What is scale AI?

Why — the first-principles explanation

AI systems need a data and evaluation loop around the model. Raw images, video, text, audio or sensor data must be selected, labeled, reviewed and tested against the behavior a customer wants. Generative systems also need prompt-response examples, preference rankings, safety tests and domain experts who can judge whether an answer is useful or harmful.

Scale’s Data Engine describes the loop as “collect, curate, and annotate data; train models and evaluate; repeat.” Its Generative AI Data Engine describes expert-curated datasets, reinforcement learning from human feedback, red teaming and model evaluation. Scale’s documentation also describes a GenAI Platform for connecting data, fine-tuning models and deploying applications with proprietary data. The company’s current About page positions it as both a data provider and a full-stack technology provider for enterprises and governments.

That makes Scale different from a foundation-model lab. Scale can supply data operations, human review, quality tooling, evaluation and application infrastructure; the customer still defines the target behavior, chooses or supplies models, approves the data, and owns the deployment decisions. “Scale powers a model” is not the same claim as “Scale owns or trained the model.” Vendor pages are useful for what the company offers, but they are not independent proof of quality, market share or customer outcomes.

Corporate facts also need dates. Scale announced in June 2025 that a Meta investment valued it at more than $29 billion, that Meta would hold a minority stake and that Scale would remain independent. On July 30, 2026, Scale announced Francis deSouza would become CEO effective August 10, with Jason Droege supporting the transition after serving as interim CEO. Those are company announcements; exact cap-table rights, future leadership and customer relationships require current primary documents.

An example that makes it click

How to do it

1. Define the job to be improved: perception, document processing, generation, preference ranking, safety, evaluation or a full application workflow.
2. List the data modalities and labels required: text, image, video, audio, 3D, sensor fusion, expert responses or pairwise preferences.
3. Specify who is qualified to review the data and how reviewers will be trained, calibrated, sampled and audited.
4. Set quality metrics before procurement: agreement, error taxonomy, coverage, latency, rework, export format and acceptance thresholds.
5. Decide whether you need managed operations, a platform, an API, an in-house workflow or an open-source stack; compare control and switching costs.
6. Map privacy and security roles. Ask whether Scale is acting as controller, processor or service provider for each data flow, and what the customer must do.
7. Test the workflow on held-out data and adversarial cases. Include prompt injection, unsafe outputs, annotator disagreement and edge cases from the real domain.
8. Confirm model and data ownership, license, retention, deletion, regional processing, subcontractors, audit rights and export before sending sensitive data.
9. Separate vendor claims from independent evidence. Treat logos, “best” language, valuation and performance claims as claims that need context, not proof.
10. Recheck products, leadership, pricing and legal terms before signing. AI data and application offerings change quickly, so preserve the dated source set used for the decision.

Key facts

• Scale’s Data Engine describes a repeatable loop: collect, curate and annotate data, train models, evaluate them and repeat.
• Scale’s Generative AI Data Engine describes expert-curated data, reinforcement learning from human feedback, model evaluation, red teaming, safety and alignment work.
• Scale’s documentation describes Scale Pro, Nucleus dataset management and a GenAI Platform for connecting proprietary data, fine-tuning models and deploying applications.
• Scale’s About page describes three broad activities: training data and RLHF, model evaluation/red teaming, and full-stack AI systems for enterprise and government deployments.
• Scale is a data, evaluation and application infrastructure provider; that does not by itself mean it owns or trains every foundation model that uses its data.
• Scale’s June 2025 announcement said a Meta investment valued the company at more than $29 billion, that Meta would hold a minority stake and that Scale would remain independent.
• Scale announced on July 30, 2026 that Francis deSouza would become CEO effective August 10, while Jason Droege supported the transition after serving as interim CEO.
• Scale’s current public materials display customer and partner logos, but a logo is not a complete, current customer roster or independent evidence of performance.
• Scale’s privacy policy distinguishes controller and processor/service-provider contexts and says customers remain responsible for policies and legal compliance when Scale processes data for them.
• The highest-value buyer questions are reviewer expertise, data rights, quality controls, evaluation coverage, privacy, exportability and measurable model improvement—not the label “AI data.”

What authoritative sources say

People also ask

The same question, asked other ways

This page answers one intent expressed in 2 phrasings. How the index is organized →

• What does Scale AI do?

Related questions