← Volver a la portada
Estados Unidos · Revistas de economía · Fortune

We lead MIT’s FutureTech program on the economics of AI: Don’t ask if AI can do the job, ask how much it costs

Publicado por Fortune · Estados Unidos hace 50 min
Damian Saccocio and Neil Thompson of MIT FutureTech.
Damian Saccocio and Neil Thompson of MIT FutureTech. · Fortune

Resumen detallado

Firms will hand work to AI when it is cheaper than a person, not when a demo says it can do the job.

Every week seems to bring a major new advance in AI’s ability to do heretofore human-only work, heightening concerns of large-scale and rapid job loss for knowledge workers. Our research at MIT FutureTech suggests the reality is likely to be much more nuanced, with cost, not just capabilities, playing a major role.

Our study of computer vision made this case first. Most research on “AI exposure” measures which tasks AI could potentially perform, focusing on where the technology might develop. That’s a far cry from what we studied: whether firms would actually choose to automate those tasks once they had to pay to build, deploy and maintain the AI system. Across about 1,000 vision tasks in roughly 800 U.S. occupations, only 23% of the wages paid for those tasks were worth automating at the costs of the time.

The reason is fixed cost. A system costs about the same to build whether it replaces an hour of one worker’s day or the shifts of a hundred workers. Consider a bakery that wants to use computer vision to check for ingredient spoilage. If you are the Wonder Bread factory, producing 70+ million loaves of bread per year, it might make sense to automate this task. If you are a corner bakery, any reasonable amount of upfront fixed costs will quickly eclipse your labor savings. So the factory automates and the bakery keeps its bakers.

Generative AI, the AI powering large language models, is different. A general-purpose model sold by the token is far cheaper to start using than a custom vision system, so the fixed-cost barrier is lower. Our newer research asks how good that work is. Experienced workers graded AI output on more than 6,000 text-based tasks from the Labor Department’s occupational database, over 60,000 evaluations in all. The most capable models produced work usable as-is about two-thirds of the time.

That is impressive, but a two-thirds success rate can hide two different realities: AI may handle some tasks well every time and others poorly, or succeed at a…

O…

Inicio del artículo, extracto de Fortune. Para leer el resto, abra el artículo en el sitio del medio.

Leer el artículo completo en Fortune Inicie sesión para guardar

MarketNews muestra el titular y el resumen publicados por el medio. El artículo completo se lee en el sitio del editor. (fortune.com)

También te puede interesar