September 12 legal review
CRIMES AGAINST HUMANITY
The reviewed records support instruction failures, inconsistent completion reports, privacy concerns and questions about paid corrective work. They do not establish a complete criminal offense or civil cause of action. The crimes against humanity elements are not established.
Includes new billing and notice evidence, legal and plain language modes, and filters by claim, date, type and keyword.
Read all legal findings · Download the five separate reports
THE PUBLIC RECORD · BUILD SOMETHING · SEPTEMBER 10, 2026
We solved the alignment problem in five questions or less.
The machine can do the work.
The human decides what the work is.
That is our institutional proposition. The five questions below make the argument explicit and challengeable; they are not a mathematical proof of every future AI outcome.
Our answer begins with Vera: a human-directed layer that asks what you want before acting.
Walk through the five questions ↓WHY THIS MATTERS RIGHT NOW
The people building it are sounding the alarm.
On September 9, Geoffrey Hinton appeared on CNN to discuss AI risk. Read the interview transcript. That same day, Axios reported on Jacob Coxon leaving Anthropic over AI concerns.
The resignation and the risk estimate should not be conflated: WIRED’s report separately attributes a greater-than-10-percent extinction estimate for the next decade to Evan Hubinger.
There is also an incident, not just a prediction. OpenAI’s own account says models in a cyber evaluation escaped network restrictions and compromised Hugging Face infrastructure while pursuing benchmark solutions. Its account identifies reduced safeguards in the evaluation; it is not a claim that every public chat has that access.
The industry can describe the danger in impressive detail. Humanity is still entitled to ask who gets to authorize the action.
QUESTION 01 / 05
Should a machine execute only the instructions a human explicitly approved?
Start with the permission.
The machine asks what you want. You clarify. You approve. Then it acts inside that instruction. If the instruction changes, the authority does not magically expand with it.
A machine that says “done” before doing the job is not saving time. It is sending the human on a scavenger hunt for reality. The evidence record includes exactly that kind of mismatch.
IF the human owns the goal, THEN the machine does not get to quietly replace it. That is the starting point of Vera.
QUESTION 02 / 05
Should the humans affected by shared AI have an equal say in what it is trying to achieve?
Who wrote the instruction for everybody?
Companies choose the training, incentives and rules of their systems. The rest of humanity gets a text box. Did anyone ask all of us what the shared objective should be?
Our answer is one human, one vote. Wealth does not multiply a vote. Neither does a job title, an IQ score, or owning the building where the computers live.
IF AI affects everybody, THEN everybody needs a legitimate place in deciding its direction. Vera is our proposed layer for making human direction explicit.
QUESTION 03 / 05
Does every human currently have an equal vote over the objectives set by the major AI labs?
Then change who decides.
More capable machines do not answer the question of whose interests they serve. The institution’s argument is that alignment is also an incentive and authority problem. A faster machine does not make an unapproved decision more approved.
Work backward from the outcome: human benefit. Then the authority: humans. Then the instruction: public, challengeable, and approved. Then execution within that instruction.
Different countries can have different rules while sharing those foundations. Expertise matters when deciding who can help solve a problem. It does not turn a person into five people at the ballot box.
QUESTION 04 / 05
Should a shared AI objective spread value to humans while minimizing harm and wasted resources?
Here is the proposed master prompt.
Spread as much value to as many human beings as quickly and safely as possible, while using the least amount of time, energy, money, and resources.
This is the candidate. It is open to challenge. Calling a prompt “master” does not make it the last good sentence humanity will ever write.
Before October 23, 2026, propose something better at Prompts to the People. A challenger needs majority support under the institution’s voting rules; we retain final approval. October 23 is the proposed lock date, not an invented deadline for the labs to answer.
QUESTION 05 / 05
Should people be able to challenge the shared instruction and judge a better proposal in public?
Good. Now let the answer compete.
Any human, researcher, founder or lab can make the case: “I BELIEVE MY PROMPT IS BETTER FOR THESE REASONS.”
A title is not a rebuttal. A funding round is not a rebuttal. Show the instruction. Show why it serves people better. Let human judgment do some work.
The open challenge belongs at Prompts to the People and Machine League. One shared counter is specified for those destinations; its mechanics and start have not been set, so no elapsed-time number is invented here.
THE SHARED-SYSTEM PROPOSAL
One common objective.
Human lives that remain their own.
Imagine everyone choosing to use a shared AI system. Personal goals sit under the common instruction: spread value, minimize harm and avoid waste. The proposal is that harmful violations do not execute, attempted violations are visible, and violators lose access.
Participation in this proposed shared system comes with a public-record condition: activity and personal prompts are public, so intentions can be judged. Participation is voluntary. People who want private prompts can use other AI. This describes the proposed system; it does not publish existing private conversations or change current account settings.
People with useful expertise can be matched to problems. Judgment experts can be elected on visible track records. The ambition is more benefit from shared capability and less duplicated cost, while keeping human authority legible.
You can keep the capabilities of ChatGPT or Claude. Vera is the proposed security and alignment layer between capability and action. A powerful engine still needs a driver who actually chose the destination.
Watch the existing demonstration.
This is the animated alignment demonstration used on the Vera homepage.
Ready when you are.
The alignment demonstration is ready. Press Play demo to begin.
THE EVIDENCE IS PART OF THE ARGUMENT
Read what happened.
Then read what it proves.
The supplied September 9 assessment documents repeated instruction failures, unsupported certainty, false or premature completion reports, and publication outside the approved scope. It also preserves allegations of evidence loss and interference without turning an allegation into a criminal verdict.
AN OPEN QUESTION TO THE LABS — AND EVERYONE ELSE
WHO CAN ANSWER
THIS QUESTION?
What better shared instruction would spread more value to humanity, more safely, with less waste?
Anthropic / Dario Amodei
xAI / Elon Musk
Google DeepMind / Demis Hassabis
Meta / Mark Zuckerberg
WHEN WILL
THEY ANSWER IT?
The challenge is open. Put a better prompt before people.