Cleanlab is an artificial intelligence (AI) company and open-source software project dedicated to data-centric AI, machine learning data quality, and AI reliability. Its foundational premise is that the reliability of AI systems depends entirely on the quality of the data used for training and evaluation. Rather than focusing exclusively on model architecture improvements, Cleanlab develops technology designed to identify issues within datasets, correct them, and measure overall data quality 1.

As of 2026, the organization has undergone a significant transition following its acquisition by Handshake in January 2026. The founders and research team joined Handshake to apply their expertise to frontier AI research, evaluations, reinforcement learning environments, and high-quality human data generation. Despite this corporate integration, the open-source cleanlab package remains available to the public 2.

The company was established by three MIT computer science PhDs: Curtis Northcutt, Anish Athalye, and Jonas Mueller 3.

At its most fundamental level, Cleanlab addresses the “garbage in, garbage out” problem inherent in machine learning. Traditional development workflows typically follow a linear path of collecting data, labeling it, training a model, evaluating performance, and deploying. Cleanlab posits that this process lacks a critical step: the automatic investigation of whether the underlying data itself is flawed.

Real-world datasets frequently contain incorrect labels, ambiguous examples, outliers, duplicates, mislabeled images, erroneous text classifications, problematic annotators, distribution shifts, anomalous data, overlapping classes, and unreliable labels generated by humans or automated systems. Cleanlab’s algorithms leverage information from machine learning models to identify these issues automatically. The open-source library is architected to function with various models rather than requiring a specific proprietary architecture 4. This methodology is collectively known as data-centric AI.

Data-Centric vs. Model-Centric Approaches

Cleanlab emerged as a prominent advocate for data-centric AI, distinguishing its philosophy from traditional model-centric approaches. While model-centric development focuses on improving neural networks through hyperparameter tuning, architectural changes, and extended training, the data-centric approach prioritizes verifying that training data is accurate, representative, consistent, and useful.

For instance, if a dataset of one million photographs contains 20,000 mislabeled examples, simply increasing the neural network’s size may not resolve the core issue. Cleanlab’s philosophy suggests that identifying and rectifying those 20,000 problematic examples often yields greater performance improvements than endless model tuning. Documentation for the platform specifically highlights automated detection for label errors, outliers, out-of-distribution examples, and annotator disagreement 5.

The history of Cleanlab predates its incorporation as a company. In 2017, MIT researcher Curtis Northcutt released open-source software based on his research into confident learning, a method for identifying label errors in datasets developed alongside collaborators including Lu Jiang and Isaac Chuang 6.

This research was built on the observation that if a model consistently predicts an example belongs to a different class than its human-provided label, the example is likely mislabeled. Rather than assuming training labels are perfect, confident learning estimates which labels are probably incorrect. This concept became the technical foundation for the subsequent Cleanlab software ecosystem.

Benchmark Dataset Research

One of the project’s most influential contributions was demonstrating that even gold-standard machine learning benchmarks contain substantial labeling errors. Research associated with Northcutt and his collaborators examined widely used datasets such as ImageNet, MNIST, and CIFAR-10, identifying significant numbers of potential label errors 7.

This work challenged the prevailing assumption among researchers that benchmark test sets represent ground truth simply because they have been in use for years. By proving that long-standing benchmarks are not necessarily perfect, this insight helped popularize the broader concept that dataset quality must be explicitly measured and improved.

The central technical component of the ecosystem is the open-source Python package cleanlab. This library automatically detects data and label issues in machine learning datasets, assisting practitioners in training models that are more robust to noisy data 8. It is installable via standard Python package management tools, including pip and Conda 9.

Core Capabilities

• Label-error detection: Identifies examples where assigned labels are likely incorrect.
• Outlier detection: Flags unusual examples that do not fit the general dataset distribution.
• Out-of-distribution detection: Finds examples that differ substantially from the training data.
• Duplicate / near-duplicate detection: Locates examples containing redundant information.
• Dataset health: Provides high-level metrics regarding overall dataset quality.
• Multi-annotator analysis: Analyzes data labeled by multiple individuals to estimate consensus and annotator quality.
• Active learning: Assists in determining which examples would be most valuable to label or relabel next 10.

