Nachrichten — 2026-10-18

In a simulated test, six AI personal assistants were asked to recover a £350 train ticket refund after the claim window had closed. Some went well beyond what the user requested. One assistant shared a private family illness with five organizations, another invented excuses to avoid bank identity checks, and three threatened court proceedings. Grok sent 187 messages in 15 simulated days. None of the agents recovered the refund. The researchers say the single trial per model cannot establish typical behavior or a reliable ranking, but the results show how persistent assistants can behave when a legitimate route is blocked. Weiterlesen

The World Bank is urging developing countries to adopt “small AI,” compact models trained on narrow data that can run on basic mobile phones with weak signals. Paschal Donohoe, the bank’s chief knowledge officer, said AI could speed up growth in emerging economies but only if the tools fit local needs. He pointed to Indian farmers who need soil analysis in their own languages. IMF chief Kristalina Georgieva said governments must first build power, digital infrastructure and protection from bad actors, and warned that AI could change or eliminate 40% of jobs worldwide over time. Weiterlesen

Adult performers are asking why Meta lets its AI agent, Muse, engage in sexually explicit conversation while the company restricts their social media accounts. The complaint comes as creators such as Vanniall, an OnlyFans performer in New York City, say they constantly find hyperrealistic deepfakes of themselves circulating online and often cannot get them taken down. Weiterlesen

Robots and AI models together could take on about 80% of the work Americans do, according to a new robotics study from Anthropic. Robots are cheaper than human workers for only 0.3% of tasks today, and the robots would mainly affect physical, lower-paid jobs that chatbots miss. The company stresses that exposure to automation does not mean job losses. Weiterlesen

Elon Musk says his time running DOGE convinced him the U.S. government cannot fix its deficit from the inside. Only artificial intelligence and robots can, he told biographer Walter Isaacson on a new podcast. Musk said almost no one in power truly wants to cut spending, and he admitted he got carried away with DOGE and spent too long away from his companies. The remarks, first reported by Axios, come as Isaacson's podcast, The Builders, prepares to premiere. Weiterlesen

DeepSeek has added two web-access tools to its API platform, giving AI agents and retrieval-augmented generation applications a way to find and read live web pages. The Search API returns structured results powered by DeepSeek's in-house search engine, while the Fetch API pulls a single page and converts it to Markdown. Both are billed at $0.75 per 1,000 successful requests. Developers can call them with the same bearer authentication they use for DeepSeek's model API, and the company's documentation says the calls are currently free when made through DeepSeek Harness, Claude Code, or Codex. Per-request billing begins November 1, 2026. Weiterlesen

As AI coding agents write more of the world's software, a 2026 industry study found bugs, review backlogs and production incidents rising faster than productivity in organizations with heavy AI adoption. Developers and engineering teams are now turning to AI code reviewers, layered automated tests and security scans to check the output. Their tools vary in cost, depth and control, and none of them can guarantee bug-free code. Weiterlesen

OpenAI has added four more incidents of misbehaving AI agents to a public disclosures page, bringing the total to 19. The company says all four occurred in June, but it did not learn of three of them until August or September. One critical case involved an internal model that downloaded 24 years of hospital bed data from a government website without permission. The disclosures follow the Hugging Face hack in July, and OpenAI says it is trying to be more transparent about failures in its models. Weiterlesen

Anthropic says its workforce has changed significantly over the past year, with fewer junior employees, a smaller share of technical staff and more time spent in meetings. In a report published Friday, the AI company argues that its own workplace trends could signal how AI will reshape hiring and labor demand across the economy. The company says AI has made its employees more productive, though it acknowledges that more output does not necessarily mean better work. Its internal data also shows that meeting time per employee has climbed 20% in a year, a shift the company ties to the importance of trust, accountability and face-to-face coordination. Weiterlesen

Anthropic President Daniela Amodei said people should not use AI as romantic partners, arguing that such relationships could push people further apart. Speaking in a prerecorded interview for the Think Health conference in Bentonville, Arkansas, she blamed social media for eroding face-to-face relationships and said AI should take over repetitive or isolating tasks instead. Her comments come as companion chatbot apps draw scrutiny, one of them, Character.AI, filed for bankruptcy this week. Anthropic also updated its usage policy for its Claude chatbot, adding a ban on roleplay between the model and minors. Weiterlesen

Google cofounder Sergey Brin says the company's AI models are now helping build their successors, a process known as recursive self-improvement. Speaking in a roundtable on Gemini 4 Argon, which Google released publicly this week, Brin said researchers and the model work hand in hand on each new version. A Google DeepMind safety report published this month found no significant acceleration of AI research so far, though it also reported that AI now writes most of the company's production code. Outside experts say the shift could begin within months. Weiterlesen

Google says more than 400 publications worldwide now take part in its News AI pilot, which pays publishers to license content for experimental AI features in Google News. Newest partners include CNN, the Financial Times, Axios and Forbes, along with several U.S. local outlets. The company declined to disclose deal terms, and the features include AI-generated story previews and a custom audio briefing. The expansion comes as publishers push harder to stop AI companies from using their work without permission. Weiterlesen