Algorithmic Transparency

[Cramulus]

I was just reading an article about facebook taking a more active role in "combatting misinformation"...

Facebook says it is taking more steps to encourage voting, minimize misinformation and reduce the likelihood of post-election "civil unrest."

..."This election is not going to be business as usual. We all have a responsibility to protect our democracy," Facebook CEO Mark Zuckerberg said in a post on Thursday.

Yeah that all sounds great. But I have zero faith that the people running Facebook are actually "protecting democracy" -- they are participants in democracy, not moderators, and have their own agenda.

To me this feels like another social control experiment I didn't provide informed consent for. And Facebook does this stuff all the time. I feel that anybody who wields that kind of power should have some kind of checks and balances. Right now, we have no way of knowing what their goals are.

If facebook's algorithm actually influences civil unrest... can we see it? can we discuss what it's doing, and if it's doing that well? When government policy does these things, we get to debate about it, vote (indirectly) on it.

Right now in 2020, Algorithmic Transparency is not a widely supported political issue.

But I think it should be.


what would that even look like?

here's my potshot at it:

I think that websites which display a newsfeed and have over 1 million registered users should be required to publish a whitepaper on their algorithm.


what do you think?

[minuspace]

Okay, it's just that with shit like back- propagation, hidden layers, and various compounded levels of convolution, I'd be surprised if anyone (persons) could actually make heads or tails of how it works.


Also, it's seems like the algorithm"s initial proof of work was the extent to which it could manipulate behavior. Given that directive, I think ITS understanding of democracy is probably more akin to crowd-control than it is to giving people a greater chance of/at autonomy.


Some algorithms may need to be redesigned, not to mention all the biases inherent to whatever datasets were used. Yeah, so adding data-set transparency is going to be sticky too.

[Cain]

I would go even further than that - I would argue that algorithms which promote extremist (in the sense of promoting violence) and misleading material in a quantifable fashion (looking at you, Facebook/Youtube/Instagram) lead to criminal charges and large fines, with the money from the latter going to fund local news organisations.

The information should definitely be made publically available, but there needs to be a stick as well. I've watched Qanon morph from "made up 8chan bullshit" to "modern day version of Aum Shinrikyo with a political party" in less than 3 years thanks to social media power and irresponsibility. That's no accident.

[POFP]

This is precisely why I've always been a proponent of CopyLeft/Open Source licenses on all Intellectual Property that the Public is dependent on. Not only are all the most hardened Algorithms for ANYTHING Open Source and available to the Public, but Intellectual Property that has large-scale impacts on Public Infrastructure (And if you think Social Media is NOT a part of Public Infrastructure at this point, you are empirically wrong. It is merely an abstraction layer on the Internet, and should be subject to the same scrutiny.) should ABSOLUTELY be available to the Public, at the very least via FOIA (Through the FCC).

Facebook can keep their Ad System Source, but something that dictates how content and social communication is programmatically served to the Public is absolutely fundamental to the nature of the Internet, and should be treated as part of it. This is why the most popular Search Engines should ALSO release their Source to the Public. They fundamentally alter the flow of information to Users, and therefore have a direct impact on how the Internet and its components are used.

[Cain]

Accidentally posted in here what I intended to post in the Trump hilarity thread, I have since split the discussion and merged it there

[P3nT4gR4m]

The problem with the algorithms at this stage in the proceedings is that most of them are utterly inconceivable to the human mind. The system has been writing itself for the best part of a decade now.

[minuspace]

That, and what if someone were to then I dunno, export them.

[Doktor Howl]

Very relevant

https://www.youtube.com/watch?v=dlXkTKxg4e4

[minuspace]

Very relevant

https://www.youtube.com/watch?v=dlXkTKxg4e4

Yup

[Cramulus]

100% yes

[Cramulus]

"A sheriff launched an algorithm to predict who might commit a crime. Dozens of people said they were harassed by deputies for no reason."

https://www.businessinsider.com/predictive-policing-algorithm-monitors-harasses-families-report-2020-9

But according to a six-month investigation published this week by the Tampa Bay Times, the high-tech tool deployed by the Pasco Sheriff's Office didn't lead to a reduction in violent crime — instead, 21 families singled out by the algorithm said they were routinely harassed by deputies, even when there was no evidence of a specific crime.

In September 2019, deputies showed up at 15-year-old Rio Wojtecki's door because the algorithm had determined Rio was one of the county's "Top 5" at risk of committing more crimes, the Tampa Bay Times reported.

Before that, Rio had been arrested only one time, a year prior, and was charged with sneaking into a carport and stealing motorized bicycles. Rio had already been assigned a juvenile probation officer — but because of the algorithm, police showed up at Rio's house to question him at least 21 times, beginning with that September visit, Rio's mother told the Tampa Bay Times.

People's criminal records — including charges that were later dropped — were fed into the algorithm to determine potential future offenders. Former employees of the sheriff's office said deputies were instructed to visit the homes of people the algorithm selected, charge them with zoning violations, and make arrests for any reason they could. Those violations and arrests were then fed back into the algorithm, according to the Tampa Bay Times.

[chaotic neutral observer]

People's criminal records — including charges that were later dropped — were fed into the algorithm to determine potential future offenders. Former employees of the sheriff's office said deputies were instructed to visit the homes of people the algorithm selected, charge them with zoning violations, and make arrests for any reason they could. Those violations and arrests were then fed back into the algorithm, according to the Tampa Bay Times.

Feeding activity generated by the algorithm back into the algorithm without qualification is just begging for a positive feedback loop.  It's like that horrible screech you get when the microphone is positioned too close to a speaker, only with cops.

In order for such an algorithm to be stable, it would be necessary for it to act only on independent activity by the "subjects" themselves, and not only cancel out the second-order effects introduced by the algorithm, but also correct for any biases in police activity (if they spend extra time watching a certain person, for whatever reason, that sampling bias needs to be accounted for).

So, to function correctly*, the algorithm would need exhaustive, detailed, and accurate data on the behaviour on each of the individual cops involved.

Heh.

*Assuming that "correct function" is actually predicting criminal activity, and not simply redirecting department funds to an IT contractor.

[Doktor Howl]

It's like that horrible screech you get when the microphone is positioned too close to a speaker, only with cops.

Newsfeed?

[chaotic neutral observer]

It's like that horrible screech you get when the microphone is positioned too close to a speaker, only with cops.

Newsfeed?

[altered]

It's like that horrible screech you get when the microphone is positioned too close to a speaker, only with cops.

Newsfeed?

Yeah, think so.