Thread by @SonguelTolan

Excited that our paper: "Measuring the Occupational Impact of AI: Tasks, Cognitive Abilities and AI Benchmarks" has been published in Journal of Artificial Intelligence Research (@JAIR_Editor).

Here's a thread (1/10)
#EconTwitter #HUMAINT #futureofwork
doi.org/10.1613/jair.1…

The paper is the result of truly interdisciplinary work with sociologist Enrique Fernández-Macías, computer scientists @NandoMartinezP , @emiliagogu and José Hernández-Orallo, and economists Annarosa Pesole and me. Also, help from @CharisiVicky, HRI researcher. (2/10)

We develop a framework with which we can measure the exposure of occupations to AI. Note: we use the term "exposure" not "substitution" to avoid feeding into that fear-mongering narrative of "robots stealing our jobs". (3/10)

So, we map a list of 328 AI benchmarks (from a compilation of various papers, AI challenges and repositories, think paperswithcode, EFF, NLP-progress, SQuAD etc.) to data on 59 work-related tasks (from #PIAAC, #EWCS, and #ONET) through a layer of 14 cognitive abilities. (4/10)

Why not just map AI benchmarks to work tasks directly? Well, because a lot of AI systems are not programmed for specific work tasks but they "learn" abilities that lets them perform on different tasks. Think NLP that can help with reading/writing emails or customer support. (5/n)

We find: most AI exposure occurs through "thought-related" abilities: memorisation, text understanding, planning & search, learning & problem solving, also visual perception. Less exposure through social abilities, e.g. emotional control, social interaction or communication. 6/10

When we plot our AI exposure score against wage percentiles we find, surprise, a positive correlation: high-wage occupations (engineers, med. doctors) have a higher AI exposure score than low-wage occupations (care workers, cleaners & helpers, waiters & bartenders) (7/10)

Beyond technical feasibility, AI adoption depends on investments in complementarity infrastructure, restructuring of business processes, the relative costs of labour vs AI, aggregate demand. Then a bunch of other economic stuff happens when AI enters the market. (8/10)

We discuss this all in the paper. We also discuss the concern that if AI adoption increases wages, we might see an exacerbation of wage inequalities as AI diffuses because of this positive correlation between AI exposure and wages. (9/10)

But we are also excited to see what other researchers might do with our data on AI benchmarks and AI exposure scores. All data, code and results can be found on github.com/nandomp/AIlabo…
(10/10)

@EU_ScienceHub