Gorilla Lobotomy

When large complex models give weird outputs, companies respond using some weird hacks to prevent repeating the mistakes

Ages back there was this controversy where Google Photos tagged black people as “gorillas”. This was a clear case of their algorithms being trained on biased (largely white people) samples, because of which they failed spectacularly when faced with an input that wasn’t in the training dataset.

There was predictable outrage, and rightly so in this case. What is interesting is Google’s response to this episode - rather than doing the right thing, which is using a representative and diverse sample of people to train its Google Photos people recognition tool, Google responded by simply “deleting gorillas”. Since gorillas as a category didn’t exist any more on Google Photos, you didn’t have the problem of classifying people as gorillas any more.

Of course, the side effect of this is that Google Photos is still unable to find and tag gorillas! Every once in a while, someone from the media picks this up and tries to get Google Photos to identify gorillas, and finds that it isn’t able to do so. The “hack” they used to get around their racist data set seems to have become permanent.

However, this response of google (to “delete gorillas” rather than train algo on a more representative data set) is emblematic of how some large corporations (which make the right kind of noises when it comes to diversity or “alignment”) deal with uncomfortable situations. Basically using hacks.

For example, recently someone found a bug on ChatGPT. If you asked it to repeat a word forever, the algo would end up exposing the training data. It’s a bizarre bug, but one that people spotted. This was duly posted on social media, and people started trying it.

Soon it was time for OpenAI (the makers of ChatGPT) to respond. How do you think they did it? Rather than changing the algo itself (which can be rather complex since it’s a bit of a black box), ChatGPT instead has marked all requests to repeat a word infinitely as “against its terms”.

I think as algorithms get more and more complex, exception handling will become more complex as well. And there is no choice for the builders of the algorithms to deal with the exceptional exceptions in a “normal” way, and they HAVE to use hacks. As the use of generative AI, and other large-scale unexplainable machine learning models, increases, one can only see the increase of such hacks.

In other words, Google’s lobotomy of gorillas is simply not going to go away. It will instead become a sustainable strategy.