Handshake

The data company training eight of the world's frontier AI labs spent its first decade as a college job board.

What Handshake actually does

Handshake started as the place college students find work. Universities hand it their whole student roster; students sign up free with a .edu email; employers pay to recruit them. That core still runs: 18 million students and alumni, more than 1,500 universities across the US and Europe, and over a million employers, including all of the Fortune 500.

Then the second business arrived. Training a frontier AI model now takes more than scraping the internet. To get better at graduate-level physics or contract law, a model needs actual physicists and lawyers to write hard problems, work through the solutions step by step, and grade what the model produces. That work is called human data labeling, and the constraint is people: experts with real credentials, available at scale. Handshake already had them. In January 2025 it launched Handshake AI, which routes verified people from its network to AI labs. The work is open to a range of backgrounds, from bachelor's-level generalists to postdocs, and pay scales with field and credential, roughly $40 an hour for general evaluation work up to about $125 for specialized domains. Current students can qualify too, not just alumni. Eight of the top labs, OpenAI among them, use the output.

The analogy

Amazon is the cleanest parallel, and the mechanism is specific. Amazon built warehouses and logistics to ship its own products, then opened that same network to outside merchants through Fulfillment by Amazon, including merchants selling against Amazon's own shelves. The hard, expensive asset built for the first business became a higher-margin service for a second set of customers. Handshake ran the identical move. The verified talent network it spent a decade building to recruit college grads is the asset; AI labs are the new customers renting access to it. The tell that the parallel holds: rival data-labeling firms now source experts through Handshake, the way third-party sellers run their stores on Amazon's rails.

Who they serve

Handshake serves two sets of paying customers off one network. On the recruiting side, the buyers are employers, from Fortune 500 firms down, and university career centers that license it as their career-services software. On the AI side, the buyers are frontier AI labs; the company has named OpenAI among eight it supplies. The supply for both is the same: students, alumni, and graduate-level experts, who join free through their schools. That free, university-fed supply is the whole point. The recruiting business and the AI business are not two companies sharing a logo. They are two ways to charge for access to one hard-to-assemble pool of verified people.

Who shouldn't use it

An AI lab that needs millions of cheap, generalist labels, such as drawing boxes around objects in images, is overpaying for a network built around credentialed experts; a high-volume vendor fits better. A lab that has already built its own in-house expert bench and annotation tooling gets less from Handshake's main convenience. On the recruiting side, a company hiring seasoned senior staff rather than students and recent grads is reaching for the wrong tool; a general professional network covers experienced hires better than a campus-rooted one. And a team that wants published, self-serve pricing should expect a sales conversation instead, especially on the AI side, which is quoted per project.

Sample customer stories

Hypothetical example: a frontier AI lab. Its newest model keeps missing graduate-level quantum mechanics questions, and the team has no way to recruit and vet fifty physics PhDs for a two-month push. Through Handshake AI, the experts arrive already credentialed by their universities, onboarded through Handshake's training program, writing worked solutions and scoring the model's answers on Handshake's own platform. The lab never runs the sourcing, vetting, or payroll. The win is the part competitors stumble on: the credentials are real, because they came from the school, not a self-reported profile.

Hypothetical example: a national retailer's early-careers team. It wants 200 interns across 40 campuses and can't fly recruiters to each one. On Handshake it targets students by major, school, and graduation year, runs virtual fairs, and fills the class without a single campus visit, work that used to mean a travel budget and a fall spent in airports.

What only Handshake can claim

Handshake didn't recruit its expert AI workforce on the open market. Universities delivered it, one .edu address at a time, over a decade, at almost no acquisition cost.

The network was built through institutions, not ad spend. Schools onboard their entire student bodies, and credentials are verified by the university rather than self-reported. Because of that, a lab paying for a molecular biologist gets someone whose degree is real, which is the difference between training data you can trust and data you have to re-check.

The same supply now feeds two demand sides, and that is the flywheel. Universities supply students; employers pay to hire them; AI labs pay to tap the experts among them; the revenue funds better tools and reach, which makes the network more valuable to the next school. The proof that the supply is the prize: rival data vendors, including Mercor and Scale AI, source experts through Handshake. When you can buy from the source, the middleman is the slower option.

The Cleanlab acquisition adds a quality layer. In January 2026 Handshake bought Cleanlab and brought on its founders, MIT computer-science PhDs whose software flags labeling errors without a second human checking. That matters to a lab because it shifts the pitch from how many experts to how good the data is.

Neutrality is a quiet advantage. Handshake isn't owned by an AI lab, so a customer's training data isn't flowing toward a competitor's model.

Breadth is optional, not a bundle. The network spans roughly 200 specialties and more than 500,000 PhDs, so a lab can pull music theorists this month and virologists the next without onboarding a new vendor each time.

Why this is hard, and why it matters now

Durable structural shift: the kind of human help AI training needs has changed. Early models learned from cheap, generalist labeling. Frontier models have run out of easy gains and now need domain experts, physicists, doctors, lawyers, to define what a good answer is and to write the problems the model still gets wrong. That turns "find some labelers" into "find verified PhDs across 200 fields, fast," which is exactly the bottleneck Handshake's decade of university relationships already solved.

