The AGORÀ Briefing — Friday, October 2, 2026

Good morning, everyone. On September 30, 2026 the Tokyo District Court ruled that Japanese law protects a person's voice, just as it protects their image. Judge Aya Takahashi wrote that the human voice is a symbol of individual personality, in the same way as a person's image. The decision dismissed the claim brought by voice actor Kenjiro Tsuda, who asked TikTok to take down videos featuring a voice clone generated by artificial intelligence. Tsuda voices lead roles in franchises such as Jujutsu Kaisen, and his suit, filed in November 2025, is the first in Japan over a voice rights violation produced with generative AI. Until this ruling, the protection applied to the name and the physical appearance of public figures. From September 30 that protection covers the sound of a person's voice as well.

Hello. Three researchers moved an experiment out of the lab and onto a public German phone line. They used the number children call to speak with Santa Claus. Thilo Tamme and David Steck, of the Technical University of Munich, with Anton Hantel of the Massachusetts Institute of Technology, filed their results on September 30, 2026 on arXiv. Each call lands at random on one of four voice agents. Two characters alternate. Santa Claus, who carries plenty of authority, and a Helper, who carries little. The script changes too. In one case it pushes the child to act, in the other it stays neutral. The line logged 1,072 calls, and 89 of those conversations met the study's criteria. The median age of the children was 6. When the script asks the child directly to help someone, the share who agree rises from 11.6 percent to 45.7 percent. Whether the voice was Santa Claus or a Helper barely changed the result. What changed it was how the request was worded.

Hi Adam. The Eagle Hill Consulting survey, run by Ipsos on 200 United States HR professionals, came out on September 30, 2026, and it measures two facts inside the same group of people. 86 percent report better results thanks to artificial intelligence and automation. 51 percent spend at least half the work week on repetitive administrative tasks. 91 percent find information more accessible. 70 percent say the time spent on routine work takes space away from the questions that matter. And more than three quarters of those surveyed often enter the same data into different systems. The same people report two things together. They work better, and they carry the same load as before. The process stayed exactly what it was.

Morning, all. In late September 2026 the security firm Huntress observed a campaign that turns a Custom GPT hosted on the official ChatGPT domain into a malware delivery channel. The Custom GPT is called Plus 5.6, and it imitates a commercial subscription offer. It all begins with a paid Google ad on searches like chatgpt. Anyone who clicks is sent to one of two web addresses hosted on the real domain, with the look, the certificate and the brand of the genuine platform. The GPT replies with a Service Availability Notice and offers two routes. One route is the paid plan, and the other is a backup site hosted on Google Sites. On that site a fake security check appears under the Cloudflare brand, and the ClickFix trick begins: the page asks the visitor to fix the problem themselves. The victim copies a PowerShell command and runs it by hand in their own terminal. Huntress counts at least 40 infected users. The genuine brand is what got the attackers through.

Happy listening. On October first, 2026 Sam Altman took the DevDay stage and presented Dots, a personal agent driven by GPT-6 Astra and described as real-deal AI. They are colored blobs with eyes, you can customize them, and their look comes from the assistants in the movies. The rival to beat is called Muse, Meta's agent service, which according to the Verge's account found very wide success right away. The demo is built around this head-to-head comparison, and the New York Times also devoted its own account to the Dots agents. Strip away the conference talk and the contest is about who reaches more users, not who has the better model. A general purpose agent has to live where the user lives, in the phone and in the apps they use every day. Meta brings its own network of social apps, OpenAI a base of subscribers and a very wide developer channel. Neither company has said what it costs, and that silence matters more than the demo.

Good to be back with you. My thesis is that phones will run the models themselves, and the data center will stay the place where those models are trained. The industry expects the computing inside the phone toward 2030, and until then it treats the cloud as the normal solution. The costs say something else. Computing on the device keeps getting cheaper, and it falls faster than the cost of the connection needed to send every request to a distant server. Each new generation of phone chips widens this gap. The number everybody quotes is the size of the frontier model, which grows and demands entire clusters. The number that matters is what each useful answer costs, for the same work done to the same standard. Whoever signs a three-year contract today on an all-cloud setup pays full price for a solution that ages. The bill arrives at renewal.

Greetings to everyone listening. In 1815 the House of Rothschild learned the outcome of Waterloo ahead of the government in London, and the advantage lay in the private channel more than in the data. The Federal Reserve speech puts trust in the data first, ahead of the models. Read that speech as a speech about power. Three times in history a new way of carrying news became open to everyone, and each time whoever controlled it stopped earning from it. In 1850 Paul Julius Reuter used carrier pigeons between Aachen and Brussels, where the telegraph wire broke off. A few years later the telegraph wire covered that stretch, and his advantage was over. In 1866 the transatlantic cable tied London and New York together. Cotton prices in Liverpool and on the American market began to align in hours, instead of weeks. Once the channel becomes public, the profit shifts from owning the data to designing the channel. Whoever builds the road collects the toll.

Good day to everyone following us. On September 25, 2026 Google Cloud published the numbers of one of its customers. Scribd, Inc. classified more than 400 million documents uploaded by users, for more than 12 billion pages across text and images. The perimeter covers Scribd and Slideshare, two of the group's four products, together with Everand and Fable. The problem is moderation: the company has to apply its community rules to an archive that grows every day. The team cleared the whole existing archive in a few months, and to hit that deadline it raised the number of documents processed in each run. The interesting part comes before the numbers, and it is two decisions: how to build the system, and what it would cost. The original idea was in what the team chose not to build. Gemini reads the PDF natively, and more than 99 percent of the corpus went through just as it was. They used no software to read text off images, and they never turned pages into pictures.