A Say in AIRead · Check · Discuss · Decide · DeliverČesky
Who gets a say in how AI is built? Our view: today, very few people.
Our view: a handful of companies decide what the most powerful AI may do and how safe it has to be, and governments are joining in slowly, mostly one at a time. We know of no place where most people, in most countries, can read what the companies committed to, check whether it happened and say what they want. This is an attempt to build that place, in the open, one working part at a time.
Five parts
One place, in the order you would use it. Two are open; the rest are marked as what they are.
- 1
Read
What the companies committed to and published, in their own words, with the source for every line.
We start with the White House accord of September 29 and each company’s own safety documents. Our plain-words reading always sits beside their words and is marked as ours.
Open nowOpen - 2
Check
What they have shown since. Commitment by commitment, company by company, dated.
The first tracker follows the accord’s four controls and its commitment to meet, across six companies. It records what is public, not what is true inside a company, and says where we looked.
Open nowOpen - 3
Discuss
Short statements you agree with, disagree with or pass on. A map will show where very different people agree.
No comment threads to fight in. You vote on one sentence at a time, or write your own. It will run in Czech and English first.
Next to openPreview - 4
Decide
A few questions that deserve a real verdict, for example whether AI should be allowed to train its own successors.
First an open vote, then a panel of people chosen by lot to reflect the wider public, paid for their time. We publish both results and the gap between them.
Planned, not built yet - 5
Deliver
The results, with how many people took part and what we cannot tell about them, put in front of companies, governments and the UN.
We publish their replies, or their silence, next to what people asked for.
Planned, not built yet
Right now11 days after the accordChecked October 10, 2026
Six companies made the same four commitments, and one together: to meet regularly. 13 of 24 commitments have something dated since the accord, and in 5 of those the company’s own words refer to the accord.
The accord sets no deadline, no penalty and no public reporting, so nothing in it lets the public check whether the commitments are kept. The tracker follows each commitment for each company, with every source.
How we hold ourselves
Aims we can be checked against. If we break one, the page that broke it will say so.
People write the values. We don’t.
AI helps us translate and summarize, and every summary links to what people actually wrote. No model votes or speaks for anyone. The one exception is the starter statements on the discussion page, which we wrote to start the conversation and say so there.
A source for every line.
If we can’t open it, we don’t say it. Numbers come from counting, never estimating.
How many took part, and what we can’t tell.
A big count of clicks is not humanity. Every result says how many people took part, in which language, and what we cannot tell about them; where we know who was not reached, we say so.
The same standard for every company.
Including Anthropic, whose model helped build this site. If you read one row to check our work, read that one.