Washington Lawyer - October 2019 - 17

someone unfairly receiving a harsher sentence or waiting longer for parole
than is appropriate."

should look for opportunities to audit the data sets, training parameters, and
assumptions built into AI solutions and not just hope for the best.

Challenging the validity of criminal justice data inputs is relatively simple in
the United States, experts say, given that errors can be easily tracked to its
long history of inequitable arrest and prosecutions due to unjust and racially
biased laws.

"We are a society of data now," says Lezak. "People try to lean data one way or
another all the time. It's as easy to manipulate data as much as it is anything
else. It comes down to honesty, and a machine cannot guarantee honesty.
What we need to do is teach people to look behind the data and at the source.
You don't take anything at face value."

"It's one of the fundamental problems of prediction in the criminal justice
system," says Logan Koepke, senior policy analyst with Upturn, a nonprofit that
advocates for equity in digital technology. "All predictions rely on historical data,
and historical data reflects race discrimination and bias across the entire system.
It's going to be very hard but necessary to find technical fixes that address the
problem."
This is especially critical when AI exceeds its original productivity mission
and is applied to the criminal justice system. Even minor flaws in the governing algorithms could exacerbate the problematic disparities that already
mark the system.

Others suggest that while the American business model uses proprietary
algorithms and software that cannot be easily accessed, open-source products
would be more effective in allowing for periodic assessments of the quality
of the data and the impartiality of its sources.
Finally, LegalBrain's Richard observes that the biases built into human decisions
are hard-wired: The known is preferred over the unknown, the familiar over the
unfamiliar. While these biases have kept the human race alive, they also have
established a pattern of fearing "the other." Machines have no legacy or context
biases, but there is always the human connection.

"When you think about data for predictive policing or determining recidivism,
most systems are going to use prior arrest records," says Koepke. "Arrests
aren't the perfect measure of any action. If anything, arrest records document
the practices and behaviors of police officers rather than being a true
measure of crime."

"I don't think we can be completely rid of implicit bias," says Richard. "You really
don't want to. It comes from a good place. What we need to do is understand
and manage those situations where biases don't work, like in the workplace or
the courtroom."

MAINTAINING VIGILANCE

Sarah Kellogg is a regular contributor to Washington Lawyer.

What experts suggest is that whoever uses AI to weed out implicit bias do so
with the necessary due diligence at every stage of the process. Law firms

Composite image, Getty Images/John M. Lund Photography Inc./DigitalVision; Darin Snyder, courtesy of
O'Melveny & Myers LLP; Adrienne Fowler, courtesy of Harris, Wiltshire & Grannis LLP

2019

FREDERICK
DOUGLASS

AWARDS

07 NOVEMBER 2019
CONRAD WASHINGTON DC HOTEL
950 New York Ave., NW
Washington, DC 20001
For additional information, visit
ww.schr.org/dinner.

PRESENTING
SPONSORS

AND

HONORING

SCOTT BUDNICK
Film Producer
Founder of the Anti-Recidivism Coalition
President and CEO of One Community, LLC

AND
THE HONORABLE
ROBERT REYNOLDS "RENNY" CUSHING
Member of the New Hampshire House of Representatives
Founder and Executive Director
of Murder Victims' Families for Human Rights

https://www.schr.org/event/23rd_annual_dinner

Washington Lawyer - October 2019

Table of Contents for the Digital Edition of Washington Lawyer - October 2019