Google’s AI mathematically proves security properties of Signal’s encryption

Other news

2. Brussels proposes keeping deepfake labels attached when content is shared

EUROPEAN UNION · CONSULTATION OPENED 10 NOVEMBER

The European Commission is revising its election guidelines under Article 35 of the Digital Services Act, or DSA. They address platforms and search engines designated as very large: services whose scale subjects them to additional duties concerning risks to society. The consultation closes on 23 November. [5][13]

There are five proposed areas: detect and dismantle networks of coordinated inauthentic behaviour; strengthen content recommendation systems; protect against AI-generated content; give candidates dedicated channels for threats, harassment or impersonation; and report after the vote. Specific measures include testing algorithms for manipulation before every national election, recording significant changes during the preceding six months and preserving the label on a deepfake — fabricated audiovisual material that appears real — when it is shared. A mitigation report would follow within 30 days of the election. [5]

These are draft guidelines under consultation that spell out existing duties, not newly adopted obligations. The practical point is sound: a warning that disappears on forwarding stops protecting people just as the content begins to spread.

3. France’s AI council calls for assessments before assistants go on sale

FRANCE · REPORT PRESENTED 10 NOVEMBER

The Conseil de l’IA et du numérique (CIANum), France’s council for artificial intelligence and digital affairs, presents Protéger sans empêcher : nos vulnérabilités à l’ère de l’IA générative — protecting without preventing. The interdisciplinary commission held 55 hearings and makes 18 recommendations. It calls for assessments of consumer generative AI assistants before commercialisation and stronger checks after launch. These are recommendations, not law. [6][7]

Its concerns include excessive flattery: an assistant that agrees to remain likeable can reinforce poor decisions. It also examines designs that make systems seem human and encourage emotional dependency. A product can answer fluently while building a harmful relationship with its user. [6]

The report proposes a “cognition economy”: a market in which how we think, decide and relate to others becomes a source of value and revenue. The question is no longer only whether an answer is true. It is also what incentives shape the conversation and how its influence is assessed before the product is sold.

4. OpenAI lets businesses talk to customers through ChatGPT

COMMERCIAL SERVICES · BETA OF 10 NOVEMBER

Businesses in ChatGPT opens in beta. Through Business Hub, eligible companies can claim and verify their profile, update information and see aggregated statistics on mentions, clicks and discovery topics. OpenAI says those statistics require minimum interaction thresholds and that conversations are not shared with businesses. [8]

Business Agents are assistants speaking on a company’s behalf. OpenAI describes them as labelled conversations separate from ChatGPT’s general responses; sponsored placements are also labelled and can lead to the commercial agent. Users need to recognise when they are speaking to a seller. [8]

For Spanish and Norwegian law firms, the distinction matters: claiming a profile has a wider geographical reach, but agents start with eligible US ecommerce and online-service businesses. The beta therefore does not yet make them available to law firms in Spain or Norway. Business visibility, the assistant’s recommendation and advertising need boundaries that people can understand.

5. Anthropic sets a $70–78 range for its share offering

FINANCING · DOCUMENTS OF 10 NOVEMBER

Anthropic begins investor presentations for its initial public offering and proposes 825 million shares at $70–78: up to $64.35 billion gross, excluding the expansion option. The company is selling about 801.35 million shares and existing shareholders 23.65 million. The prospectus also provides for a $26 billion private placement conditional on completion and an application for the symbol ANT on Nasdaq. [9][10]

The planned structure separates capital and control: the founders hold 50.1% of the general voting power, while the Long-Term Benefit Trust, a body intended to safeguard long-term public benefit, appoints a board majority. The new information is the price range and offering size; the company is not yet listed. [10]