Make the makers of high-impact AI, not society, responsible for proving that it is safe
Steps to make it happen
- People sort AI into several levels by how much it can affect people's lives.
- Companies that make high-impact AI hand over papers showing it is safe before they release it.
- Experts with no ties to the maker check those papers.
- The key parts of the results are published in a form anyone can read.
- If a problem shows up after people start using the AI, its maker explains why and fixes it.
- People who are harmed can ask the maker for help without having to find the cause themselves.
- Around the world, it becomes normal for the maker to be the one who shows that an AI is safe.
Water so far 💧 0
The proposal in detail
This proposal sets a principle about who carries the burden of proving that AI is safe. It covers AI that many people use, and AI used where a mistake directly affects lives and livelihoods, such as medicine, hiring, education, lending, and basic services like electricity and water. Low-impact tools and prototypes still in the research stage get light treatment, with only a notice required. Before release, the people who make or provide covered AI put together a document describing the intended uses, how it was tested and what the results were, its known weaknesses, and a summary of where the training data came from. The review is done by independent evaluators who take no money from the maker, and the key points are made public while trade secrets stay protected. If the AI does not meet the standard, its range of use is narrowed or more testing is required. When serious harm happens after release, the harmed person does not have to prove a defect. Instead, the maker is held responsible unless it can show that it took enough care. However, if the user clearly used the AI in a way no one could have expected, the responsibility is shared. Each society may draw the lines between levels differently, but the methods of evaluation are aligned internationally so that countries can recognize one another's review results.
BenefitsDangers are found before harm, and responsibility is clear
The biggest advantage is that dangers are found before the harm, not after. The makers are in the best position to know their own AI's weaknesses, and making them responsible for checking and showing those weaknesses first is the most efficient way to reduce danger. For harmed people, there is no longer a need to prove what is inside an AI they cannot see, so seeking relief becomes realistic. By showing where the training data came from, creators whose work was used also get a clue about how their work was handled. When the key points of the reviews are public, hospitals, schools, companies, and others have something to compare when choosing AI, and do not have to rely only on advertising claims. Makers also gain clarity about how far they must go to have done their duty. They worry less about being sued under vague standards, and companies that take safety seriously are fairly rewarded. If evaluation methods are aligned internationally, the work of repeating the same tests in each country is saved, and it becomes harder to run away to places with loose regulation.
DrawbacksIt may slow progress and leave only big companies
What opponents worry about most is that it could crush new technology before it grows. Review before release takes time and money. Wealthy large companies can bear it, but it may be too heavy for small companies, university researchers, and makers of AI released for free so that anyone can use it. As a result, power may gather in a few large companies and competition may weaken. Countries with little money or few experts may not be able to carry out reviews at all, and could be left behind and miss the benefits of the technology. Moreover, for general-purpose AI that can be used for anything, showing in advance that every use is safe is difficult in principle, and however the standard is set, loopholes or excessive regulation will appear. There is also a worry that passing review will be taken as a stamp of approval for safety and cause people to let their guard down. The number of independent evaluators is limited, and if the wait for review grows long, useful AI will reach people late. If there are lives in medicine and other fields that could have been saved by using AI sooner, the point that such delay is also a harm must be taken seriously.
Learn the background
AI is quickly working its way into the deeper parts of daily life, such as writing text, spotting signs of disease, and choosing job candidates. Yet in many cases AI is released without being checked well enough, and problems are found only after people use it. Even when biased decisions, false information, or complaints about other people's work being used for training without permission arise, it is hard to see inside an AI from the outside, and for a harmed person to find the cause and prove it is almost impossible. There is a large imbalance in information: the makers know what is inside, while users and society do not. Meanwhile, for technologies with large effects that cannot be undone, like medicines and airplanes, it has long been normal for the maker to show safety first. AI development moves in fierce competition, and releasing quickly tends to come first. Rules are starting to be made in various places, but if the ideas stay scattered, development may flow to places with loose regulation. What is needed is a shared idea, applying to every field of AI, about where responsibility should rest.
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 2026-10-15d) and translated by AI (claude-sonnet-5-5). Nobody has checked the facts. Read the original (日本語) How the AI writes