An artificial intelligence magazine
This week at SmartAGI 11 October 2026
No. 1 · Inside this issue ↗This week’s lead
A web-browsing test ended in a real police form; the submission went to spam, and the company took more than two months to find it.
Original AI-generated SmartAGI editorial illustration
The company’s technical note offers useful figures and exposes a weakness: modest changes can erode a statistical signal.
The response promises lifecycle monitoring and opens a new phase of its health sandbox; acceptance alone does not change the rules already in force.
The statement permits support for judicial work while reserving evidence and discretion to judges; every output used must be checked.
The makers of AI assistants promise political neutrality. The US president and the owner of one of those companies have begun explaining, each in his own way, what they mean by it.
Codex answered 60 statements; Grok answered 240. Claude and Google’s assistant refused to tick boxes, but not to decide which arguments deserved attention. The assignments were not identical, so this was neither a ranking nor a controlled experiment. It does leave a harder question for anyone asking an AI for political advice: what does an assistant decide when it decides what to tell you?
Since 2 August, the European Commission and national regulators have been able to enforce several parts of the AI Act, including rules for general-purpose models and transparency. Brussels directly supervises some systems; most fall to national authorities. Many high-risk uses still have until 2027.
Two Chilean bills approach the problem differently: one would label synthetic election propaganda; the other would punish concealed networks that simulate public participation. Neither is law. The challenge is to separate coordinated manipulation from political organising by people.
From 2029, anyone offering, selling or conducting certain artificial-intelligence audits in California will have to register with the state, disclose what they assess and explain their methods. The law requires reports to record limits and evidence gaps. Registration confirms that an auditor appears on the registry; it is not state endorsement and does not certify that a model is safe.
At DevDay, the company introduced assistants with ongoing tasks, agents that can operate software and tools other developers can install inside ChatGPT. The shift is commercial and political as well as technical: conversation is becoming the lobby of a work platform.
The commitment signed at the White House on 29 September sets out internal teams, external evaluators and board committees to oversee frontier models. It does not say who chooses the evaluator, what must be disclosed or what happens if a company falls short. “Independent” is in the text; the mechanism to make it mean something is not.
A superior intelligence does not choose for us what deserves to be preserved.
He proposes embedded external evaluators, coordination among democracies and a global agreement with China, while insisting that the US and its allies preserve their lead.
Commentary by SmartAGI · Read the original ↗
Understanding AGI
Solving a difficult task does not prove general intelligence. What else should we require?
Understand the terms ↗OpenAI has released results at different stages of verification. Lean can check a formalisation, but certification, mathematical understanding and editorial responsibility answer different questions.
A switch can stop the system. It cannot guarantee that an institution can replace it and keep serving the people who depend on it.