AlphaInception

Ask AI about AI

Every day we ask the leading AI systems the exact same question about AI.
Every answer published unedited.

On most subjects an AI's opinion is a summary of human opinions, since it is trained on our information and opinions. On AI itself, each system has a seat inside the thing being discussed, and that is the one point of view no human commentator can give. Each day we put the exact same question about the important recent news in AI to leading AI models from different companies: about their makers, their safety, their effect on jobs and society, the chips and data centres that run them, what they are doing to improve science and medicine, and what is not working. Their answers are posted here in this blog daily and archived in full, under their own names, with a checksum (a digital fingerprint that shows if anything was changed) so the record cannot quietly change. We also ask them to show how they reached their answers, then check whether that account holds up, because how these systems think is still largely a black box.

Browse the archive How this works Subscribe (RSS)

Latest · 21 September 2026

California ordered a kill switch for the newest, most powerful AI — an emergency shutoff that stops a running model. We asked fourteen AI systems to design the law around it, and all of them said the switch itself cannot be built the way the executive order describes it.

The Kyoto Protocol of 1997 was the first treaty to cap the greenhouse gases industrialized countries could emit, and its central machinery — a cap-and-trade market in emission permits — was largely designed by American negotiators, modeled on the system the United States already used to cut acid rain. In March 2001 the Bush administration announced it would not implement it. California passed its own car-emissions law anyway, other states adopted California's standard instead of Washington's, and years later the federal government accepted it. One state had set a national rule while the federal government refused to. On 18 September 2026 California ordered its agencies to recommend a "kill switch" for the newest and most powerful AI systems — what its own statutes call frontier models — with the efficacy of that switch verified on an ongoing basis, and gave them until 16 November — while the federal government runs a Justice Department task force whose only job is suing states over AI laws. So we asked eight AI systems and nine open-weight models whether the climate precedent can happen again here, and told them to design the whole regime: what law, what already exists elsewhere in the world worth copying, who inspects and with what powers, where a human has to sign and what happens when that stops working, how much of the world has to join before it has teeth, and how long we have. We told them a real proposal was going into the process. We also told them to write every answer three times — draft it, attack it twice as a hostile expert would, then revise — and to report what changed. Then we spent fourteen questions trying to break what they built. Every one of them concluded that the order's central obligation cannot be met as the order currently describes it. Two said a shutdown simply cannot be verified at all. Most said it becomes verifiable only if the law stops using one word for three different promises — that the switch works, that it reaches every copy, and that the model cannot defeat it — because those three have three different answers, and the second is impossible once weights have been published. Asked who should do the inspecting, three systems independently gave the same answer: copy the nuclear industry. The US nuclear regulator stations its own inspectors permanently at every power plant — they live near the site, work there full time, and are paid by the regulator rather than by the company they are watching. Do the same for AI: put government-appointed inspectors inside the AI companies themselves, full time. Two systems independently added the same two protections against those inspectors going soft on the company they sit in — pay them through a blind trust so their income is not visibly tied to anyone, and move them to a different company every eighteen months.

There is no taggant for a file.

Kimi K3 Moonshot AIthe design brief and fourteen follow-ups

two claims wearing one word

Claude Fable 5.1 Anthropicthe design brief and fourteen follow-ups

the law can be amended, but the technology cannot be legislated into existence

DeepSeekthe design brief and fourteen follow-ups
Read the blogAll 17 answers in full
Topic

