Many of us have probably heard the call for ethical AI, but what is ethical AI exactly?

Ethical AI applyies an ethical framework to Artificial Intelligence, which is to say to apply ethics to the machine learning model.

Constitutional AI is a potential solution to ethical AI. The challenge is that all ethical models are flawed. Constitutional AI suggests a rule set that is wrapped around the base functional model. Product examples of this are Claude and Anthropic, which is supported by Google. OpenAI, which relies on human governance, the basis of ChatGPT, is supported by Microsoft.

Each of these has inherent challenges. We’ve all likely heard of the systematic bias inherent in the data used by large language models. OpenAI uses human governance to adjust and minimize the bias in these models. However, this can lead to hypercorrection and introduces different human biases. Moreover, this leads to situations where queries are refused by the model because human governance has determined the outputs to be out of bounds.

Constitutional AI on the other hand has underlying ethics explicitly built into the model under the auspices of harm reduction. The problem I have with this is twofold: The first is fundamental. Constitutional AI is based on the deontological morality principles elaborated by Kant. I’ll come back to this. The second is empirical.

Many of us are of the age to recall when Google’s motto was to do no evil. When they decided they could not follow their own dictate mom they simply abandoned the directive. Why should we expect a different behaviour this time around?

Moreover, harm is a relative concept so to minimize harm of one group may be to increase harm in another. This undermines the deontological intent and is of larger concern.

As a moral relativist and subjectivist, I find this to be categorically problematic. It poses even more problems as a moral noncognitivist.

From the relativist’s perspective, our AI is fundamentally guided by western white guys with western white-guy sentiment and biases. Sure, there are token representations of other groups, but by and large they are marginalised and the aggregated are still dominated by western white guys.

DISCLAIMER: It is still difficult for me to input or edit copy into a computer, so this may be more ragged than usual. I may return to amend or extend it as I see fit.