Ontario prison AI assigns black prisoners harsher living conditions | Hacker News Reader

Ontario prison AI assigns black prisoners harsher living conditions

breachmedia.ca

9 pointshn_acker5 comments

The original title is:

> Black prisoners are assigned harsher living conditions in Ontario jails—thanks to AI

> The ministry responsible for Ontario’s prisoners agrees. It wrote in internal training documents viewed exclusively by The Breach that “Indigenous and racialized individuals face systemic discrimination in our justice system … As a result, assessments like SAFER would likely contribute to the overrepresentation of Indigenous inmates in maximum security.”

If "Indigenous and racialized individuals" (whatever the fuck that means) are more violent (even in prison), then it would make sense to put them in higher security prisons to lower likelihood of harm to other prisoners and/or guards. But that isn't what's happening, as the race of the individual prisoner is not even inputted into the algorithm, only their behavior is: "SAFER inputs a prisoner’s personal information—including arrests, charges, and disciplinary records—into an algorithm."

You're ignoring the garbage-in-garbage-out problem that forms the basis of the article.

An individual person's previous criminal behavior is not "garbage" for predicting their future behavior -- in fact prior behavior is one of the best predictors for future behavior that we have. Even the Prison Policy Initiative's own data shows this.

>> The Security Assessment for Evaluating Risk (SAFER) program has been operating quietly in Ontario’s jails since early 2021. SAFER inputs a prisoner’s personal information—including arrests, charges, and disciplinary records—into an algorithm.

> An individual person's previous criminal behavior is not "garbage" for predicting their future behavior

Just as in your first reply, you're ignoring the main problem posed by the article, that the information getting inputted is not truthful for some individuals and not equally truthful across different individuals. People of different races are not getting the same information put in for the same crimes because a person of one race gets a charge or conviction where a person of another race might not [1]. People of one race tend to get more severe punishments for the same type and level of crime compared to people of another race [2]. ([1] and [2] are linked in TFA.) Disciplinary actions (and thus, disciplinary records) toward detainees (before/no conviction) and inmates (after conviction) by jail and prison staff respectively are arbitrary and subject to in-the-moment biases that might normally get diluted (not always eliminated) during the process between charge and conviction. If all of these data were fed into an algorithm as if equally weighted and equally truthful, it's a given that such an algorithm would produce biased outcomes, even if the algorithm were not the source of the bias. Another problem emphasized in the article is that the public doesn't know how the SAFER algorithm works nor how the government inputs data.

The Breach's investigation suggests that at least one (possibly both) of the following are happening:

1. Black people are getting put in higher level "security" (meaning isolation) than they should be.

2. People of at least one other race are getting put in lower level "security" than they should be.

> in fact prior behavior is one of the best predictors for future behavior that we have. Even the Prison Policy Initiative's own data shows this.

(For clarity: statistical predictions are correlations, not causation.) Here you're overstating how meaningful the correlation is for the purposes of deciding both the dangerousness of a person with multiple convictions and the appropriate different severity of punishments for people with different numbers of convictions [3]:

> An analysis of prior convictions shows that even those with many prior convictions are still unlikely to be rearrested for a new violent crime while on release. Figure 1 divides all those released into four groups based on their number of prior convictions. We then calculate the fraction of each group that was rearrested for a violent crime. About 1.5% of those with no prior record were rearrested, a number that rose only slightly for defendants with one conviction. On the other end of the spectrum, about 2.5% of those with four or more prior convictions were rearrested. Thus, the rearrest rate was not quite twice as high for those with many convictions as for those with none. A single past conviction, though, hardly appears to predict any increased dangerousness. Repeat offenders—those who have four prior convictions—were only arrested for pretrial violent crime in about 1 in 30 instances.

[1] https://www.ohrc.on.ca/en/impact-action-final-report-anti-bl...

[2] https://www.justice.gc.ca/eng/rp-pr/jr/rrbb-bbrr/conclusions...

[3] https://static.prisonpolicy.org/scans/Baradaran_predictingvi...

Reply on news.ycombinator.com