Francois Chollet, the deep learning expert who created Keras, put it neatly in a Tweet:

“I'm so old I remember when fully autonomous cars were going to be ready for mass deployment by late 2017”  

Autonomous vehicles are one of the best examples of the tendency for technology to over-promise and under-deliver, and I’d put customer service chatbots in a similar category. The tools that we call AI, for now and the foreseeable future, can be extremely good at performing very specific tasks, but they don’t think. There is still nothing close to the AI “general intelligence” you might see in science fiction (the closest thing I know of is GPT-3), nor even a consensus on how (or if) building one might be possible.

The core problem, she argues, is what she calls “technochauvinism”—the belief that technology is always the solution to every problem. This manifests in the regular spectacle of silicon valley entrepreneurs “inventing” products that have existed for years, such as “reusable tissues”.

That’s quite funny, and not doing anyone any harm, but the same outlook applied to machine learning approaches to problems that have a real impact in society can be much more damaging. If you believe, for instance, that AI is a better way to diagnose disease, or make decisions about early release of criminals, or to sift job applications.

All AI tools require data, usually mountains of data. Where does it come from?

“…data always comes down to people counting things…data is socially constructed.”

This is often overlooked, but enormously important. First of all, it means that technochauvinists tend to prioritise things which are relatively easy to measure. It’s very difficult to measure quality, for instance, but very easy to measure popularity. To most of us it’s fairly obvious that there’s a distinction between those two things, although perhaps we’d be hard put to define exactly what we mean by “quality”.

In practice it’s very common for AI applications to treat popular as a synonym for good, such as the app which promised to rate your photos for quality, but ended up rating them based on the extent to which they resembled an attractive 20-something white woman.

The other point about all that data gathered up by humans, is the amount of work that goes into it. Where the data exists, great, why not make use of it. But don’t pretend that machine learning can operate in a vacuum without all the human-generated data. As Broussard comments on the headline-grabbing AlphaGo algorithm:

“Millions of hours of human labor went into creating the training data – yet most versions of the AphaGo story focus on the magic of the algorithms, not the humans who invisibly and over the course of years worked (without compensation) to create the training data.”

We’re in such a hurry to heap praise onto the robots that we sometimes forget to give ourselves enough credit. Broussard gives the example of a tool as everyday as Google, which to a large extent works as well as it does because we have learned how to use it well. Googling effectively is a skill, and that’s a really good example of how the best technology solutions come about from fusing together the complementary skills of humans and computers.

Very much the same principle is likely to apply to the world of autonomous vehicles, which (as Francois Chollet alluded to) seem in many ways as far away as ever.

The Hollywood vision of AI coupled with technochauvinism has led to the rushed deployment of machine learning approaches, launched with hyperbolic claims, that are simply not delivering.

“…we are so enthusiastic about using technology for everything…that we stopped demanding that our new technology is good.”

That’s not to say that AI doesn’t have potential, it does, but it makes a lot more sense to see it as a tool that humans can use, rather than as an autonomous agent that can step in to replace human decision-making in most cases.