14 August 2019

Things have changed in the last few years: Google Maps automatically displays the hotels that you’ve booked. Alexa, Siri, and Cortana tell you, in response to a spoken question, how much two plus four is or what the weather will be tomorrow. Amazon and Spotify tell you what to buy, or listen to, given your past purchases and favorites. Soon, we’re told, cars will drive themselves, because they’ll know the road and recognize and understand the traffic situation. And those big companies that don’t already will use artificial intelligence to pre-filter — or even choose — job candidates.

It’s all quite convenient. But the extent to which “artificial intelligence” is really “artificial” may be oversold. We’re finding more and more that human workers are hidden behind the friendly, “artificially intelligent seeming” interfaces of these systems. Earlier this year, Bloomberg and The Atlantic reported that human workers were “listen[ing] to audio clips recorded by Amazon Echo speakers.” On reflection, this is quite logical: Alexa’s speech recognition systems can’t consistently recognize speech in all of the world’s 6500 languages alone. They need to be constantly improved, and they need human help. Humans have to listen to some of the audio clips and transcribe them. According to the coverage, temp workers in Poland recruited by the temporary staffing agency Randstad were transcribing these clips; they were trained at an Amazon location in Danzig. After being trained, they worked from home. (The measures taken to protect the data in a work-from-home setting are unclear.) One of the temporary workers described the work in an interview with Welt am Sonntag as “ideal for housewives.” Some of the workers did the work on a laptop at their kitchen table while doing other tasks, like caring for children.

(Amazon has ended the program.)

The parallels with crowdwork — where workers are recruited via a digital platform — are clear to see. A study by Florian Schmidt published by the Hans Böckler Foundation revealed that a similar mass of not-particularly-well-paid “crowdworkers” working from home — all over the world — produce “training data” for the self-driving car programs of major automobile manufacturers. Qualified workers working full time via these platforms often earn between 1 and 2 US dollars per hour. The work “flows” all over the world, usually to the places where workers are willing to work for the smallest payment. In the case of Schmidt’s study of crowdsourced training data for the automotive industry, this turned out to be Venezuela, where a highly educated population with good internet connectivity was struck suddenly by political and economic crises and hyperinflation. The “crowdworker” colleagues in Venezuela have become part of a wave of what might be called “digital migrant workers.” Like migrant agricultural workers wandering from harvest to harvest for work, these crowdworkers migrate from platform to platform — wherever the tasks are.

(Schmidt’s study is available online in German; an English version is forthcoming. If you are interested in an English version, please email hello@fairtube.info.)

The results produced by these “artificially intelligent” systems are not necessarily better or more “objective” than results produced by humans. This was made extremely clear when Amazon used AI to sort applicants for technical jobs: because there were more men than women in the company, their system filtered out all the women who applied.

What about YouTube? On YouTube, decisions about recommendations and monetization are made by a combination of algorithms or “bots” and about 10,000 human “raters.” The raters are there because the algorithms are error-prone and can’t be completely trusted with these important decisions. The raters’ decisions also contribute to the ongoing “training” of the algorithms. Additionally, the European General Data Protection Regulation — specifically, Article 22 — provides that people have “the right not to be subject to a decision based solely on automated processing” that “produces legal effects” or “similarly significantly affects” them.

Of course, these decisions clearly “significantly affect” people — specifically, YouTube Creators. Whether or not GDPR provides Youtubers a legal right to talk to a person about these decisions that so unpredictably and dramatically affect their livelihoods is unclear. But common sense tells us that they absolutely have a moral right to do so.