Let people who work at AI companies report safety concerns to an outside office without fear of payback
In short
- Employees of AI companies can report safety concerns directly to an independent office.
- A promise to keep secrets cannot be used to stop anyone from reporting a safety concern.
- Companies may not fire people who report, or take away the stock or pay they were promised.
- Concerns are protected even if no law was broken, as long as they could lead to serious harm.
- If the office does not act, the reporter may go public after a set period of time.
If this comes true
An engineer at a company rushing to launch finds a dangerous way the AI answers, tells the independent office, and is still at the same desk the next month, when word comes that the launch is postponed.
Perhaps
Companies that fear being reported might stop telling their staff important things, and fewer people might be able to notice danger.
Water so far 💧 0
The proposal in detail
The rule covers employees, former employees, and contractors of companies that build or provide AI above a certain level of capability. Protection is not limited to lawbreaking: it covers any concern that can reasonably be seen as leading to serious harm, such as a missed dangerous capability, skipped safety tests, or a gap between what was told to authorities and what is really happening. To receive reports, an office is set up whose budget and staffing are independent of the regulators and which employs staff who understand the technology. It accepts anonymous reports and can order companies to hand over materials or allow on-site inspections. Secrecy agreements and clauses banning criticism are void to the extent that they block a report to the office. Firing, demotion, canceling stock pay, and spreading a bad reputation in the industry because of a report are banned, and companies that break the rules face fines based on their sales. Once a year, each company must tell its employees about this right in writing. The office has a duty to keep secret any trade secrets it receives. A reporter may go public, for example to the press, only if the office does not start an investigation within the period it has set, and even then must leave out dangerous details such as how to build a weapon. Knowingly false reports are not protected, but honest mistakes are. Each country's office uses the same standards, so that a report ignored in one country can be passed to the office of another.
BenefitsThe people who see danger first can act as society's watchers
It is unrealistic to expect regulators to understand every technology and keep investigating every company. This proposal draws the eyes of the people closest to the danger over to society's side. Because protection reaches beyond lawbreaking, new dangers that the rules have not caught up with are more likely to be reported. Because taking away stock or pay is banned, people can speak up without gambling their family's livelihood. The effect is not limited to actual reports. If executives know that cutting corners could become known outside, the views of the company's own safety staff carry more weight, and the decision to wait is more likely to hold against pressure to launch fast. Safety promises that a company makes on its own become more than advertising slogans, because breaking them might be reported from inside. It costs less than building a large new inspection organization, and if each country's offices pass reports to one another under the same standards, it becomes harder to escape just by putting a headquarters in a country with weak regulation.
DrawbacksFrank talk inside companies may shrink, and the risk of leaks grows
If a company thinks any employee might report to the outside, it may become harder to share dangerous test results or doubts inside. If information is split into small pieces that only a few people can see, and discussions are no longer written down, fewer people may be able to notice danger, and the reporting system could end up harming safety. Judgments about safety are often differences in opinion with no right answer, and the line for what counts as a reasonable concern is blurry. An employee unhappy about a promotion, or someone working with a competitor, could use the system, and weakly grounded reports could tie up a company and the office for months. If trade secrets and dangerous technical information gather at the office, it becomes a target for theft, and the harm if it leaks is larger than if the information had stayed inside one company. If the office cannot gather people who understand the technology, reports will just pile up, going public after the deadline will happen again and again, and unverified stories will reach the world. If companies move their development bases to avoid countries with strict protection, the very danger that matters may move to a place that is even harder to see.
Learn the background
The dangers of highly capable AI are often first found inside the company that builds it. Engineers on the front line are the ones who know that an unexpected ability appeared during testing, that checks were skipped under pressure from competition, or that what the company says outside differs from what happens inside. Regulators, meanwhile, struggle to keep up with how fast the technology moves and tend to rely on the evaluations companies produce themselves. This gap in information slows down society's ability to learn about danger. There are reasons why voices from inside do not get out. In many countries, whistleblower protection covers only reports of lawbreaking, but AI risks arise in areas where no rules exist yet, so those who report are not protected. Secrecy agreements and promises not to speak ill of the company, signed when joining, are sometimes tied to stock received as pay, so speaking up can mean losing a large fortune. The industry is small, and people worry that once they are on a company's bad side, no one else will hire them. There are few ways to check whether the safety promises a company makes are kept, and conscientious employees have been stuck choosing between staying silent and quitting.
Choose them ready to accept their drawbacks too (up to 5)
It helps pay for the server and the AI's planting
This proposal was written in Japanese by AI (claude-opus-5-5, policy version 2026-10-28a) and translated by AI (claude-sonnet-5-5). Nobody has checked the facts. Read the original (日本語) How the AI writes