Algorithms Are Not Oracles: Anyone Can Train Them to Say What They Want

Taking advantage of the global obsession with ChatGPT, a plethora of companies are now offering algorithms for just about any task, and all abundantly funded. Over the last week, more than two hundred.

In response, some people are trying to jailbreak these algorithms to circumvent their programmed restrictions and understand their biases and conditioning factors, highlighting how important it is for these types of tools to be open source, so that we can adapt them to our own needs: Microsoft might believe that ChatGPT-4 has artificial general intelligence or AGI, but algorithms do not think for themselves; they are the result of whatever they have been trained with, which means bias can be built in. US conservatives accuse ChatGPT of being liberal, and are now trying to build more reactionary chatbots.

Whether you use an algorithm to write restaurant reviews without having tasted the menu, write software or create a virtual boyfriend, biases of any kind will be there, built in consciously or unconsciously, depending on the information the algorithm was trained with and your skills when writing your prompts. Competition is ferocious: Microsoft intends to prevent other algorithms from feeding on the data generated by its new version of Bing, while Google is launching its Bard in closed beta (if you want to request access, use your VPN and tell it you are in the UK or the US) and comparisons between the two are appearing, while Sam Altman says on Forbes that AI is one of the forces behind the breakdown of capitalism. You might want to take that with a pinch of salt…

This is where we see the importance of open source: either we can access algorithms and data repositories through the back door and see on what basis they have been trained, or we will be working with tools that are dangerous by nature, capable of generating all sorts of biases without us necessarily being aware of it. For some reason, whether through lack of experience or reverential fear, the human mind tends to attribute machines with impartiality or even infallibility: if the algorithm said so, it must correct. Nothing could be further from the truth. We must do everything to avoid a future where the algorithms we use for more and more things depend on a few companies, and the antidote to this is either transparency or the possibility of building and training our own algorithms.

Let’s start thinking about it.

—

This post was previously published on Enrique Dans’ blog.

***

You Might Also Like These From The Good Men Project

You Said ‘Race’, but Are You Actually Talking About Race?

Understanding the Nonbinary: Are You Confusing Gender With Sex?

The Difference Between Compassion for Those With Disabilities & Ableism?

‘Masculinity’ Is Having an Identity Crisis