<s>
Correctional	B-Algorithm
Offender	I-Algorithm
Management	I-Algorithm
Profiling	I-Algorithm
for	I-Algorithm
Alternative	I-Algorithm
Sanctions	I-Algorithm
(	O
COMPAS	B-Algorithm
)	O
is	O
a	O
case	O
management	O
and	O
decision	B-Application
support	I-Application
tool	I-Application
developed	O
and	O
owned	O
by	O
Northpointe	O
(	O
now	O
Equivant	O
)	O
used	O
by	O
U.S.	O
courts	O
to	O
assess	O
the	O
likelihood	O
of	O
a	O
defendant	O
becoming	O
a	O
recidivist	O
.	O
</s>
<s>
COMPAS	B-Algorithm
has	O
been	O
used	O
by	O
the	O
U.S.	O
states	O
of	O
New	O
York	O
,	O
Wisconsin	O
,	O
California	O
,	O
Florida	O
's	O
Broward	O
County	O
,	O
and	O
other	O
jurisdictions	O
.	O
</s>
<s>
The	O
COMPAS	B-Algorithm
software	O
uses	O
an	O
algorithm	O
to	O
assess	O
potential	O
recidivism	O
risk	O
.	O
</s>
<s>
According	O
to	O
the	O
COMPAS	B-Algorithm
Practitioner	O
's	O
Guide	O
,	O
the	O
scales	O
were	O
designed	O
using	O
behavioral	O
and	O
psychological	O
constructs	O
"	O
of	O
very	O
high	O
relevance	O
to	O
recidivism	O
and	O
criminal	O
careers.	O
"	O
</s>
<s>
General	O
recidivism	O
scale	O
The	O
General	O
recidivism	O
scale	O
is	O
designed	O
to	O
predict	O
new	O
offenses	O
upon	O
release	O
,	O
and	O
after	O
the	O
COMPAS	B-Algorithm
assessment	O
is	O
given	O
.	O
</s>
<s>
In	O
July	O
2016	O
,	O
the	O
Wisconsin	O
Supreme	O
Court	O
ruled	O
that	O
COMPAS	B-Algorithm
risk	O
scores	O
can	O
be	O
considered	O
by	O
judges	O
during	O
sentencing	O
,	O
but	O
there	O
must	O
be	O
warnings	O
given	O
to	O
the	O
scores	O
to	O
represent	O
the	O
tool	O
's	O
"	O
limitations	O
and	O
cautions.	O
"	O
</s>
<s>
A	O
general	O
critique	O
of	O
the	O
use	O
of	O
proprietary	O
software	O
such	O
as	O
COMPAS	B-Algorithm
is	O
that	O
since	O
the	O
algorithms	O
it	O
uses	O
are	O
trade	O
secrets	O
,	O
they	O
cannot	O
be	O
examined	O
by	O
the	O
public	O
and	O
affected	O
parties	O
which	O
may	O
be	O
a	O
violation	O
of	O
due	O
process	O
.	O
</s>
<s>
Additionally	O
,	O
simple	O
,	O
transparent	O
and	O
more	O
interpretable	O
algorithms	O
(	O
such	O
as	O
linear	B-General_Concept
regression	I-General_Concept
)	O
have	O
been	O
shown	O
to	O
perform	O
predictions	O
approximately	O
as	O
well	O
as	O
the	O
COMPAS	B-Algorithm
algorithm	O
.	O
</s>
<s>
Specifically	O
,	O
COMPAS	B-Algorithm
risk	O
assessments	O
have	O
been	O
argued	O
to	O
violate	O
14th	O
Amendment	O
Equal	O
Protection	O
rights	O
on	O
the	O
basis	O
of	O
race	O
,	O
since	O
the	O
algorithms	O
are	O
argued	O
to	O
be	O
racially	O
discriminatory	O
,	O
to	O
result	O
in	O
disparate	O
treatment	O
,	O
and	O
to	O
not	O
be	O
narrowly	O
tailored	O
.	O
</s>
<s>
The	O
team	O
found	O
that	O
“	O
blacks	O
are	O
almost	O
twice	O
as	O
likely	O
as	O
whites	O
to	O
be	O
labeled	O
a	O
higher	O
risk	O
but	O
not	O
actually	O
re-offend	O
,	O
”	O
whereas	O
COMPAS	B-Algorithm
“	O
makes	O
the	O
opposite	O
mistake	O
among	O
whites	O
:	O
They	O
are	O
much	O
more	O
likely	O
than	O
blacks	O
to	O
be	O
labeled	O
lower-risk	O
but	O
go	O
on	O
to	O
commit	O
other	O
crimes.	O
”	O
They	O
also	O
found	O
that	O
only	O
20	O
percent	O
of	O
people	O
predicted	O
to	O
commit	O
violent	O
crimes	O
actually	O
went	O
on	O
to	O
do	O
so	O
.	O
</s>
<s>
A	O
subsequent	O
study	O
has	O
shown	O
that	O
COMPAS	B-Algorithm
software	O
is	O
somewhat	O
more	O
accurate	O
than	O
individuals	O
with	O
little	O
or	O
no	O
criminal	O
justice	O
expertise	O
,	O
yet	O
less	O
accurate	O
than	O
groups	O
of	O
such	O
individuals	O
.	O
</s>
<s>
COMPAS	B-Algorithm
,	O
by	O
contrast	O
,	O
has	O
an	O
accuracy	O
of	O
65	O
percent.	O
"	O
</s>
