Does AI inherently propogate bias and inequality? — A Code Like A Girl debate

Source: Artificial Intelligence on Medium

For the Negative

(i.e. arguing that, NO, AI techniques like Machine Learning and Deep Learning DO NOT inherently propagate biases and inequalities in the world)

Let me start by saying that I think we should regulate the use of AI techniques.

I say this upfront because a lot of people seem to think we need to argue that AI techniques are somehow inherently biased in order to justify regulating them at all.

However, there are plenty of things we regulate, without believing that they are somehow inherently bad for society. Chemicals are a good example. No-one thinks that industrial chemicals are inherently bad, but that doesn’t stop us regulating them very heavily.

In fact, almost any major technological advance is greeted with questions about whether it’s a threat to society in some way: in the 19th century it was chemistry, in the early twentieth century, cars were considered so dangerous that the law required a person to walk in front of the car with a flag to warn pedestrians of its approach.

So, yes, of course AI techniques can perpetuate biases, both accidentally and on purpose. And we should regulate them because of that risk, but they don’t inherently perpetuate biases or injustice. If they did, we couldn’t change whether they perpetuate biases, and we can. I’ll even argue they can help our societies overcome our biases and injustices.

Firstly, you’re blaming the wrong thing.

If you see machine learning and deep learning techniques as inherently perpetuating bias, you’re blaming the model or the neural network, when what happens on either side of it is far more important.

The two biggest sources of bias in any AI model are the historical data it’s trained on, on the one hand, and interactions it has with humans on the other.

Let’s start with the data:

A great example here is word embeddings — natural language processing models which encode relationships between words as vectors. Word embeddings are critical for AI systems to understand analogies — like apple is to skin as banana is to peel, or you’re the king of the castle. Word embeddings are trained on millions of articles (like Wikipedia dumps or Google news articles). Unfortunately word embeddings trained in this way usually exhibit gender biases, so a model trained on this data will infer that man is to doctor as woman is to nurse.

The problem isn’t that the model is flawed. On the contrary, the model is too accurate to the data it is trained on.

The problem is “garbage in, garbage out”.

The solution is, of course, not to throw out the technique that created the model. The solution is to de-bias the data.

A model built on de-biased data will not propagate those biases. There is nothing about the technique that created the model that is inherently biased. Use the same technique on unbiased data and the problem goes away. This is why, yesterday, Google removed gendered labels from its Cloud Vision API to ensure that its AI isn’t fed gender-biased data.

Which brings us interaction bias. This is the bias that is introduced when we let an AI system loose in the wild — and when the patterns of behaviour it finds in its interactions with humans re-shape the way it behaves. The classic example is Microsoft’s Tay, which was a Twitter bot that the company described as an experiment in “conversational understanding.” The more you chat with Tay, said Microsoft, the smarter it gets. Unfortunately, within 24hrs, Tay was talking like a racist, white supremacist. No one had encoded this behaviour in Tay, and it wasn’t trained on these kind of comments, but as soon as it was released, people set about trying to find out what offensive thing they could get it to say. Tay was trained to maximise its engagement, and the people it was interacting with rewarded it for being offensive.

As an aside: it’s interesting to realise that the story about biased data and the story about interaction bias are the same story. The only difference is “when” — biases in data are simply an historical record, traces of biased interactions in the past.

AI models are not born biased, they have biases thrust upon them.

They are trained to be biased, either before or after their release into the wild. Either way, if they are biased it’s because they are successfully mimicking past or present human behaviour.

And here’s the rub : biases in AI are not a bug in the program, they are an achievement. AI research stalled several times during the 20th century, because we could figure out a way of encoding ordinary human commonsense into an AI system. We discovered, somewhat paradoxically, that it was easier to teach a system to do university maths than to teach it things any child will intuitively know by the time they get to preschool.

The big advance of machine and deep learning techniques is to give up on encoding commonsense, and just to — as the name says — learn commonsense from exposure to volumes of experience, in the form of data or interactions.

The trouble is, when we train an AI system to mimic common sense it turns out we don’t like what we see. Artificial intelligence, as it stands today, is in fact, all too human.

The problem does not lie in the techniques of machine and deep learning, the problem lies in getting clear on what counts as commonsense — which we want to perpetuate — and what counts as bias — which we don’t.

And what happens when a society of people starts asking that question? Well, they find themselves wondering whether what they thought was common sense was really that common after all, or whether there aren’t hidden, implicit biases in interactions that otherwise seem perfectly normal.

A society that develops the desire to become fully conscious of its biases, more humble about its own intelligence, and with the right tools — tools like deep and machine learning techniques that can discover and expose implicit deeply held biases — can begin to de-bias it’s own behaviour, promote inclusive design and development, precisely to ensure that its future (and its AI) doesn’t simply mimic its past.

So, rather than vilifying AI (all the while — literally — projecting our own biases onto it), we should see this as an opportunity to discover and dampen those biases that are most invisible to us.