Thread by @mmitchell_ai

I really love the active discussion abt the role of ethics in AI, spurred by Google Gemini's text-to-image launch & its relative lack of white representation. As one of the most experienced AI ethics people in the world (>4 years! ha), let me help explain what's going on a bit.

One of the critical pieces of operationalizing ethics in AI dev is to articulate *foreseeable use* (including misuse): Once the model we're thinking of building is deployed, how will people use it? And how can we design it to be as beneficial as possible in these contexts? 2/

This question overlaps with "Responsible AI" as well. Where ethics begs the question "should we build this at all?", once the decision at your company is YES, Responsible AI and Ethical AI share a lot of similarities. 3/

What you first do is figure out ways it's likely to be used, and by whom. This is harder for some people than others -- ppl with expertise in social & cog sci really shine at this (interdisciplinarity in tech hiring ftw!)
To help, I made this chart: Fill out the cells. 4/

Now here's the same chart filled out for text-to-image. The "green" cells are those *where beneficial AI is most likely possible*, not where it always will be beneficial. Within intended uses are "historic" depictions of reality -- such as white popes 5/

-- as well as "Dream world" depictions, where popes come in all shapes, colors, sizes. 6/

When designing a system in light of these foreseeable uses, you see that there are many use cases that should be accounted for:
- Historic depictions (what do popes tend to look like?)
- Diverse depictions (what could the world look like with less white supremacy?) 7/

Things go wrong when you treat all use cases as ONE use case, or don't model the use cases at all. 8/

That can mean, without an ethics/responsible AI-focused analysis of use cases in different context, you don't develop models "under the hood" that help to identify what the user is asking for (and whether that should be generated). 9/

We saw this same error in the generation of TSwift pornography: They forgot to have models "under the hood" that identify user requests that *should not* be generated. 10/zdnet.com/article/micros…

In Gemini, they erred towards the "dream world" approach, understanding that defaulting to the historic biases that the model learned would (minimally) result in massive public pushback. I explained how this could work technically here (gift link):
11/wapo.st/3OSmXo4

With an ethics or responsible AI approach to deployment -- I mean, the expert kind, not the PR kind -- you would leverage the fact that Gemini is a system, not just a single model, & build multiple classifiers given a user request. These can determine: 12/

1. Intent
2. Whether intent is ambiguous
3. Multiple potential responses given (1) & (2). E.g., Generate a few sets of images when the intent is ambiguous, telling user you're generating both the world *as the model learned it* and the world *as it could be* (Wording TBD). 13/

And further -- as is outlined in AI Safety, Responsible AI, AI ethics, etc., we're all in agreement on this AFAIK -- give the user a way to provide feedback as to their preferences (within bounds defined by the company's explicitly defined values). 14/

I think I've covered the basics. The high-level point is that it is possible to have technology that benefits users & minimizes harm to those most likely to be negatively affected. But you have to have experts that are good at doing this! 15/

And these people are often disempowered (or worse) in tech.
It doesn't have to be this way: We can have different paths for AI that empower the right people for what they're most qualified to help with. Where diverse perspectives are *sought out*, not shut down. 🩵 /end