Idaho legislators approve law requiring transparency for risk assessment tools | Hacker News Reader

Idaho legislators approve law requiring transparency for risk assessment tools

muckrock.com

48 pointsdanso15 comments

This is lawyer trickery. Do not think that this transparency law is meant to ensure fair justice, expect the opposite. This is the equivalent of the common "open file" rule of prosecutors. Rather than sift through and release relevant material to defendants (aka Brady material) most prosecutors just declare their files open for inspection. It is then on the defense to do the sifting, to slog through the growing pile of irrelevant material. So too with this law. Rather than properly manage the system, this is the legislature throwing its hands in the air. The burden will now be on defendants to become AI experts prior to making challenges to this system. The poor defendant seeking bail will have to point our the flaws in the code.

You make a valid point that this places additional burdens on defense. However, I feel like such openness is still better than allowing prosecutors to cherry-pick information to provide, or "accidentally" leave out a chunk of data, or require repeated specific motions to request each detail as the defense becomes aware of it.

One benefit of the "old" way is that if the prosecution omits evidence, it can be grounds for an appeal. If the defense skips something, there's no grounds and they're just out of luck.

Sorry, but your analogy/comparison makes no sense. Could you provide a citation/reference to an "open file" rule that you've seen abused in such a way that it would be better to not have that rule at all?

According to TFA, the Idaho state law mandates that records/data used in the design and testing of these risk assessment tools to be made public. Presumably, this information was not required to ever be disclosed upon public request.

So how is "These records must be made public upon request" a net negative compared to the status quo of "The government is not required to disclose these records"?

Yeah, the problem with this is that it's woefully insufficient, not that it's not a step in the right direction.

Except that once a flaw is identified it can be shared with future defendants. They do not use a unique AI for each person. This will result in the AI actually being fixed to prevent people from "getting off".

Except that under Brady, Giglio, and their progeny, "open file" policies are legally insufficient to meet the requirements imposed by those cases. Although a few states have not passed laws or rules to enforce those court opinions, most have either explicitly or through standard discovery rules. Furthermore, most prosecutors, myself and my office colleagues included, actively provide any and all non-privileged information we obtain in a case and flag anything that is obvious Brady or Giglio material.

Note: the law was passed by both legislative chambers in March, and was then signed by the governor on March 28th, to go into effect in July 2019:

https://legislature.idaho.gov/sessioninfo/2019/legislation/h...

>It is also the first to set requirements for public access to the design and data used to build the tool’s predictions, requiring that records created during a risk assessment tool’s development, validation, and use to “be open to public inspection, auditing, and testing.”...

Now these are the kinds of laws we need! This is the sort of thing we should be doing all across the US for a lot of software being used, allegedly, "in the public interest".

I'm surprised this didn't make any local news here in Idaho. And such an openness requirement seems at odds for Idaho's usual politics (cf. Idaho's Ag Gag laws).

The only coverage I saw was from Nathan Brown at the Idaho Press.

https://www.idahopress.com/eyeonboise/bill-to-regulate-bail-...

This is why they happily get my subscription dollars every month.

Things happen like this when it's not a issue established by any parties yet.

this is the exact thing that Rule Lists can do as proven by Cynthia Rudin and her team show https://www.youtube.com/watch?v=MjxcwKN2dXs and https://today.duke.edu/2017/07/opening-lid-criminal-sentenci... She also has an open source code for the rule lists here https://users.cs.duke.edu/~cynthia/code.html

Isn't this rather difficult? I thought tracking back an output of something like a neural net to the salient input features was a non-trivial process.

Reply on news.ycombinator.com