GlassBox:

GlassBox at MIT PolicyHackathon April-2019. Bringing transparency and auditability to recidivism risk-assessment and parole/pre-trial decision-making process.

Inspiration:

Making AI more interpretable and transparent to the judges while adjudicating criminal cases. Currently available recidivism assessment tools (e.g. Compas) provides black-box recommendations to judges.

Built With:

• Scikit-learn
• Numpy
• Python v3.6
• Bootstrap
• HTML 5
• CSS 3

Description:

• Software transparency

Recidivism risk-assessment software must provide statistically significant score followed by a list of contributing factors, allowing judges to tweak the inputs to influence the score

• Decision-making process auditability

Changes to original score must be justified by the judges, subject to appeal court’s audit

What it does:

• Displays the likelihood of recidivism (in %)
• Offers the choice to pull out Top-K(1<=K<=30) factors that contribute to the likelihood of recidivism score
• Allows toggling amongst all of the factors to see the score deviation

Acheivements:

• We won the 'Artificial Intelligence' category prize in this MIT policy hackathon in April 2019
• We were one of the Top 5 teams to compete for the 1st prize!

What's next for Glass-Box:

• Training Glass-Box to surpass Compas accuracy
• Providing statistics and benchmarks to judges(individual trajectory, benchmark with state national level)