As companies across multiple industries invest in new artificial intelligence technology, a growing gig economy targeting students to train these models has emerged.
Students across a wide variety of disciplines have started contract jobs through companies such as Mercor and Handshake, training AI models for clients in areas ranging from sports knowledge to math equations.
However, this growing market — which has trained AI models such as Anthropic’s Claude and OpenAI’s ChatGPT — is raising concerns as the rise in AI has influenced a decrease in entry level positions.
“I understand that people are concerned about (AI) in some ways … and I totally get that,” said Lucy Lawlor, a senior at UC Berkeley contracted with Handshake as a marketing AI trainer. “My thing is, I think … it’s inevitable … it’s coming down the line, so you’re going to have to learn how to work with it.”
Positions at Mercor pay anywhere from $23 an hour for general data annotation to $200 for doctors and other industry experts for specialized model training projects.
Data annotation jobs from Mercor and other recruiter platforms involve training and evaluating AI models’ responses and content generation capacity.
UC Berkeley sophomore Carter Capetz, who used to be contracted with Mercor as a machine learning engineer, said the jobs’ monetary appeal for students seeking part-time work was “very much so designed.”
The growth of this gig economy amid concerns about AI-induced unemployment has additionally led to questions about the end goal of these job listings. Many of these positions are specifically targeting students with “only an undergraduate degree, no advanced education.”
“A lot of startups these days value the speed of their execution over everything, (because) they need to show results to their investors,” said Avi Gupta, a UC Berkeley sophomore currently contracted with Mercor. “If that means losing the restrictions on what it means to be an expert to train your model, so be it. Undergraduates have a lot of time, generally, compared to working class professionals … I think it’s very difficult to lean on this contract life as full-time employment.”
AI model training involves two phases according to electrical engineering and computer sciences professor Alexandros G. Dimakis. In the pre-training phase, models are fed huge amounts of data such as raw text and media, and then they enter the post-training phase, which provides guidance to specific responses and behaviors.
This second phase is crucial to making AI models “align” with human objectives, such as by blocking AI from responding to dangerous objectives. More human input is required here, as the raw text used in the pre-training phase cannot give directions to guide behavior.
Despite these benchmarks, concerns around cybersafety and alignment have only grown with the development of new technology, with Anthropic reporting that its own model Claude was used in an attempt to infiltrate data from companies and government agencies, and Mercor being hit by a security breach late March.
In an email statement to The Daily Californian in response to this controversy, Mercor spokesperson Heidi Hagberg said “nearly every customer has continued operating as normal and has started new projects as our investigation continues.”