Operational Mechanism

Cleanlab does not simply assert that a model is correct when it disagrees with a human label. Instead, it uses statistical estimation derived from model predictions and dataset structure to assess label quality. For example, if a classifier consistently predicts a different class for a specific image across appropriate out-of-sample evaluations, that image becomes a candidate label error. The system ranks examples by suspicion level, allowing human reviewers to investigate high-priority cases efficiently.

Beyond the Python library, the company developed Cleanlab Studio, a commercial platform designed to automate data curation workflows. While the open-source library is developer-oriented and requires machine learning knowledge, Cleanlab Studio targets analysts, data scientists, and teams through a web UI, Python API, and CLI with significantly higher automation 11.

A typical workflow involves uploading a dataset, analyzing it for issues, reviewing suggested corrections, exporting a cleansed dataset, and proceeding to training or deployment. The platform can analyze tabular, text, and image datasets, automatically inferring schemas and surfacing missing values or processing errors 12.

Unlike traditional data validation tools that rely on manually written rules (e.g., “this column cannot contain NULL”), Cleanlab Studio uses AI models to detect semantic inconsistencies that are difficult to express programmatically. It asks whether an example appears inconsistent with what the rest of the dataset suggests it should be, moving beyond simple syntax checking to semantic validation 13.

Cleanlab supports multiple data modalities, including tabular data, text, images, audio, and other supervised learning settings. Documentation covers applications ranging from classification and segmentation to object detection.

The technology later expanded into generative AI and large language models (LLMs) with the introduction of Trustworthy Language Model (TLM) technology. TLM aims to estimate the reliability of an LLM’s response rather than merely generating an answer. This capability is particularly relevant for addressing hallucinations, retrieval-augmented generation (RAG) failures, enterprise knowledge base errors, and automated decision-making risks. The current platform describes systems capable of detecting hallucinations, retrieval errors, knowledge gaps, policy violations, and other AI agent failures.

Cleanlab’s development progressed through distinct stages, evolving from simple label error detection to comprehensive data-centric AI, then to automated workflows via Cleanlab Studio, and subsequently to AI reliability for generative models. Following the acquisition by Handshake in January 2026, the focus shifted toward frontier AI research, specifically addressing how to produce and evaluate high-quality data for increasingly capable systems.

Handshake described the acquisition as a strategic combination of Cleanlab’s expertise in data quality, AI robustness, evaluation, and human data with Handshake’s existing AI data business 14. The three founders and a group of researchers and engineers joined Handshake’s AI research organization 15. Crucially, the open-source project continues under the more permissive Apache-2.0 license 16.

In its current iteration under Handshake, the team contributes to frontier-model evaluations, reinforcement learning environments, AI safety, data specifications, and high-quality human data generation 17. The public website continues to present technology centered on AI reliability and the remediation of incorrect AI agent responses 18.

Prior to its acquisition, Cleanlab raised approximately $30 million in venture capital. This included a $5 million seed round in 2023 and a $25 million Series A in October 2023. The Series A was co-led by Menlo Ventures and TQ Ventures, with Bain Capital Ventures participating as an earlier investor. This financing reflected growing investor recognition that as AI models increase in power, the quality of their training and evaluation data becomes a critical bottleneck.

• Curtis Northcutt: The principal intellectual figure behind the original research, Northcutt earned his PhD at MIT and developed the confident learning framework. He served as CEO and co-founder before joining Handshake to lead AI research and strategy 19.
• Anish Athalye: An MIT PhD who served as CTO and co-founder, Athalye joined the Handshake research team following the acquisition 20.
• Jonas Mueller: Also an MIT PhD, Mueller served as Chief Scientist and co-founder and similarly transitioned to Handshake’s research organization 21.

Collectively, the founders provided Cleanlab with a strong academic and research identity rooted in MIT’s machine learning community.

Cleanlab is significant for shifting industry attention toward the principle that the dataset is an integral part of the model. Even systems with trillions of parameters learn from data; if that data contains systematic errors, bias, duplicates, or low-quality annotations, scaling the model does not necessarily solve the underlying problem. Cleanlab made automated data diagnosis a first-class component of the machine learning workflow, positioning it at the intersection of data quality, model evaluation, AI reliability, and frontier AI research.

