OpenAI wants your private data to train smarter AI

Public internet data is drying up, and AI labs are getting desperate for new training fuel. OpenAI is now making a formal pitch to anyone holding proprietary archives: hand over your data, and we will help you turn it into AI.

OpenAI just launched OpenAI Data Partnerships, a program aimed at working directly with organizations, companies, and governments to build both open-source and private datasets for training future frontier models.

Deep archives over internet scraps

Instead of scraping random internet forum posts, OpenAI is hunting for high-quality data that expresses deep human intent, like long-form writing and multi-turn conversations across any language, modality, and topic.

They are accepting text, images, audio, and video. Even better for non-tech organizations, OpenAI is offering its own in-house AI stack to help clean things up. That includes using OCR to digitize mountains of old PDFs and automatic speech recognition to transcribe audio tapes, alongside data-cleaning tools to strip out personal details and sensitive third-party information.

This is not entirely theoretical. OpenAI previously partnered with the Icelandic government to level up GPT-4's Icelandic skills, and teamed up with the non-profit Free Law Project to ingest massive legal archives.

Public commons or private vault

Partners can choose between two distinct tracks depending on how protective they are of their files.

The first track is an open-source archive. Data donated here will be made entirely public for anyone in the AI community to train on, and OpenAI plans to use it to safely train open-source models of its own.

The second track is for proprietary datasets. If a company wants AI that masters its specialized domain without leaking crown jewels to competitors, OpenAI will build private datasets with custom access controls strictly for foundation, fine-tuned, and bespoke models.

Why it matters

We think this move shows that the low-hanging fruit of internet scraping is officially gone. Frontier AI development is transitioning from scraping whatever is publicly online to negotiating access for what lives behind paywalls and offline vaults.

For specialized institutions, minority language communities, and historical archives, this is a tempting proposition. Digitizing and cleaning obscure archives usually takes immense capital and technical effort, which OpenAI is now offering in exchange for training rights.

However, partnering with the dominant AI player comes with real trade-offs. Once proprietary domain knowledge is baked into foundation model weights, maintaining true data sovereignty is tricky, even with contractual privacy guarantees.

It also sets the stage for an aggressive land grab. As raw web text becomes saturated with AI-generated sludge, whoever locks down the world's offline legal, medical, and cultural archives will likely own the smartest models.