For early-career scientists, the cost of AI can be a problem

From ScienceMag:

Earlier this year, Saad Bhamla, who leads a biophysics and engineering lab at the University of Colorado Boulder, sent a poll to the nearly 2 dozen members of his lab: Which artificial intelligence (AI) tools were they using? How much were they paying?

The answers caught him by surprise. “I realized, holy moly, everybody’s paying anywhere from $20 to … $200” per month, Bhamla says. “I was like: ‘This is a tax.’”

It’s not a unique situation. As AI gets better at the work involved in science, more scientists see it as essential rather than a nice to have. But researchers tell Science access to the best tools today depends less on need than on who can pay—and that the costs of AI are falling especially hard on early-career scientists, many of whom are reluctant to raise the issue because it feels taboo.

In a survey of more than 10,000 researchers around the world, carried out in March by Nature, 41% said they paid for their main AI tool themselves, compared with just 11% who said their institutions covered the fees. Those who relied most heavily on AI to do their research were more likely to be paying out of their own pockets, and heavy AI users were also disproportionately early in their careers.

Some universities have struck campuswide deals with AI firms. And companies such as OpenAI, Anthropic, and Google also have programs to give tens of thousands of scientists free or reduced-price access to their tools. But the versions universities make available to students are often too limited for heavy coding or data analysis, and it’s unclear how long any giveaways from the companies will last. 

In Bhamla’s case, the results of his poll caused him to ask his university whether it could cover his lab’s bills. No, he was told, because it already gives students and faculty access to tools, including ChatGPT Edu. Ultimately, Bhamla decided to cover the costs of the AI his team says it needs out of his own funding, costing thousands of dollars a year.

Elsewhere, many graduate students are shouldering the costs themselves. Haris Riaz, a Ph.D. student in his final year at the University of Arizona, says he often pays $300 or more per month for Anthropic’s Claude Code and ChatGPT’s Codex, especially during crunch periods ahead of a deadline. For Riaz, who studies how large language models (LLMs) absorb knowledge, “all aspects of the research process … basically involve heavy usage of AI tools.” His university does provide access to LLMs, but it doesn’t offer the most up-to-date models and it doesn’t include agents—models that carry out multistep tasks on their own, such as writing and running code.

Riaz says he’s “lucky” that industry internships have helped him cover some costs; if he had to depend on his stipend alone, it would have been a “significant hit.” Yet he’s steadfast in his belief he needs the latest tools. “There’s no middle ground. It’s either you spend the money … or you work on some very small old problems.” 

Omar Choudhry, a Ph.D. student at the University of Leeds using AI for medical diagnosis and care, was able to cancel a personal subscription to Claude after his lab joined Anthropic’s discounted plan for scientists. It gives a research group up to 25 seats that are free, or priced at $15 per month for higher usage. Because Choudhry’s lab has fewer than 25 people, he says members realized they could each use several seats. He still has to pay the $15 for each seat he takes himself, but it’s a better deal than he had before. Still, he doesn’t expect the discounts to last—and he also pays $200 per month out of pocket for ChatGPT. “We all know that this is just temporary,” Choudhry says. “Eventually, all these prices go up.”

Support from institutions can also be fragile. Jeffrey Backus spends much of his day chatting with Claude or ChatGPT—he has subscriptions to both. He’s in the fourth year of a theoretical particle physics Ph.D. at Princeton University but works mostly from the nearby Institute for Advanced Study (IAS). Backus paid for his own $20-per-month ChatGPT subscription, but IAS covered his heftier $200-per-month Claude plan during a 3-month, computationally heavy project. That support didn’t continue, however, and he now pays for Claude himself. “This is so essential right now to what I’m doing that I would definitely fork over the money for sure,” Backus says.

(Princeton does offer more advanced AI tools to graduate students if a faculty member sponsors them, according to the university’s website. The university did not respond to a request to clarify who pays.)

Sometimes, the early-career researchers most likely to fall through the gaps are those least willing to say so. “AI usage is looked down upon,” says another Princeton Ph.D. student who asked to remain anonymous. “There’s a level of embarrassment, I’ve noticed at least, when people admit to having used AI for whatever research product they’re showing.”

The cost of accessing tools and “tokens”—the basic unit of AI computing power—raises questions about whether fees are shaping who gets ahead, says Igor Chirikov, a researcher at the University of California, Berkeley who runs an annual survey of students at research universities. The person “who has access to those tokens will probably be a better scholar,” he says, “as opposed to, if you’re a smart person, you’ll be a better scholar.”

In nations where funding is scarcer, the AI divide can be especially stark. Giovani de Almeida Valdrighi is a computer science Ph.D. student at the State University of Campinas who studies fairness in machine learning algorithms. He pays about $20 per month for Claude. Valdrighi says he can afford it only because the stipend he receives through his scholarship is “one of the highest” in Brazil. But for a student on the standard federal doctoral stipend of 3100 reais per month ($615) the same plan would consume about 3.5% of their income, and many of his friends go without.

Top-tier tools are further out of reach. Ana Clara Zoppi Serpa, a Ph.D. student living in Belo Horizonte, Brazil, who tests LLM safety, pays $10 per month for a cheaper service that gives her access to several models, topping up with credits when she hits her limit. “I would love to have a subscription of Kimi or Claude, but I end up finding them more expensive,” she says. At $200 per month, a top-tier plan is “almost the price of the monthly rent of a small apartment” near the university, she says. She could afford it, but isn’t willing to. And others, she notes, “can’t afford that, so it’s not accessible.”

For a long time, Valdrighi says he used Google’s free AI tools but switched to paid tools when he saw peers with better models moving faster. “I did not decide to start paying because I like the idea of the product,” he says. “It was more because of the pressure that everyone is using it, and you have to use it also.”

On an internship in Germany this year, Valdrighi was surprised to find students paying for their own subscriptions as a matter of course. “I was shocked when I saw this difference,” he says. “I believe that’s kind of unfair because the research becomes pay to win.” Labs in different countries have always had unequal equipment, he notes. “But the difference, I believe, is that the university is paying for this equipment, while with AI, actually it’s the researchers that are paying.”

Valdrighi isn’t alone in his unease. Several researchers whose access to AI is largely paid for by others told Science they worry about colleagues slipping through the cracks, whether down the corridor or overseas. “I’m very advantaged,” Backus acknowledges. “But it does really bother me.”

For Bhamla, the solution starts with having open conversations about AI’s costs. His lab now has an AI tools channel on Slack, and one researcher has even made an AI bot to answer questions team members have. “We are openly discussing and embracing becoming ‘AI-native,’” Bhamla says. In many cases, however, he thinks the relationships between faculty members and their students is a bit like those parents have with their kids. When it comes to AI, he says, “the parents are a little bit out of touch with what the kids are doing.”

Read More