Air Datasets · Synthetic Personal Data for Privacy AI

We help you solve the hardest problems.

Give your research room to breathe.

Research data without the year of legal negotiation. We build realistic, entirely artificial datasets you can just use: safe to share, clearly licensed, made in Switzerland. And when your work needs real data, we connect you with partners who have it.

15M+dataset downloads

131publications logged

9dataset releases

What becomes possible.

The best way to explain what we do is what happens when personal data stops being the obstacle.

Without a single patient record

A hospital NLP team trains and evaluates its masking models on synthetic personal data. Production-grade privacy tooling, and no real identity was ever at risk.

Without a data-sharing agreement

Personal data cannot leave the building, but synthetic personal data can. Train, mask, and evaluate across teams and borders without a single real identity in the set.

And when research matures beyond synthetic, we connect teams with real-data partners for the escalation stage. Why synthetic data works this way.

The datasets: AI4PRIVACY.

How do you build privacy AI when you are not allowed to look at the personal data it protects? You train it on data that never belonged to anyone.

A redacted document with one black bar lifting to reveal glowing synthetic text

AI4PRIVACY flagship

The largest open datasets of realistic, entirely artificial personal data. Downloaded over 15 million times and included in OpenAI's latest open source release.

TextTeach models to find and mask personal information, in many languages.

AudioTeach models to find and mask personal information in speech.

Explore AI4PRIVACY → · Ai4Privacy.com →

The Air Data Gazette.

Fascinating research the community has built on these datasets: logged, linked, and worth celebrating. A few recent favourites from our register of 137 external publications:

Google DeepMind · University of Maryland · Oregon State Privacy Backdoors: Enhancing Membership Inference through Poisoning A genuinely clever adversarial study - it made the whole field rethink what a poisoned checkpoint can leak. uses PII-Masking-200k

Indiana University Bloomington · JD Cloud · Stanford The Janus Interface: How Fine-Tuning in Large Language Models Amplifies Privacy Risks Beautifully uncomfortable finding: fine-tuning can resurface what a model was supposed to forget. uses PII-Masking-300k

Mitsubishi Electric Research Laboratories · Penn State Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Models Turns machine-unlearning from a compliance chore into a capability - elegant work. uses PII-Masking-200k

Northeastern University Rescriber: Smaller-LLM-Powered User-Led Data Minimization Puts minimization in the hands of the user, with small models - practical privacy at its best. uses PII-Masking-200k

Nanjing University The Fire Thief Is Also the Keeper: Balancing Usability and Privacy The rare paper that treats usability and privacy as one problem instead of a trade-off. uses PII-Masking-43k

TU Wien Privacy Agents: Utilizing Large Language Models to Safeguard Contextual Integrity Brings contextual integrity - the right privacy theory - into working LLM agents. uses PII-Masking-200k

Built where clarity meets responsibility.

The emblem is air.

Freedom for the expression of problems that are difficult to access due to legal or physical constraints.

Not a platform.Not a replacement for your data.

A new type of solution.For hard problems real data cannot safely reach.

When we want security, it's not a shield or a weapon. We want to take a deep breath and have a calm mind.