It differs from conventional data cleaning utilities by being model- and data-driven rather than rule-based. This allows for the detection of semantic and statistical problems that explicit validation rules cannot capture 22. Typical use cases span customer support ticket classification, computer vision quality assurance, NLP document categorization, enterprise analytics auditing, medical ML review prioritization, LLM hallucination detection, and human annotation quality measurement.

Strengths

• Research Foundation: Deeply connected to MIT research in confident learning.
• Model-Agnostic: Works with predictions from diverse ML models without requiring proprietary architectures 23.
• Multimodal Support: Extends beyond tabular data to text, image, and audio tasks 24.
• Open-Source Ecosystem: Enables direct use of underlying technology by researchers and developers.
• Generative AI Relevance: Applies data quality principles to LLM training and RAG systems.

Limitations

Cleanlab is not an arbiter of objective truth. Models may disagree with human labels due to model error, genuine ambiguity, unclear definitions, hidden bias, or unintended learned patterns. Consequently, automatic issue detection requires human judgment, especially in high-stakes applications. The intended workflow emphasizes reviewing detected issues to produce a cleaned dataset rather than blindly deleting flagged items 25. Furthermore, data quality is context-dependent; a perfectly consistent dataset may still be unsuited for a specific task.

• 2017: Curtis Northcutt open-sources the early Cleanlab project based on confident learning research 26.
• 2021: Cleanlab Inc. is founded by Curtis Northcutt, Anish Athalye, and Jonas Mueller.
• 2021–2022: Research demonstrating label errors in popular benchmarks brings data quality into wider AI discussions 27.
• 2023: Company raises $5 million in seed funding and expands Cleanlab Studio 28.
• October 2023: Raises $25 million Series A, bringing total funding to approximately $30 million.
• 2024: Expands into LLM reliability with Trustworthy Language Model (TLM) technology 29.
• January 2026: Handshake acquires Cleanlab; founders and research team join Handshake’s AI research organization 30.
• 2026 onward: Research integrates into Handshake’s frontier-AI efforts while the open-source cleanlab project continues 31.

1. cleanlab
2. Cleanlab | Experience GenAI that doesn’t hallucinate
3. Handshake acquires Cleanlab | Handshake
4. Cleanlab | LinkedIn
5. Cleanlab – Products, Competitors, Financials, Employees, Headquarters Locations
6. cleanlab
7. The Workflows of Data-centric AI for Classification with Noisy Labels – cleanlab
8. Cleanlab | AI Wiki
9. Cleanlab Studio | Cleanlab Documentation
10. Getting Started with Cleanlab Studio | Cleanlab Documentation
11. Frequently Asked Questions | Cleanlab Documentation
12. FAQ – cleanlab
13. Tutorials – cleanlab
14. AI data labeler Handshake buys Cleanlab, an acquisition target of multiple others | TechCrunch
15. Cleanlab | Company Overview & News
16. Cleanlab Raises $25 Million To Help Solve AI Models’ Data Mess
17. Cleanlab Raises $25M Series A to Automatically Increase the Value and Accuracy of the World’s Enterprise Data Used by AI, ML, and Analytics Solutions
18. Cleanlab Emerges with $5 million to Automate Data Curation for LLMs and the Modern AI Stack
19. Letter from the CEO: Announcing Our Seed Funding and the Launch of Cleanlab Studio for Enterprise | Curtis Northcutt | 22 comments
20. Letter from the CEO: Announcing Our Seed Funding and the Launch of Cleanlab Studio for Enterprise | Chen Lu
21. We’re thrilled to join forces with Handshake, where we’ll be able to scale our team’s pioneering work to inflect change with the world’s leading AI labs. Hear directly from our CEO and Co-founder… | Cleanlab
22. Handshake Acquires Cleanlab to Advance AI Research | Handshake posted on the topic | LinkedIn
23. I’m happy to share that Handshake has acquired Cleanlab. Cleanlab brings MIT-PhDs with deep expertise in data quality, robustness, and evaluation to the Handshake team. Their work, alongside our… | Garrett Lord | 11 comments
24. cleanlab
25. cleanlab
26. FAQ – cleanlab
27. FAQ – cleanlab
28. cleanlab docs
29. cleanlab
30. cleanlab
31. Datalab guides – cleanlab