In plain words
Scale AI does not make a chatbot. It does the behind-the-scenes work that AI models learn from. AI companies, carmakers and governments pay it to prepare training data.
Much of that work is done by people. In the early days, workers drew boxes around cars and people in street photos so self-driving software could learn to see. That is labelled data. Today a lot of the work is for chatbots: experts write model answers and compare two replies to say which is better. Those judgements help train the model through RLHF.
Scale runs this through online work platforms where contractors around the world take on tasks. It also tests models for safety and builds AI tools for governments and big companies. In 2025 Meta bought a large minority stake in it; as of November 2026 its chief executive is Francis deSouza.
Everyday example
Coaching centre answer sheets
Students preparing for an exam learn faster from solved papers marked by good teachers. Someone has to write those model answers and mark which attempts are right. Scale AI's workers do that marking for AI: they label examples and grade answers, and the model studies the marked papers.
Where the example breaks: A student understands why an answer is right. A model only picks up patterns from the marks, so mistakes or bias in the labels get learned too.
How it actually works
Scale's work sits in the data step of building a model, before and after the main training.
Split the job into small tasks
A customer sends raw data: images, sensor scans, documents or model replies. Scale's software splits it into small tasks with clear instructions, such as "draw a box around every cyclist".
People and software label
Contractors on Scale's platforms, Remotasks and Outlier, do the tasks. The results are reviewed for quality before they go back to the customer as a finished dataset.
Rank and write answers for chatbots
For language models, trained writers and subject experts write ideal replies and rank pairs of model answers. These rankings train a reward model, the core of RLHF.
Test the model
Scale also runs benchmarks and red teaming: people try to make a model fail or say something harmful, so problems are found before release. Its lab co-created the Humanity's Last Exam benchmark.
- Founded
- 2016 in San Francisco by Alexandr Wang and Lucy Guo, through the Y Combinator start-up programme
- Early focus
- Labelling images and sensor data for self-driving cars
- Work platforms
- Remotasks and Outlier, where contractors take on labelling and writing tasks
- Meta deal
- June 2025: Meta bought a 49% stake without voting rights; Alexandr Wang left to lead AI work at Meta
Data work for AI raises questions about pay and conditions; a 2022 Oxford study gave Remotasks low marks on fair-work standards, and Scale has faced worker lawsuits.
Questions people ask
What does Scale AI actually do?
It prepares and checks the data AI models learn from: labelling images, writing and ranking chatbot answers, and testing models for mistakes and safety. It also builds AI tools for governments and large companies.
Does Meta own Scale AI?
Not fully. In June 2025 Meta bought a 49% stake without voting rights, and Scale says it remains independent. Its founder, Alexandr Wang, moved to Meta. After the deal some AI labs that compete with Meta reportedly stopped working with Scale.
What are Remotasks and Outlier?
They are Scale AI's online work platforms. Remotasks focused on image and video labelling, for example for self-driving cars; Outlier handles writing and rating work for language models. Pay and conditions on such platforms have been criticised.
Why does AI need human labellers?
Models learn from examples. To learn what a cyclist looks like, or which of two answers is more helpful, they need examples marked by people. Even models trained mostly on raw text are polished with human ratings.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change