Anyone who has spent a few minutes playing with ChatGPT will understand the worries and hopes such technology generates when it comes to white-collar work. The chatbot is able to answer all manner of queries—from coding problems to legal conundrums to historical questions—with remarkable eloquence.
Assuming companies can overcome the problematic way these models tend to “hallucinate” incorrect information, it isn’t hard to imagine they might step in for customer support agents, legal clerks, or history tutors. Such expectations are fueled by studies and media reports claiming that ChatGPT can get a passing grade on some legal, medical, and business exams. With companies like Microsoft, Slack, and Salesforce adding ChatGPT or similar AI tools to their products, we are likely to see the impact on office life soon enough.
A couple of research papers posted online this week suggest that ChatGPT and similar chatbots may be very disruptive—but not necessarily in the ways you expect.
First up, Edward Felton and colleagues at Princeton University try to identify the professions most likely to be affected by ChatGPT. They used a benchmark called the AI Occupational Exposure, which maps occupational tasks to the capabilities of various AI programs, to see which jobs are most vulnerable to chatbots with strong language skills.
The results suggest big changes may be ahead for those in some professions, including telemarketers, history teachers, and sociologists, while people with more physical occupations, such as brickmasons, dancers, and textile workers, may not need to worry about ChatGPT showing up at their place of work.
But a second study suggests that people in language-centric jobs are not necessarily destined for replacement. Shakked Noy and Whitney Zhang, graduate students at MIT, looked at what happens when you put ChatGPT in the hands of office workers. They asked 444 college-educated professionals to complete a series of simple office tasks, including writing press releases and short reports, drafting emails, and creating analysis plans. Half of them got to use ChatGPT.
The study found that people with access to the chatbot were able to complete the assigned tasks in 17 minutes, compared to an average 27 minutes for those without the bot, and that the quality of their work improved significantly. Participants who used ChatGPT also reported higher satisfaction with their work. Although the study involved asking experts to judge the quality of participants’ work, the paper does not say if that included looking for the types of “hallucinated” errors that can creep into ChatGPT’s output.
These two studies hint at ways things might play out, but they are just early (and not yet peer-reviewed) attempts to figure out where ChatGPT is taking us. It’s notoriously difficult to predict how new technology will impact work, and economic research related to ChatGPT is appearing rapidly.
It’s also ironic that textile workers were identified as potentially immune to ChatGPT, since those who fret about the impact of AI on labor are sometimes branded as Luddites, in reference to the nineteenth-century movement in which English textile workers smashed looms to protest automation.
In fact, Luddites were by some accounts, more worried about who controls automation than the existence of automation itself, aiming their rage at employers who used automation to avoid paying workers fairly.
It might be a good idea for workers to take the initiative and start using ChatGPT to make themselves more productive. Just don’t tell my boss, OK? (I’m joking—WIRED just published a new policy on use of generative AI that says we won’t publish AI-generated text except where it’s part of a story).
My first attempt at automating my own work was a false start. When I asked ChatGPT to find some links for this week’s newsletter, it suggested a bunch of stories from 2021, which makes sense when you remember that the AI model was trained on data taken from the web some time ago 🙄. Newsletter writers may not see a big boost to their productivity just yet.