OpenAI is paying outsiders to test AI policy

OpenAI is putting money behind policy ideas that can actually be tested.

The company is awarding 14 independent projects a total of $1 million, plus up to $1 million in API credits, with work spanning the US, EU, Brazil, Singapore, and South Korea. The projects cover jobs, taxes, benefits, energy, healthcare, AI safety, and government use of AI.

That scope stands out. OpenAI already says ChatGPT has 1 billion users, so questions about who benefits from AI are moving from conference rooms into workplaces, hospitals, and government offices.

I think the useful part is the emphasis on testing specific policies instead of publishing another broad set of principles. Ideas like occupation-triggered “livelihood insurance” and privacy-preserving simulations of government benefits can now be examined against real constraints.

The projects run for six months, with results expected in 2027.

The harder question comes after that: which findings will governments actually use?

Claude’s text will carry a detectable pattern

Future Claude models will leave a statistical pattern in the words they generate.

Anthropic says it is adopting a version of Google DeepMind’s SynthID-Text to comply with the EU AI Act. Readers won’t see extra characters or labels on the page, and Anthropic says the watermark adds no tokens or additional cost.

The interesting part is what the watermark can actually establish. Someone with Anthropic’s key can estimate whether Claude was involved in a piece of writing, but the system cannot identify a user, prove human authorship, or reliably detect short passages.

I think that makes this useful for provenance, with clear limits. Heavy rewriting can remove the signal, while light proofreading may leave too little Claude-written text to detect in the first place.

Anthropic plans to offer a detection API and is launching the watermark globally because it cannot yet reliably limit it by region.

How people interpret a positive detection may matter as much as the technology itself.

China wants its data inside global AI models

China is preparing to export the material that AI systems learn from.

Beijing wants China to become a global data supplier by the end of 2028, with high-quality datasets covering more than two dozen strategic fields. It is already releasing government-backed collections through platforms including GitHub and Hugging Face.

One example is WanJuan, a dataset spanning history, law, medicine, current events, and other subjects that is explicitly aligned with “mainstream Chinese values.” It is also available in Arabic, Korean, Russian, Thai, and Vietnamese.

I think this makes training data a much more visible part of the AI competition. Researchers already found that ChatGPT and Claude gave more favorable answers about Beijing in Chinese than in English, likely because Chinese-language sources contain large amounts of state media.

Cheap Chinese models could make these datasets especially attractive in developing markets.

The important question is who decides which sources an AI learns to trust.

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.