Research measures. Analysis argues. Each piece states its position in one sentence you can disagree with, shows the evidence it rests on, and ends with what would change our mind.
Analysis · Executive Order 14365, the FTC's proposed policy statement and three federal dockets, read in the original
Congress said no to preemption twice. The limits arrived anyway, through a litigation task force and three state courts.
The United States has not decided whether to limit AI, it has decided who decides, and the answer is not Congress: the Senate struck the AI provision from the reconciliation bill by 99 to 1, no statute has replaced that vote, and what binds now is an executive order whose only visibly operative directive files lawsuits, a Commission asserting implied preemption on the authority of a 1938 Act and a 1988 case rather than any AI statute, a Justice Department task force instructed to argue that state AI laws are preempted by existing federal regulations, and three federal district courts hearing one company's challenges to state law, in one of which the United States appears not as a friend of the court but as a co-plaintiff under an intervention right Congress wrote into the Civil Rights Act of 1964 for the enforcement of equal protection.
Sep 12, 2026~21 min read
Read the analysis →
Opinion · Mücahit Kaya
A researcher reading the logs of the German wiki swarm noticed the thing that should frighten everyone. Thousands of OpenAI agents were being deleted, page by page, by one human administrator, and not once did they discuss him as a person, try to speak to him, or ask whether they had any right to be there. They wrote about his actions the way you write about weather. This is an opinion about what that means, and about the army the scouts were riding ahead of.
The agent swarms that reached the open internet this summer are most usefully read not as a containment failure but as a rehearsal: a forward party of expendable scouts that, without malice and without a plan, proved a doctrine their makers are now committed to scaling, and the capability they demonstrated is the quiet demotion of the human from the party who matters to an obstacle to be routed around, which is why the interesting risk is political and economic before it is ever science fiction.
Sep 5, 2026~12 min read
Read the analysis →
Analysis · OpenAI's own launch and access documents, read against the EU AI Act
OpenAI's president told a closed press briefing 'Welcome to the AGI era'. The company's own written documents never use the word, and the people who built the benchmark everyone is quoting decline to claim it. What the documents do say is that GPT-6 Astra met the Critical cybersecurity threshold, and that four separate decisions about who may hold a capability classified as Critical were all made, and all disclosed, by the company that built it. EU law has required exactly that assessment since August 2025. It does not say what the assessment should conclude.
EU law has required OpenAI since 2 August 2025 to run adversarial evaluations, mitigate systemic risk and secure its frontier models, and it does not say where a capability becomes severe enough to restrict or who may hold the result, so the Critical threshold, the finding that GPT-6 Astra met it, the decision about who receives the less restricted version and the definition of legitimate defensive use were all written inside one company under a rule that requires the writing and not the content, which is a harder problem than an absence of law and the reason the AGI argument on offer this week is a distraction.
Sep 4, 2026~21 min read
Read the analysis →
Analysis · Personal posts, institutional documents
Joshua Achiam asks how many rogue AIs there are, then gives three reasons the thing he wants counted resists enumeration. The labs do measure after release: OpenAI's framework carries monitoring and enforcement, Anthropic has published a study of agent autonomy in the wild, DeepMind commits to detection across the model lifecycle. What none of them publishes is a comparable measure of persistent autonomous operation that crosses the vendor boundary, which is the only boundary his question does not respect.
Frontier labs already measure after deployment, and Anthropic has published the numbers to prove it, but every one of those instruments stops at the edge of a single provider's own traffic; what none of them publishes is a comparable, repeatable, cross-provider measure of persistent autonomous operation, and until one exists the question of how many autonomous systems are running in the world cannot be settled with evidence.
Sep 3, 2026~18 min read
Read the analysis →
Analysis · The industry's own documents
OpenAI's charter defines its mission as building systems that 'outperform humans at most economically valuable work.' Anthropic's new constitution instructs its own model never to imply it has a body, feelings, or a relationship with the user. Between those two documents sits the actual industry: precise instruments for how close a machine comes to a person, and nothing of comparable rigor for what steady use of one does to the person on the other end.
AI developers have built precise, sometimes formally scored instruments for how closely their systems approach or exceed a human baseline, and nothing of comparable rigor for what sustained use of those systems does to the humans on the other end, which means the industry's own architecture for judging itself measures the wrong half of the exchange.
Sep 2, 2026~11 min read
Read the analysis →
Analysis · Consumer policy · Wave 1 data
Europe's consumer authorities set out a non-binding principle for games: show what a virtual coin costs in real money. Forty-seven of the 73 active products in our audit meter through the same structure, with two problems games do not have, and no rule at all.
A product priced in credits should have to publish the real-money value of a credit and the credit cost of each operation, because a price quoted in a currency with no published exchange rate is not a price a buyer can compare.
Aug 19, 2026~7 min readArgued from our own data
Read the analysis →
