Study: Scientists, researchers and technologists most at risk from LLMs?
Researchers from OpenAI and the Wharton School have investigated how high the labor market impact of LLMs could be. They are affected themselves.
(Image: Bild erstellt mit KI in Bing Designer durch heise online / dmk)
The reaction of the working population to large-scale language models, which have been ubiquitous since the release of ChatGPT in November 2022 at the latest, is extremely mixed. There are those who are happy that AI systems can seemingly take over routine tasks. Others fear the collapse of existing employment relationships, particularly in the area that was previously considered safe: jobs behind a desk that usually require a degree of education.
Computer colleague, take over!
The truth probably lies somewhere in between. This is the result of a study published in the journal Science by Tyna Eloundou, a member of the OpenAI Technical Staff, together with colleagues from the ChatGPT manufacturer and researchers from the Wharton School at the University of Pennsylvania. They created a framework that could be used to assess the potential impact of large language models (LLMs) on the world of work.
They assume that LLMs currently still have a rather small impact. With a total of 1.8 percent of all jobs, it is conceivable that at least half of these people's tasks would at least be "affected" by LLMs, provided they retain their current rather simple interface and non-specialized training data. But it won't stop there. "If you look at current and likely future
developments in software that complement LLM capabilities, this proportion rises to 46 percent of all workstations." It is high time to carry out "robust societal evaluations" - including political measures to address these potential effects.
80 percent of current employees affected
Based on a total of 923 popular jobs taken from the O*NET-27.2 occupation database and described therein along with their content, it was shown that the use of an LLM at the level of a GPT-4 would reduce the required working time by at least half. At the same time, the quality allegedly does not suffer or even improves.
Estimates made by the researchers as part of their framework assume that for 80 percent of all employees, at least 10 percent of their tasks could also be performed or improved by LLMs. 18.5 percent of today's employees would be faced with an LLM takeover rate of 50 percent if their employers were to come up with the idea of implementing this. There are three employee groups that are particularly affected - and they are those that are seen as particularly promising for the future: "scientists and researchers" and "technologists".
Crafts with a golden bottom?
It doesn't have to stop there. Those who retire to the skilled trades, for example, could also be affected by advances in robotics. "Tasks that are currently considered unattainable may become feasible thanks to future innovations," write the researchers. Conversely, however, it could also happen that areas that seem ripe for disruption "encounter unexpected barriers". Ergo: new technical breakthroughs could change everything once again. "Predicting the development of LLM applications is difficult due to emerging capabilities, changes in human perception and technological advancements." One important insight is that there are always components of jobs that almost all of us perform that could at least be improved by LLMs, if they are not too hallucinatory.
The US bank Goldman Sachs recently predicted that AI tools could affect 300 million full-time jobs worldwide. These include data analysis and accounting, customer service, financial advice, legal services and the creation of media content. Meanwhile, Eloundou - who had already looked into the topic last year- and her colleagues like to drink their own Kool-Aid in their work: the researchers' paper ends with the words that GPT-4 and ChatGPT were used for authoring, coding and as "formatting assistance" at work.
(dahe)
