Amazon helps and teaches bomb makers

Amazon’s recommender algorithm works pretty well: if people start to gather together ingredients needed for making a thermite bomb, Amazon helpfully suggests other items that may be needed to make it, including hardware like ball bearings, switches, and battery cables. What a great teacher!

It is disturbing that this seems to imply that there are enough people ordering such things for the algorithm to recognize a pattern. However, it would seem remarkably dumb for a determined terrorist to leave such a (figuratively and literally) blazing trail behind them, so it is just as likely to be the result of a very slightly milder form of idiot, perhaps a few Trump voters playing in their backyards. It’s a bit worrying, though, that the ‘wisdom’ of the crowd might suggest uses of and improvements to some stupid kids’ already dangerous backyard experiments that could make them way more risky, and potentially deadly.

Building intelligent systems is not too hard, as long as the activity demanding intelligence can be isolated and kept within a limited context or problem domain. Computers can beat any human at Go, Chess, or Checkers. They can drive cars more safely and more efficiently than people (as long as there are not too many surprises or ethical dilemmas to overcome, and as long as no one tries deliberately to fool them). In conversation, as long as the human conversant keeps within a pre-specified realm of expertise, they can pass the Turing Test. They are even remarkably much better than humans at identifying, from a picture, whether someone is gay or not. But it is really hard to make them wise. This latest fracas is essentially a species of the same problem as that reported last week of Facebook offering adverts targeted at haters of Jews. It’s crowd-based intelligence, without the wisdom to discern the meaning and value of what the crowd (along with the algorithm) chooses. Crowds (more accurately, collectives) are never wise: they can be smart, they can be intelligent, they can be ignorant, they can be foolish, they can even (with a really smart algorithm to assist) be (or at least do) good; but they cannot be wise. Nor can AIs that use them.

Human wisdom is a result of growing up as a human being, with human needs, desires, and interests, in a human society, with all the complexity, purpose, meaning, and value that it entails. An AI that can even come close to that is at best decades away, and may never be possible, at least not at scale, because computers are not people: they will always be treated differently, and have different needs (there’s an interesting question to explore as to whether they can evolve a different kind of machine-oriented wisdom, but let’s not go there – SkyNet beckons!). We do need to be working on artificial wisdom, to complement artificial intelligence, but we are not even close yet. Right now, we need to be involving people in such things to a much greater extent: we need to build systems that informate, that enhance our capabilities as human beings, rather than that automate and diminish them. It might not be a bad idea, for instance, for Amazon’s algorithms to learn to report things like this to real human beings (though there are big risks of error, reinforcement of bias, and some fuzzy boundaries of acceptability that it is way too easy to cross) but it would definitely be a terrible idea for Amazon to preemptively automate prevention of such recommendations.

There are lessons here for those working in the field of learning analytics, especially those that are trying to take the results in order to automate the learning process, like Knewton and its kin. Learning, and that subset of learning that is addressed in the field of education in particular, is about living in a human society, integrating complex ideas, skills, values, and practices in a world full of other people, all of them unique and important. It’s not about learning to do, it’s about learning to be. Some parts of teaching can be automated, for sure, just as shopping for bomb parts can be automated. But those are not the parts that do the most good, and they should be part of a rich, social education, not of a closed, value-free system.

Address of the bookmark:

Original page


Update: it turns out that the algorithm was basing its recommendations on things used by science teachers and people that like to make homemade fireworks, so this is nothing like as sinister as it at first seemed. Nonetheless, the point still stands. Collective stupidity is just as probable as collective intelligence, possibly more so, and wisdom can never be expected from an algorithm, no matter how sophisticated.

Bigotry and learning analytics

Unsurprisingly, when you use averages to make decisions about actions concerning individual people, they reinforce biases. This is exactly the basis of bigotry, racism, sexism and a host of other well-known evils, so programming such bias into analytics software is beyond a bad idea. This article describes how algorithmic systems are used to help make decisions about things like bail and sentencing in courts. Though race is not explicitly taken into account, correlates like poverty and acquaintance with people that have police records are included. In a perfectly vicious circle, the system reinforces biases over time. To make matters worse, this particular system uses secret algorithms, so there is no accountability and not much of a feedback loop to improve them if they are in error.

This matters to educators because this is very similar to what much learning analytics does too (there are exceptions, especially when used solely for research purposes). It looks at past activity, however that is measured, compares it to more or less discriminatory averages or similar aggregates of other learners’ past activity, and then attempts to guide future behaviour of individuals (teachers or students) based on the differences. This latter step is where things can go badly wrong, but there would be little point in doing it otherwise. The better examples inform rather than adapt, allowing a human intermediary to make decisions, but that’s exactly what the algorithmic risk assessment described in the article does too and it is just as risky. The worst examples attempt to directly guide learners, sometimes adapting content to suit their perceived needs. This is a terribly dangerous idea.

Address of the bookmark: