On AI-driven Job Apocalypses and Collective Bargaining for Information

There has been another, fresh news cycle triggered by high-profile people in AI commenting on the potentially very disruptive economic impacts of AI. Of particular note, Dario Amodei, the CEO of Anthropic recently spoke very bluntly about his views on the potential impact of AI on jobs and the economy, inviting responses from many high-profile figures.

The Sedan Plowshare Crater. Wikimedia Commons.

Part 1: Links Round-up

First, let’s just review some links. This may be handy if you want to catch up on various interrelated new essays and news coverage (and I will update this, with a change log, if anything else comes up!)

There’s also been several widely circulated essays that are highly related:

And importantly, there’s also been pushback to the narrative of inevitability

Part 2: Adding Collective Bargaining for Information into the Conversation

What do I want to add to this conversation?

First, while I think the totality of above essays/articles/posts already provides a very comprehensive set of perspectives, considerations, and possible interventions, I of course want to reiterate the potential role of data leverage, and more specifically, “collective bargaining for information”:

Second, I want to provide another stab at a bullet point level analysis of the current evidence and the “mechanistic argument” for why economic power concentration is possible and even likely. I’ll also touch on how this prediction is compatible with “hype concerns” and the possibility that in some domains AI will face some data-related challenges.

I certainly think considering both theoretical and empirical work to understand the potential impact of AI on jobs and power will be critical. It’s also important to keep in mind that most people making predictions have wide bounds on their estimates right now (note Amodei’s “1-5 years” qualifier), and that there’s a split between people who want to focus more heavily on the data (the “it’s not something to freak out about until we actually see sectoral unemployment spiking” stance) vs. the theory (the “it’s something to freak out about because of the nature of information, cognitive labour, compute, power accumulation feedback loops” stance).

Part 3: Another simple model for thinking about AI impacts

Part 4: We should keep in mind the goal of the AI field and the plausbility of “augmenting AI”

So, large scale economic disruption from AI is possible. This disruption should be roughly forecastable based on task-specific and job-specific data availability (and it’s great to see more research on these topics, e.g. work from Labeskin et al. extending the “GPTs are GPTs” paper from Eloundou et al.), and there are some levers (see also a recent AI Now report that discusses, among other things, labour organizing for this purpose).

Part 5: In conclusion

Much of the AI field is focused on taking records of human work and creating compressed artifacts that can replicate the “output sequence of information” actions that workers must take regularly to maintain the leverage needed to keep their jobs.

One of the main goals of the field is to get better at replicating these sequences!

So if the field is successful, this will disrupt the economy.

AI might be limited in certain domains because of data availability (and in particular, will be limited in where it can be deployed because of eval data leverage).

But the core challenges in designing markets for information create conditions where powerful actors with existing capital needed to operate AI systems can create feedback loops to accumulate more information and build more powerful AI systems.

We should work to prevent this.

Recap of all links above:

Change log

1

While putting this together, I was repeatedly reminded of Jasmine Sun’s post on “The Aesthetic Genealogy of the Beige Tech Microsite”, which touches on all of the above, and feel called about recently reworking my personal site to use the Crimson Pro font…

2

Note this is simpler than the conceptual frameworks in other, more elaborate work: