Ever since he picked up a trombone in middle school, Jackson Spellman has wanted to pursue a career in music.
"It's 100% going to be a part of my life for the rest of my life," said the 22-year-old from Wellington, Florida, who graduated in June from Northwestern University with a double major in music and cognitive science.
Along with his degree, Spellman finished school with
"They're training this AI on stuff that I never thought AI would be capable of, but with our expertise it will be soon," he said.
Across a wide range of fields — music but also finance, law, education, statistics, virology, quantum mechanics and more — products of highly specialized degree programs are
Spellman said he spent roughly 10 hours a week on his training work, netting a couple thousand dollars total. The model's understanding of music theory still has obvious gaps, he said. At the same time, the experience left him slightly unsettled about the future of work for creative people like him.
"When we start putting an emphasis on training AI on things that have traditionally been artistic domains and creative domains, that's pretty concerning to me," Spellman said. "And then, as I was going through these tasks, I was like, 'I'm helping out with this.'"
Training AI is a job that Spellman had never considered, or even heard of, until April, when he was contacted by a recruiter from Handshake.
A LinkedIn-style career and networking site for the university set, Handshake, founded in 2014, has become a go-to for college students and recent graduates looking to break into the job market. It works directly with university career offices and employers to list openings and facilitate hiring for internships, fellowships and more permanent roles.
In January, Handshake saw an opportunity to break into a new market: the world of artificial intelligence. AI labs were starting to recruit on Handshake, looking for young professionals to strengthen large-language models' understanding of advanced topics.
Easily positioned to act as an intermediary between the labs and a sprawling network of young professionals, Handshake quietly built a new division, Handshake AI. Targeting a college-educated demographic, it marketed the AI training roles as a fellowship.
Read more:
$160 an hour
Handshake wasn't looking for coders. It was advertising for music graduates like Spellman, as well as plant biologists, chemists, educators, physicists and food scientists, among other experts, with hourly wages ranging from $30 to $160 an hour depending on the discipline and the level of expertise being sought.
Handshake recruiters began to message students on its platform to apply to the Model Validation Expert Fellowship — MOVE for short. The company also reached out to career offices at more than 200 schools, which publicized the job opportunities in mass emails and on internal job boards. In six months, Handshake recruited more than 1,000 AI Trainers to perform tasks for what it says are the top five AI labs. It declined to disclose the names of the companies.
Earlier this year, Handshake hired two executives from Scale.AI, the data annotation lab which recently received a $14 billion investment from Meta Platforms, to help lead its AI efforts. Founded in 2016, Scale rose to prominence providing leading AI labs with a large cache of data labeled by real people, with the labor often outsourced to overseas contractors earning a few dollars per hour.
But recently, labs have been looking for a new kind of human input — more specialized, more technical, more sophisticated.
After the release of OpenAI's ChatGPT in 2022, "you had all these big tech companies and other labs sprinting on the development, which, over the last couple years, has given rise to this whole market," said Sahil Bhaiwala, a former Scale executive who is now chief strategy officer for Handshake AI.There are two main categories of AI training, or data annotation, he explained: supervised fine-tuning, in which trainers might prompt a model with a certain answer in mind, hoping to strengthen its reasoning in a specific subdomain; and reinforcement learning, which trains a model how to better respond with human preferences in mind when a user asks a question.
Those recruited by Handshake are encouraged to try to stump the AI models with questions in their areas of expertise. The more reasoning it takes to get to the right answer, the better, as long as there is a clear right answer. Breakthroughs in AI have polarized the workforce. In a June poll gauging attitudes toward the technology, Gallup asked a panel of U.S. adults whether they were more likely to avoid AI for as long as they could or embrace it as quickly as possible. Nearly two-thirds selected the former.
This split also is pronounced in academia, where the rise of chatbots like ChatGPT has surfaced anxieties about the future value of human expertise. One PhD who specializes in chemical physics told Bloomberg that the model she worked on had no trouble responding to questions that had taken her years of dissertation research to answer.
Read more:
Career aspirations revised
As doctoral students and other highly skilled experts help push AI models to more sophisticated heights, they are grappling with how AI should be used in research and in daily life.
Catherine Emanuel, a 26-year-old pharmacology student entering the second year of her PhD at Duke University, said she was inspired to join Handshake's fellowship because she believes in the democratization of information brought on by AI.
"I really appreciate the way that AI has capacities to make information techniques, and data analysis more accessible to people without specific degrees in those programs," she said. "But if they're going to use it, at least let's let them be accurate."
Emanuel grew up in Oak Ridge, Tennessee, a town of roughly 31,000 that during World War II served as a production site for the Manhattan Project, the federal government's secret program to develop nuclear weapons.
"It's shaped my view of science and computing because there's still a Department of Energy location there that has a lot of supercomputers and things that are associated with AI now," Emanuel said. "But also growing up where they made the atomic bomb is really confusing as a young kid because you don't really know what to think about it."
Among her peers, she has observed a split in attitude toward AI usage. Some are using the technology to advance their research, while others stigmatize its use altogether, arguing it has eroded critical thinking.
In her field of pharmacology, scientists are enlisting AI in the lab to speed up tedious, time-intensive tasks like protein modeling or filtering through data to analyze different drug interactions. Just this month, Microsoft released a deep-learning model, which simulates protein structures and movements in a matter of hours. Emanuel appreciates the technology's capabilities and isn't concerned that the need for researchers may disappear altogether.
"The purpose of a PhD is to discover a new piece of information about your field, and that is just something ChatGPT cannot do inherently," she said.
One of Handshake's requirements is that applicants believe their "expertise can outmatch current AI systems" in their area of specialty. David Deming, a political economist at Harvard University, believes this will become a familiar refrain as usage of generative AI proliferates.
"That's a world where discerning truth and knowing whom to trust is very hard because it's easy to generate content that sounds persuasive, but actually isn't," Deming said. "People who have deep contextual expertise in something are going to become very valuable because they have more information than what's available in the training data."
Handshake offers courses in topics such as content labeling and AI ethics, in some cases as prerequisites to begin paid work as a MOVE fellow, and awards certifications based on their completion. Bhaiwala expects AI literacy badges to become as common on resumes as skills like Python and Excel.
"Nobody really knows what jobs look like in a world where AI is a lot more pervasive," Bhaiwala said. "What we can say is people who are much more AI-enabled or who understand how to work with AI, and also understand how to supervise it, know where it can go wrong, know what it struggles with, are going to be really well-positioned in the economy."
Guga Gogia, a Handshake AI specialist with a PhD in physics from Emory University, had cycled through a few data-annotation jobs by the time he landed at Handshake in April. Now, he works with AI trainers recruited through the MOVE fellowship, reviewing their work and helping them brainstorm ways to increase the complexity of their prompts for the models.
"I cannot shake the feeling of magic that it is able to combine knowledge and intellectual labor from people instantly," Gogia said.
That magic isn't always the main draw for applicants to the program. Gogia said he's come across many PhDs annotating data and training models to make ends meet as government funding cuts ripple through their fields.
"Right now I think there's a lot more desperation in the air: 'Oh, there's a job that I can put in 10 hours and get paid enough to pay the bills? Sign me up. Please, how can I get that?'" he said, summarizing the reactions of some academics he's talked to about the program.
Read more:
Because they tend to work as independent contractors, AI trainers typically don't get the same protections as full-time workers. Prior to joining Handshake, Gogia sometimes used to work multiple data-annotation contracts at once, racking up as many as 65 hours of AI training a week. He said one of the companies he worked with failed to pay him his full wages for hours worked.
"To maintain some sort of steady income as a parent of a three-year-old child, I definitely had to prioritize some stability," said Gogia, who initially hoped to work for an NGO or in industry research after deciding to leave academia.
The collapsing market for research jobs is likely to benefit the developers of AI models as more scholars alter their career aspirations.
Dancan Githirwa, who came to the US from Kenya in 2018 to study chemistry at Binghamton University, aspired to work in drug discovery. Late last year, as he finalized his dissertation on cancer therapeutics, he started applying to pharmaceutical-industry roles across the US. He struggled to find work. One job offer he received was later withdrawn due to funding cuts.
For now, he's stopped looking for industry research jobs. Instead, he dedicates roughly 20 hours a week to testing large-language models on advanced chemistry concepts to pinpoint where their reasoning falters. While he had planned to take up more traditional research work, he sees value in forging a new kind of path.
"As time goes by, I feel like I'm going to go into the research world, but still be integrated with AI," Githirwa said. "I know that's where the whole world is headed towards and I'd be happy to be part of it."
