Uploaded October 2024 | Updated September 2026, 2 weeks ago
Andrea Roth, University of California, Berkeley
The use of algorithms in the criminal justice system to predict dangerousness (for purposes of pretrial and post-conviction detention) or flag potential crimes in progress (through algorithm-enhanced predictive policing using facial recognition and other technologies, for purposes of justifying a stop or arrest) has been the subject of much recent criticism, based on concerns about inaccuracy, feedback loops, racial disparities, lack of transparency about the system's inputs, and more. What has received less attention is the use of algorithms to generate claims that are introduced as evidence of guilt or innocence at trial. For example, a criminal defendant in 2024 might be accused not only by live witnesses but by a probabilistic genotyping program's "likelihood ratio" claiming that a DNA mixture is 10 trillion times more likely if he was a contributor; a machine-learning classifer's determination that he authored a Tweet confessing to murder; FitBit data suggesting that he was asleep at the time he claims to have done something; software claiming his blood-alcohol level was .09; or Google Earth results alleging he was in a particular location. Most of these algorithms are developed by private industry and are guarded as proprietary "trade secrets," with companies not only refusing to disclose the inner workings of the program, but even refusing to allow independent researchers to purchase licenses to conduct audits themselves. Meanwhile, the same concerns about algorithmic accuracy and fairness surround the use of this proof at trial. In one recent homicide case, two algorithms came to diametrically opposed results in a homicide case when interpreting the identical DNA mixture data. Courts and lawmakers seem to be vaguely aware of the potential abuses of this regime, but current rules of evidence and case law do little to regulate machine-generated proof. This seminar will offer an overview of the use of machine-generated proof, the gaps in existing rules, potential solutions and challenges for courts and lawmakers, and suggestions for how the research community can help improve justice outcomes.
Learn more, follow us on social media and check out our podcasts:
https://linktr.ee/sfiscience
Andrea Roth, University of California, Berkeley
The use of algorithms in the criminal justice system to predict dangerousness (for purposes of pretrial and post-conviction detention) or flag potential crimes in progress (through algorithm-enhanced predictive policing using facial recognition and other technologies, for purposes of justifying a stop or arrest) has been the subject of much recent criticism, based on concerns about inaccuracy, feedback loops, racial disparities, lack of transparency about the system's inputs, and more. What has received less attention is the use of algorithms to generate claims that are introduced as evidence of guilt or innocence at trial. For example, a criminal defendant in 2024 might be accused not only by live witnesses but by a probabilistic genotyping program's "likelihood ratio" claiming that a DNA mixture is 10 trillion times more likely if he was a contributor; a machine-learning classifer's determination that he authored a Tweet confessing to murder; FitBit data suggesting that he was asleep at the time he claims to have done something; software claiming his blood-alcohol level was .09; or Google Earth results alleging he was in a particular location. Most of these algorithms are developed by private industry and are guarded as proprietary "trade secrets," with companies not only refusing to disclose the inner workings of the program, but even refusing to allow independent researchers to purchase licenses to conduct audits themselves. Meanwhile, the same concerns about algorithmic accuracy and fairness surround the use of this proof at trial. In one recent homicide case, two algorithms came to diametrically opposed results in a homicide case when interpreting the identical DNA mixture data. Courts and lawmakers seem to be vaguely aware of the potential abuses of this regime, but current rules of evidence and case law do little to regulate machine-generated proof. This seminar will offer an overview of the use of machine-generated proof, the gaps in existing rules, potential solutions and challenges for courts and lawmakers, and suggestions for how the research community can help improve justice outcomes.
Learn more, follow us on social media and check out our podcasts:
https://linktr.ee/sfiscience
