Valuation Pipelines in AI

I was invited to write a short public-facing piece about my research for the APA (American Philosophical Association) Blog, which you can check out here!

You can see me start to publicly reckon with the fact that in an important sense, my job as an AI ethicist isn’t really real.

Ethicists like me show up way too late to the conversation, when the metrics have already been chosen, the tools have already been developed, and the value conversation has already been shaped.

Here’s an ugly slide I made for a recent PowerPoint to try to convey just how out of the feedback loop AI ethicists and regulators are:

Not pictured: Me

Ok so your Funder might be Meta or the DOD, but either way they’re paying a bunch of eager technicians the salary and compute to get to work constructing whatever new AI models they’re asking for.

In turn, these Labs pay for their funding by churning out improved benchmarks and stunts to demonstrate that they’re making enough progress to merit more funding.

You see a so-so breakdancer, the DOD sees a future Terminator.

So Funders are paying AI Labs to do stuff they’d find useful, for either market leadership or (here’s the DOD term) American Preeminence.

And then the labs pay them in more capable models.

But why would we expect these models to be ethically bound in any way?

And where do these benchmarks even come from?

You can read more here and let me know what you think!

Anyway I gotta go, NBA playoffs are starting...