Handshake has acquired Cleanlab, bringing the MIT-rooted data quality startup and nine of its researchers into the company’s expanding AI labeling operation. The deal, described by people familiar with it as an acqui-hire, adds co-founders Curtis Northcutt, Jonas Mueller, and Anish Athalye to Handshake’s research ranks to sharpen the company’s quality controls for human-in-the-loop data production.
The move aims squarely at one of the most stubborn bottlenecks in modern AI: mislabeled and messy data. Cleanlab’s software audits labels and flags likely errors automatically, a capability that could help Handshake scale its high-stakes expert labeling work for customers building frontier models.
Why Cleanlab’s Data Quality Tech Matters
Northcutt and co-authors helped popularize “confident learning,” a technique that estimates which examples in a dataset are likely mislabeled. Their team’s widely cited work found notable label error rates in benchmark datasets, including roughly 6% in ImageNet and higher rates in others—missteps that can mute gains from larger models and longer training runs.
In production, these errors can be costly. Gartner has estimated data preparation and cleaning can swallow up to 80% of a data team’s time. Cleanlab’s approach automates much of the auditing loop, cutting the need for a second human reviewer while preserving traceability. In practice, teams often recover measurable model accuracy and reduce the rework that slows iteration cycles.
Consider a medical coding project that draws on radiology notes and imaging reports. Even a small mislabel rate can bias downstream classifiers. Cleanlab detects likely mislabeled cases and proposes fixes, while Handshake’s network of domain experts—physicians, lawyers, and scientists—can resolve edge cases that automation cannot yet handle safely.
Deal Highlights and the Talent Rationale
Terms were not disclosed. Cleanlab had raised $30 million from investors including Menlo Ventures, TQ Ventures, Bain Capital Ventures, and Databricks Ventures, and at its peak employed more than 30 people. Nine key team members, including the founders, are joining Handshake’s research organization.
According to Northcutt, Cleanlab drew acquisition interest from multiple AI data labeling companies. Cleanlab chose Handshake in part because even competitors such as Mercor, Surge, and Scale AI often source highly specialized experts through Handshake’s platform. Folding Cleanlab’s algorithms into that expert marketplace could create a fused pipeline that audits data at scale before it reaches training stacks.
Handshake’s push into AI data labeling operations
Founded in 2013 as a platform connecting college graduates with employers, Handshake moved into human data labeling roughly a year ago to meet surging demand from foundation model developers. The company was last valued at $3.3 billion in 2022 and was forecast to close 2025 at a $300 million ARR run rate; it is reportedly tracking toward “high hundreds of millions” in ARR this year.
Handshake says it has supplied data to eight top AI labs, including OpenAI. By combining a large, vetted pool of expert annotators with automated error detection, Handshake can promise tighter quality guarantees and faster delivery windows—capabilities prized by teams fine-tuning models on sensitive legal, medical, and enterprise data.
Market and regulatory context for AI data labeling
AI builders are confronting diminishing returns from scale alone, as noted in recent Stanford AI Index reports and public guidance from major labs; attention is shifting toward data quality, provenance, and evaluation. Cleanlab’s audits offer an evidentiary trail that dovetails with emerging frameworks like the NIST AI Risk Management Framework and compliance expectations tied to the EU AI Act.
The labeling landscape is crowded, with players such as Scale AI, Surge AI, Labelbox, and Sama vying for enterprise and government contracts. Handshake’s differentiator has been access to credentialed professionals at scale; Cleanlab gives it a measurable edge in quality control, potentially lowering costs by catching issues earlier and reducing redundant reviews.
What to watch next as Cleanlab integrates with Handshake
The near-term question is integration speed: how quickly Cleanlab’s auditing will be embedded into Handshake’s tooling and rolled out across customer workflows. Expect Handshake to highlight quality metrics—error rates detected, examples remediated, and model lift after cleanup—as proof points.
Another watchpoint is whether Handshake keeps Cleanlab’s software exclusive or continues offering it to third parties. Keeping it in-house would strengthen Handshake’s moat; offering it broadly could open a new revenue stream. Either way, the acquisition signals that in the race to build better models, the next frontier is not just more data—it’s cleaner data, delivered with verifiable confidence.