What they did: They looked at the impact of the release of ChatGPT on freelancers on Upwork.They found strong indications that it lowered freelance employment and wages, and that high-quality workers might have even been hit worse by it.
Some comments and caveats: Freelance work of this sort is not only a plausible early indicator (having by nature a much faster turnaround than formal employment) but it's also likely to be specially vulnerable to this sort of AI technologies (always keep in mind that 2023-style generative AI is just a subset of generative AI, and generative AI is just a subset of AI):
- Freelance work is usually weakly coupled to company specifics: Freelance workers have by definition little knowledge of the specifics of a company's culture, processes, market, and proprietary technology. Much like general models, they are familiar with the generalities of a discipline, not with the particularities of the organization.
- Freelance work on these platforms is rarely highly-specialized. Very specialized work usually have their own formal and informal recruiting networks with deep information and reputation systems that are poorly or not at all replicated by generic platforms, so neither employers nor freelancers seek work there.
- Freelance work on these platforms is rarely at the far tail of the skill curve. This doesn't mean that there aren't highly skilled workers there, but there's a clear trend for people, as they gain experience and reputation, to either develop more stable work relationships or to build reputation networks that channel opportunities away from "flat" highly competitive markets.
In a way, AI generative models are a good competitive fit against freelance workers in Upwork-style platforms precisely because they are the underlying archetype of what those platforms try to provide to employers: low cost scalable commodity expertise on demand. And the vulnerability of freelancers to these substitution is just another iteration of what were already long-term eroding wages; anything that makes you searchable makes you fungible, and if you're fungible you're sooner or later cheap.
On the other hands, companies that takes this as good news should be mindful that low cost scalable commodity expertise on demand gives the advantage to incumbents and newcomers with incumbent-sized wallets. Catching up and winning starting from a smaller position and less capital requires out-skilling (apologies for the term) competitors, which by definition means anything a generative AI can do is far below the minimum you need to aim for if you want to be able to survive against competitors that can deploy the same AIs at a much larger scale and with much larger ancillary resources and networks.