Current shift: in June 2025, Meta took a 49% stake in Scale AI for $14.3 billion and hired its founder. Overnight, labs competing with Meta had to weigh whether to keep sending training data through a vendor a rival now part-owned. Demand swung toward independent suppliers, and a neutral network with verified experts was suddenly worth a great deal more.

What people would use instead

Without Handshake, an AI lab has three options, each with friction. It can use a high-volume labeler like Scale AI, strong on scale but, since the Meta deal, no longer neutral. It can use an expert network like Surge AI or Mercor, both formidable, though Mercor and others lean on self-reported profiles and, in Mercor's case, source some experts through Handshake anyway. Or it can recruit and manage experts itself, which is the fragmented, slow path that became the industry bottleneck in 2024. On the recruiting side, the fallback is a general job board like LinkedIn or Indeed, which reaches everyone but lacks the campus depth, or sending recruiters to campuses by hand, the expensive old way Handshake was built to replace.

Competition

• AI data labeling: Scale AI (now part-owned by Meta), Surge AI (bootstrapped past $1B in revenue), Mercor (also past $1B in gross annualized revenue), Invisible, and Prolific.
• Early-career recruiting: LinkedIn and Indeed at the broad end; Symplicity, 12twenty, and RippleMatch in campus career services; JobTeaser in Europe.

The data vendors are built around volume, speed, or a roster recruited on the open market. Handshake is built around a verified network it acquired through universities and now sells on two sides. In recruiting, LinkedIn and Indeed are built around the open professional web; Handshake is built around the campus and the institution's stamp on each profile. The unusual position is that several of its data-labeling rivals are also its customers, sourcing the experts they resell.

Website analysis

The most remarkable fact about Handshake is missing from the homepage a job seeker lands on. The page leads with "Let's find your next job," a search box, and three tiles: get hired, work with AI, grow your career. To a visitor, that reads as a nicer job board. What it doesn't say is the thing that makes the AI gigs real and well-paid: this is the network eight frontier labs, OpenAI included, rely on to train their models, and the network rival data companies quietly source from. A high-credential student deciding between Handshake and a generic site has no way to see that the specialized AI work here pays up to $125 an hour and routes to the top labs. This is a buried-moat problem. The proof of seriousness, the labs, the expert pay, the fact that competitors buy from the same pool, lives in the blog and the press, not on the page where it would change a sign-up decision. The fix is structural: put the moat on the homepage.

Website rewrite

This is a both-broken case for the AI-economy story: the hero undersells, and the subhead leads with a number that doesn't carry the pitch.

• Current hero (verbatim): "Let's find your next job"
• Current subhead (verbatim): "1M+ companies ready to hire. 100+ AI specialist roles across all levels."
• Current CTA (verbatim): "Get started"
• Rewritten hero: "Get paid to teach AI what you know."
• Rewritten subhead: "The labs building frontier AI hire through Handshake. No AI experience required; specialized fields pay up to $125 an hour."
• Rewritten CTA: Keep as is. "Get started" fits a free sign-up.
• Reasoning: the original hero treats Handshake as interchangeable with any job site, spending its best line on a generic promise; the rewrite names the one thing no other career site can offer a credentialed student. The subhead drops the "1M+ companies" stat, which says nothing a competitor couldn't, and swaps in the proof that makes the offer credible and the pay that makes it worth a click. Job-to-be-done: for the high-value visitor this page wants, the problem is awareness, not positioning, so the hero leads with the offer rather than claiming a category. Hero and subhead complement rather than echo: one makes the promise, the other backs it with who's hiring and what it pays.

Messaging to consider

A gift to the team: three lines Handshake has earned the right to say.

1. The network frontier AI labs train on.
2. Verified by the university, not the profile.
3. Even our competitors source from us.

Lead with the third. It passes ownability most cleanly: it's backed by Cleanlab's founder on the record saying rivals source experts through Handshake, and no competitor can claim the inverse. It re-categorizes Handshake out of the crowded "data labeling vendor" set and names it the source the middlemen buy from.

Likely next questions a prospect would have

• For an AI lab: how is a project scoped and priced, and what's the turnaround to stand up a team of domain experts?
• How is expert quality measured, and what does the Cleanlab error-detection layer actually catch?
• How is data security and confidentiality handled, given competitors share the same supply network?
• For an employer: what's the difference between the free tier and the paid Talent Engagement Suite, and where does pricing start?
• For an expert: how do I qualify, how is pay set by field, and what kinds of projects will I be screened out of?
• How separate are the recruiting and AI businesses, and does signing up for one expose my data to the other?

Sources

Company sources: https://joinhandshake.com/ , https://joinhandshake.com/blog/our-team/introducing-handshake-ai/ , https://joinhandshake.com/ai/ , https://joinhandshake.com/ai/opportunities/ , https://joinhandshake.com/employers/

Third-party sources: https://techcrunch.com/2026/01/28/ai-data-labeler-handshake-buys-cleanlab-an-acquisition-target-of-multiple-others/ , https://sacra.com/c/handshake/ , https://www.lennysnewsletter.com/p/inside-handshake-garrett-lord , https://www.cnbc.com/2025/06/12/scale-ai-founder-wang-announces-exit-for-meta-part-of-14-billion-deal.html , https://en.wikipedia.org/wiki/Surge_AI