Even one of the best artificial intelligence agents are pretty hopeless at on-line freelance work, in accordance with an experiment that challenges the concept of AI changing workplace employees en masse.
The Distant Labor Index, a brand new benchmark developed by researchers at knowledge annotation firm Scale AI and the Heart for AI Security (CAIS), a nonprofit, measures the flexibility of frontier AI fashions to automate economically useful work.
The researchers gave a number of main AI brokers a spread of simulated freelance work and located that even one of the best may carry out lower than 3 % of the work, incomes $1,810 out of a attainable $143,991. The researchers checked out a number of instruments and located probably the most succesful to be Manus from a Chinese language startup of the identical identify, adopted by Grok from xAI, Claude from Anthropic, ChatGPT from OpenAI, and Gemini from Google.
“I ought to hope this provides rather more correct impressions as to what is going on on with AI capabilities,” says Dan Hendrycks, director of CAIS. He provides that whereas some brokers have improved considerably over the previous yr or so, that doesn’t imply that this may proceed on the similar charge.
Spectacular AI advances have led to hypothesis about AI quickly surpassing human intelligence and changing huge numbers of employees. In March, Dario Amodei, CEO of Anthropic, prompt that 90 % of coding work would be automated inside a matter of months.
Earlier waves of AI have impressed misplaced predictions about job displacement, for instance regarding the imminent replacement of radiologists with AI algorithms.
The researchers generated a spread of freelance duties by way of verified Upwork employees. The duties span a spread of labor together with graphic design, video modifying, sport growth, and administrative chores like scraping knowledge. They mixed an outline of every job with a listing of recordsdata wanted to carry out the work and an instance of a completed mission produced by a human.
Hendrycks says that whereas AI fashions have gotten higher at coding, math, and logical reasoning lately, they nonetheless battle to make use of totally different instruments and to carry out advanced duties that contain quite a few steps. “They do not have long-term reminiscence storage and may’t do continuous studying from experiences. They can not choose up expertise on the job like people,” he says.
The evaluation presents a counterpoint to a benchmark of financial work supplied in September by OpenAI known as GDPval, which purports to measure economically useful work. Based on GDPval, frontier AI fashions equivalent to GPT-5 are approaching human talents on 220 duties throughout a spread of workplace jobs. OpenAI didn’t present a remark.
