This week, an AI summit took place in Paris - with no real consequences. The event felt like a big promo show for France and its unpopular president. The outcome is a one-and-a-half pager with nothing of substance in it, and which major players (including the US and the UK) did not sign.
Even Claude agrees: “While the document brings together a broad coalition of over 100 countries and organizations, most of its content consists of general statements about AI governance principles that have been expressed in previous international declarations.” (emphasis added). In other words, many people signed something that has already been agreed to a hundred times before.
This would be funny, if we weren’t living through one of the most radically transformative periods in human history.
This week was also the week of the 9th edition of AMLD - the Applied Machine Learning Days - at EPFL, which I help organize. It just finished a few hours ago, and thus I can’t report in detail what I saw, but at least now you know why this weekly summary will be relatively short. 😅
What Claude is used for
One thing I did find intriguing this week though was that Anthropic, the maker of Claude, released a fascinating report on how Claude is being used. Here are a few summary statistics:
Software development and writing tasks dominate current AI usage and account for nearly half of all interactions
~36% of occupations are using AI for at least a quarter of their tasks, showing strong penetration across the workforce
Peak AI usage occurs in mid-to-high wage occupations requiring bachelor's degrees, with lower usage in both very high-wage (eg physicians) and low-wage positions
57% of AI interactions show patterns of augmentation (humans collaborating with AI) while 43% demonstrate automation (AI completing tasks directly)
So, rather than wholesale automation of jobs, we’re seeing selective adoption of AI for specific tasks within occupations. This makes a lot of sense. I think it’s useful to think of a job as a set of tasks and to recognize that AI - or automation more generally - can replace tasks. The question of job evolution then hinges on whether a job involves many tasks and whether it can easily take on new one.
Take the job of a phone operator in a call center. It consists of a single task: helping callers with their problems. Once AI can automate that, no tasks remain, and we don’t typically consider additional responsibilities for phone operators. This is why these jobs will entirely disappear.
The situation is completely different for, say, a head teacher. Their responsibilities are far broader: planning and coordinating lessons, managing the budget, making personnel decisions, communicating with parents, resolving conflicts, developing the school’s pedagogical focus, complying with legal regulations, representing the school, supervising and evaluating teachers, organizing events, working with authorities, overseeing school infrastructure, and much more. Not only are these tasks difficult to automate, but even if they were, many new ones would likely emerge.
But back to the report from Anthropic, which includes the following section:
Automation versus augmentation. We also looked in more detail at how the tasks were being performed—specifically, at which tasks involved “automation” (where AI directly performs tasks such as formatting a document) versus “augmentation” (where AI collaborates with a user to perform a task).
Overall, we saw a slight lean towards augmentation, with 57% of tasks being augmented and 43% of tasks being automated. That is, in just over half of cases, AI was not being used to replace people doing tasks, but instead worked with them, engaging in tasks like validation (e.g., double-checking the user’s work), learning (e.g., helping the user acquire new knowledge and skills), and task iteration (e.g., helping the user brainstorm or otherwise doing repeated, generative tasks).
This is an insightful result. Of course, we should first remind ourselves that, for the vast majority of tasks, AI isn’t used at all. But even when it is, a slight majority of cases involve augmenting the human in performing the task rather than fully automating it.
In other words, jobs are more likely to evolve than disappear entirely - especially those that can easily take on new tasks. At last, we have some empirical data to better understand AI’s real-world impact on work.
CODA
This is a newsletter with two subscription types. I highly recommend to switch to the paid version. While all content will remain free, all financial support directly funds EPFL AI Center activities.
Podcasting: I’m hosting an AI podcast at the EPFL AI Center called “Inside AI” (Apple Podcasts, Spotify), where I have the privilege to talk to people who are much smarter than me.
