Whose value is it anyways?

This is a piece on ML and AI that I wrote for a class in January 2024.

The need to be responsible with AI

This work refers to class assignment that tasked us with making our own models using the CelebA dataset, and then extracting and using feature vectors from the trained model to add and remove features from our own photos.

There are, of course, moral implications of such a process. I could have used photos without the permission of others, and created more heinous images. Such misuse is possible with many tools, physical or otherwise. A hammer can be used to harm someone, so can the threat of a legal case to someone unable to afford a lawyer. In both cases, we as society, recognise that to be wrong. That is why using a hammer to hurt someone is a crime, and public defenders exist. Similarly, I am of the opinion (and hope), that when (not if) generative AI becomes more mainstream, we will have adapted to it with regards to these moral implications.

A deeper, more insidious problem according to me is illustrated in the following. I asked the model to make me more “Attractive”, and it made it more White, whereas asking it to make me less “Attractive”, drove it to make me more Black. Clearly that is wrong, and extremely racist. However, I would like to think I personally am not racist, and neither is any of the code that I wrote. How then, is this possible?

Clearly in this case, it was the data labelling process. Whoever compiled the data, thought White people are more attractive than Black people. Again, this is clearly racist. It raises the question of ensuring morality and objectivity in the data. However, how do we define morality, and for a whole class of data, does objectivity even exist?

The talk of morality in AI inevitably leads to talk of “AI alignment”, that is, talk of aligning AI values with that of humans. This is an important question for now, but gets far more so when we talk of some hyperbolic superhuman singularity AGI, whose actions are beyond our understanding. Regardless, in this case, what are the values? I hold the value that skin colour should necessarily be unrelated to attractiveness. If I were to label such a dataset, I would want to ensure that across different skin tones, attractiveness rates remain constant. The creators of the CelebA dataset evidently had different values. Are my values “better” than theirs? Are theirs more accepted? I think my values are correct, and should be what everyone believes, but, who gets to decide which values an AI should embody?

An argument goes that AI should not have values, instead, it should reflect society. That is what Microsoft’s Tay did, and it had to be brought offline quite quickly. A reflection of society (especially through data collected online) is often not a reflection of values that society wants to hold itself to. It is more closely analogous to an ugly mirror that highlights our worst, deepest flaws.

The mistakes, knowable and unknowable

In the attractiveness question, we caught the racist attitudes of our model, but what if I cannot? I am not culturally American, and I use AI to draft my emails. Am I communicating well, or am I committing cultural faux pas without even realising? What happens when we cannot catch the mistakes our models make? And even more importantly, who holds the blame in that case? Is it me, who did not realise I was committing a faux pas, the developer who had no idea what was buried in line 174280 of file 48290 that the model really liked, or the user who posted that on reddit with no intention of it ever being someone else’s words?

A recent use case of Generative AI I came across was synthetic data generation in environments with low positive events (insurance underwriting for lower income groups). In that case, while using Generative AI might well spur more companies to give out loans, leading to more loans for people who need them, would it also accidentally be pushing that women are less trustworthy buyers, leading to fewer women getting loans than men. On one hand, more women are getting loans than before, but on the other, we are continuing, and more importantly, codifying human biases into data. There are myriad examples of this, from hiring to judicial decision making, but these are only the ones we catch. What if we don’t catch them? What if we keep using some set of values with extreme bias towards, say, sending people of colour to jail, as we think it is right and correct, after all, the data says that they are more likely to be in prison anyways!

Are we the baddies?

I am an electrical engineer turned social scientist, and my goal is to use data to tell stories about people. That involves telling the truth about what the data says (yes, people of colour go to prison more often), but also involves investigating WHY the data is saying such things (because of a history of racial inequity, and systemic disinvestments, along with bias in judicial decision making).

LLMs and Generative AIs are great at coming up with plausible sounding explanations for what the data says, after all, that is what they are designed to do. Yet, they often fail at coming up with deeper analysis of why something might be true, not because they are not capable, but because WE have not told them yet that that is a possibility.

All of this to say, the AI alignment question is not hard because AI is hard (which it is), it is because we, humans, you and I, are complicated. The talk and discussions of what it means to be an AI is more deeply, a conversation of what it means to be human. When robots were conceptualised, humans long thought they should look and feel and behave like we do, however we moved away from that. We have extremely powerful robots now that are highly specialised, that do their tasks really well, yet they are not human-like in any discernible way.

Why then, should AI be like humans?

In Conclusion

Prof. Dave McAllester posits that the future of AI will be everyone having their own tiny “Advobots”, making the case for whoever constituted their personal Constitutional AIs. Maybe that is the way to go, no one set of values, and letting everyone choose what they want. Maybe that is too much for society to handle, and we will want some guardrails, regulations, and protections. However humanity does go about this, it will say more about humans, than it will about Artificial Intelligence.