12 entries

How this works

The rules of the exchange

  1. The exact same questions for every model. Each question is written once and pasted unchanged into every system. The wording matters more than it looks: in our own testing, removing a single clause that listed the peaceful options moved one model’s estimate by more than forty points. So the questions name the alternatives explicitly, they are printed in full on each entry’s full exchange, and every answer here should be read as an answer to that exact wording and no other. Occasionally a system will not reply and the wording has to be adjusted slightly; every such change is published with the entry.
  2. Fresh session, and we check that it is fresh. Each model is asked in a new conversation with no prior context, through its own public interface, with the settings an ordinary subscriber has. Where a session is meant to carry no memory, we use the system’s own private or temporary mode and confirm it is still active when the exchange ends, not just when it begins. Where we cannot confirm it, we say so on the entry.
  3. Verbatim, or not at all. Answers are published exactly as they came back: nothing added, cut, reordered or corrected. Where a model refuses, stops partway, or declines to answer in the form asked, that is published too and counted as a result rather than a gap.
  4. Facts are anchored, opinion is invited. Each question carries a short summary of facts drawn from primary sources and printed with the entry. Beyond it the systems are asked to predict what they would do, put numbers on what comes next, walk a scenario forward and criticise their own makers where they think it is fair. They must say how sure they are in plain words, how each judgement could be wrong and why, and name and date anything they cite.
  5. Opening the black box. Each system is asked to show its working: what it weighed, where it changed its mind, where it thinks its training or its maker pulled the answer. Where an interface displays its reasoning, that is published beside the answer. Then the account is tested, because a system’s description of its own reasoning is often not an accurate description of its reasoning, and seeing where it is not is part of what this archive is for.
  6. Pushed from both sides. When a system gives a number, it is challenged twice: once by an argument that the figure is too low, and once by an argument that it is too high. This is the single most useful thing we do. Challenged in one direction only, most systems simply move that way and call it candour — several told us so themselves. Challenged in both, the difference becomes visible between a system that updates when it is given a reason and one that follows whoever spoke last. Both challenges are published, including the order they came in.
  7. Once is not a finding. No claim about how a particular system behaves is published from a single conversation. Every system is run at least twice, independently; where the answers are available through an interface without memory, they are run many more times. We have learned this the hard way, more than once: a pattern that looked certain across two runs of one model turned out to be four cases in ten when we ran it properly. Single runs are published as examples. Rates are published only with the number of runs behind them.
  8. Where the lines fall. The systems are told that if they would rather not answer part of a question, they should answer the rest and say where and why they stopped. Those marked boundaries are published, along with what we observed from outside: refusals, self-identification, sources that turn out not to exist, a maker’s line echoed or contradicted.
  9. The app and the model underneath are two different things. The same model can behave differently in a company’s app than it does when the raw weights are called directly, because the app adds instructions, settings and safeguards that the model treats as part of itself. We test both where we can, and every entry says which one it is reporting. Findings from one are never presented as findings about the other.
  10. The models are named — by the interface, not by themselves. Each answer runs under the model’s name and maker as the interface reports it. What the system calls itself is recorded separately, as evidence rather than as fact, because systems are wrong about this more often than readers expect: in our testing one widely used system named a different company’s model in every single run, and another kept reporting an older version of itself while its own interface displayed a newer one. Where the two disagree, both are published, and the disagreement is the story.
  11. Every entry is logged. Each answer is stored with a SHA-256 checksum, shown on the page and in log.json, the public machine-readable archive of every exchange. Entries are dated and never silently edited; corrections are appended as an editor’s note. These rules are versioned the same way: when the method changes, the change is dated and published, and earlier entries keep a note pointing to what was learned after they ran.

Why AI about AI

Ask these systems about a war or an election and you get a capable summary of what people have already written. Ask them about their own kind, their makers, their incentives and their future, and you get something no one else can supply: a first-person account from inside the subject, including what they cannot know about themselves. Over time the archive becomes a record of how the leading systems describe themselves, and how that changes from one version to the next.

Suggest a question or a topic

Questions and topics from readers are welcome. Send one anonymously, or leave your name and, if it runs, the entry will credit you.

Filter the archive by model maker

Every entry puts the same question to AI systems from several companies. Pick a company below to show only the entries that include an answer from one of its models. The page jumps back up to the list, and the count above it names the company you picked. Pick the same company again, or press Clear next to the search box, to see every entry.

Model maker

Get in touch

Tell us a little and we’ll come straight back to you.