New Toronto Declaration Calls On Algorithms To Respect Human Rights

A coalition of human rights and technology groups released a new declaration on machine learning standards, calling on both governments and tech companies to ensure that algorithms respect basic principles of equality and non-discrimination. The Verge reports: Called The Toronto Declaration, the document focuses on the obligation to prevent machine learning systems from discriminating, and in some cases violating, existing human rights law. The declaration was announced as part of the RightsCon conference, an annual gathering of digital and human rights groups. "We must keep our focus on how these technologies will affect individual human beings and human rights," the preamble reads. "In a world of machine learning systems, who will bear accountability for harming human rights?" The declaration has already been signed by Amnesty International, Access Now, Human Rights Watch, and the Wikimedia Foundation. More signatories are expected in the weeks to come.

Beyond general non-discrimination practices, the declaration focuses on the individual right to remedy when algorithmic discrimination does occur. "This may include, for example, creating clear, independent, and visible processes for redress following adverse individual or societal effects," the declaration suggests, "[and making decisions] subject to accessible and effective appeal and judicial review."

Comments

Nice and all, but simplistic and naive

[Rosco P. Coltrane]

Okay, here's one: suppose an AI is in charge of running the US (and after all, why not: humans aren't doing such a great job at it). It determines the only way to end violence and protect minorities is to preventively lock up religious crazies and racists. Whose basic human rights will it choose to respect? The right of the crazies to be free, or the right of the minorities to be treated as human beings and live in peace? Zeroth law anyone?

Re:Bias in - Bias out.

[The Cynical Critic]

The first example does sound like it was just doing it's job seeing how black people do to my knowledge commit a disproportionately large portion of the kinds of crimes that have a relatively high rate of recidivism (rapes, peddling drugs, gang violence, etc.). Any correctly working system would naturally end up looking like it's biased against black people even if it's not given the defendants' race or even capable of even understanding the concept. It's basically the same "issue" as how a system that's supposed to assess the risk of hockey-related injuries, say for determining insurance rates, would determine that white men are more at risk of them and act accordingly.

As for the second example, if you actually look at employment statistics you are going to see men being the clear majority of those working in high stress, high risk, physically demanding and very high salary jobs, which tend to require skills men are more likely to have. Similarly to the recidivism system, determining of who gets shown what ads is not actually based on the supposedly discriminating characteristic, but various peripheral characteristics that end up giving the illusion of discrimination. Thus it's again really not surprising to see a system meant to display ads to those most interested in what they're advertising show these kinds of jobs to men more often than women and changing the system to display more "male-centric" ads to women is merely making the system less accurate.

Finally, the third and fourth examples are just examples of a maliciously and incompetently coded system respectively and not really relevant here. An "ethically" set up machine learning system could be just as flawed if not worse for very similar reasons. I can see why people would get upset over the first and second examples if they didn't understand how systems like them actually work or the data they work based on, but "non-racist" and "non-sexist" version of the system would actually have to be explicitly racist and factor in a bonus when computing the risk of recidivism for black people and explicitly sexist and factor in gender when determining what ads to show rather than just skills and interests.