So, a company sets up some service that they claim is automated and runs on 'AI.' Behind the scenes, they are running some imperfect and noisy data through an ML model; this ML model, being imperfectly reliable, is backed up by some large volume of invisible, precarious human labor.
At which point do we call this a mechanical turk? What's the hit rate the ML needs to achieve before we're no longer allowed to say said ML service is "just some guy?"
More importantly, why would we cede discursive power to these companies by going along with their assertion that their 'AI' is doing something? Why would we join them in asserting that the labor that backs up this service is in some sense being 'done by' an AI, rather than adopting the posture that humans are doing this labor in a way that is mediated through an ML system?
Of course, companies like Amazon do not divulge the success rate of their systems; they don't divulge how much human intervention is required. In the case of 'Just Walk Out', it sounds like rather a lot of human intervention was required.
But whether human review was being used in 90% of cases, 50% of cases, 10% of cases or 1% of cases: the reality is that it is absolutely fair and correct to say that this 'AI' was Just A Guy. Because ultimately this system would not have worked at all without the human labor; automation was not actually removing the need for human labor, it was just... automating the more obvious parts of the task.
I need people to stop being so reflexively afraid of being 'unfair' to fucking Amazon that they lose sight of the actual realities at play here. Fundamentally, an ML system was being used to render human labor obscure and invisible, in the interest of disinforming everyone involved so that this labor could be rendered precarious and disposable. Fundamentally, claims of automation were being exaggerated by a corporation with an interest in building this type of potemkin technology.
It is also important not to get dazzled by technical bullshit. "Ah, but those workers were merely 'training' the system." That is Amazon's claim. There are two problems with the way people are taking this claim at face value.
First, it is an unsupported technical claim that this 'training' would ever lead to a system that would function without human intervention. Are we meant to simply take Amazon's word for it on this?
Given that Amazon themselves seem to be abandoning this model, perhaps we shouldn't take their claims about its future capabilities (not even current capabilities! future!) at face value.
Furthermore, very clearly the human labor was required for the system to function satisfactorily at all in its present state. We don't have to countenance bullshit claims about future capabilities.
One way or another, to 'train' an ML model is still meaningfully to do the labor required to make that model function. Once again, Amazon's framing of the circumstances here is meant to make those workers disposable; they are an on-ramp to full AI management of the system. When we accept Amazon's framing and decide that we need to be 'fair' to them in this way, we aid them in their intent to use up these workers and then discard them.
Fundamentally, when an Amazon spokesperson says something, you don't have to immediately assume that they are lying but you do have to recognize that they do intend to mislead you. They are trying to enforce an epistemology of the world that is favorable to them and disfavorable to humanity. Please remember who you are hearing from and remember that they are not a good-faith participant in discourse.
Amazon's intent was absolutely to get people to think (investors, consumers, regulators, etc) that this was a closed loop with no human intervention, that they had automated away the need for grocery checkout workers. And that is simply not true, and it is right and fair to point out that it is not true. You are not 'correcting misinformation' by pointing out that the ML system obviated the need for direct human intervention in some number of cases, you are missing the point entirely.
So yes, Just Walk Out was literally Just A Guy; to shriek at people that they Don't Understand the Technology for saying so just makes you a sophist doing Jeff Bezo's work for him.
tl;dr I need people to stop thinking of 'AI' as a 'technology' and start thinking of it as what it is: a rhetorical strategy used to change how we perceive and talk about a related group of technologies.
One of the problems with the term is that I would argue that there are 3 kinds of technologies that count as "AI"
-
Variations on regression analysis. Basically, tools that optimize a model's error with some kind of known set of samples. In practice these basically all take some form of a matrix that you then do calculus on to get maxima or minima. None of this is new in terms of technique, which is why its uses tend to be either in established scientific analysis1, or in fields where its use is wildly inappropriate (chatgpt, etc.).
-
Secret outsourcing, in keeping with the even older tradition of the Mechanical Turk. The computer might do something but most of the magic is outsourced labor halfway across the country and none of the technology would scale if you were paying people in the supervisory roles a living wage.
-
It doesn't actually matter what the computer does as long as we can claim it's not actually our biases justifying our decisions. The Lavender stuff being used to classify enemy combatants in occupied Palestine, for example. The computer doesn't matter because a computer doesn't operate on ethical principles; they've already decided they'll ignore international law of armed conflict, so what does the computer need to do besides serve as a macguffin that always says whatever they were going to do anyway is a priori justified, because math?
-
I was once asked if it was possible to add AI to the analysis tools that I wrote for my job. I informed them that from my perspective that was what I was doing, as I had a set of known variables, an objective function (signal curves) and standard optimization libraries I was using to fit the one to the other. From my perspective the only thing that could make it more like the public perception of AI would be to come up with a way to aggregate the result of multiple data trials into a single set of values. Because our data is intended to measure changes in specific hardware over time, this isn't inherently a better technique for managing our data -- newer data is presumed to be more reflective of the state of the instrument, and due to the sheer volume of it under analysis, is unlikely to suffer issues of bias from noise that this sort of analysis would be useful for.
