.

Can You Spot the Criminal? Testing Memory and Machines

The researchers pitted human memory against facial recognition technology in a two-part experiment, and here is how Kleider and colleagues (2024) did it:

1. Human participants watched crime scene videos online. After each video, they completed a distracting task before looking at a lineup with either the culprit or similar-looking innocent people; then, they had to identify the perpetrator or indicate they weren't present, along with their confidence level.
   
2. Secondly, the same videos and lineups were fed into the FRS, and just like law enforcement might do, the FRS grabbed a single frame of each culprit's face and compared it to every face in the lineup. It then generated a similarity score for each comparison.

Who Won the Face Recognition Showdown?

Kleider and colleagues (2024) concluded from their results:

1. The FRS generally did better than people, especially with blurry videos.
2. FRS and memory accuracy dropped for unclear videos, stressing the importance of good footage.
3. The tested FRS showed no racial bias, and surprisingly, eyewitness performance was alson't affected by race, possibly due to the study's diverse participants.
4. 
   The FRS ideally identified faces with the highest similarity scores, which suggests that prioritizing high-similarity matches in investigations could reduce mistakes.

The researchers caution that these results are based on one FRS and need confirmation with other systems and more video variations.  They also reveal that their research has implications for law enforcement:

• FRS as a Tool: FRS can be valuable, especially with low-quality videos. However, limitations and potential errors require cautious interpretation of results.
  
  
  
  

Interpreting FRS Outputs: Police officers should consider FRS similarity scores and other evidence.

Regulations and Training Needed: Law enforcement's widespread use of FRS demands clear rules and training programs to ensure proper use.

References

Kleider, Heather & Stevens, Beth & Mickes, Laura & Boogert, Stewart. (2024). Application of artificial intelligence to eyewitness identification. Cognitive research: principles and implications. 9. 19. 10.1186/s41235-024-00542-0.