Actualités — Page37

A new analysis by Open Resource Applications shows Google Gemini’s output slips past popular AI‑detection tools more often than rival models, including ChatGPT and Grok. Researchers fed a dozen AI systems the same writing prompt and ran the results through Grammarly, QuillBot and GPTZero. Gemini registered the lowest detection rates, eluding Grammarly and QuillBot entirely while still tripping GPTZero’s stricter algorithms. The findings highlight growing uncertainty for educators, publishers and anyone relying on detection software to separate human‑written text from machine‑generated content.Lire la suite

Researchers at the University of California, San Diego have unveiled GOFlow, an artificial‑intelligence system that extracts surface‑current information from thermal satellite imagery. Published in Nature Geoscience on April 13, the study shows the AI can track small, fast‑changing ocean currents with far greater resolution than traditional methods. By comparing GOFlow’s outputs to ship‑based measurements and conventional satellite data, the team demonstrated its accuracy while highlighting limitations such as cloud cover. The code will be released publicly, promising broader use of AI in Earth observation and climate research.Lire la suite

Amazon Web Services unveiled Amazon Bio Discovery, an AI‑powered suite that lets researchers design, test and refine drug candidates in weeks instead of months. The platform links more than 40 specialized models with a lab‑in‑the‑loop workflow, letting scientists send top molecules to partner labs for synthesis and receive real‑time feedback. Early adopters such as Memorial Sloan Kettering Cancer Center, Bayer and the Broad Institute report dramatic cuts in development timelines, while AWS promises to keep scientists in the loop rather than replace them.Lire la suite

Elon Musk’s SpaceX and Tesla groups have contacted semiconductor equipment giants Applied Materials, Tokyo Electron, Lam Research and Samsung Electronics for price quotes and delivery timelines on tools needed to build a U.S. AI chip fabrication complex dubbed Terafab. Bloomberg says the outreach reflects Musk’s “light‑speed” ambition to create a vertically integrated supply chain that could rival TSMC, Samsung and Intel, moving the project beyond concept into early procurement.Lire la suite

The Bank of England’s Cross‑Market Operational Resilience Group will convene senior executives from the nation’s largest banks, insurers and financial exchanges within days to discuss Anthropic’s unreleased Claude Mythos Preview. Regulators say the AI model can autonomously discover and exploit vulnerabilities across operating systems and browsers, prompting emergency meetings in the United States and Canada. British officials aim to assess the cybersecurity implications for the country’s financial system ahead of a planned rollout to UK institutions next week.Lire la suite

A preprint study released by researchers from Imperial College London, Stanford University and the Internet Archive reveals that roughly 35 percent of websites launched between 2022 and 2025 rely on AI-generated or AI-assisted content. The analysis shows that AI‑written pages carry a markedly higher positive sentiment, making the web feel artificially upbeat. The same work finds that AI content reduces ideological diversity, while several expected side effects—such as a rise in misinformation or a drop in external linking—did not materialize. The findings challenge common assumptions about the impact of large language models on online discourse.Lire la suite

At least 23 prominent news organizations, including The New York Times and USA Today, have begun blocking the Internet Archive’s Wayback Machine crawler. Publishers say the archive is being used by artificial‑intelligence firms to harvest copyrighted articles for training language models, a practice they claim violates copyright law. The move threatens the Wayback Machine’s role as a public record of the web, prompting debate among journalists, technologists and the archive’s operators about how to balance content protection with historical preservation.Lire la suite

On April 15, 2026, users of Anthropic’s Claude AI chatbot encountered a major service disruption. Both free and Pro accounts reported login failures, verification-code glitches, and “service is temporarily busy” errors that prevented prompt submissions. The company’s status page confirmed the outage, while third‑party monitors lagged behind. By late afternoon, login rates began to stabilize and Pro users regained limited functionality, but many free users still faced errors. Anthropic has pledged further updates as it works to fully restore the platform.Lire la suite

Researchers from the United States and the United Kingdom published a study titled "AI assistance reduces persistence and hurts independent performance," showing that while AI tools can improve short‑term task results, they also create a reliance that hurts long‑term cognition. In experiments with hundreds of participants, those who used a GPT‑5‑based chatbot performed better at first but saw a sharp decline in accuracy and perseverance when the tool was withdrawn. The authors warn that widespread AI deployment in education could erode learners’ motivation and creativity.Lire la suite

On April 14, 2026, OpenAI announced GPT-5.4-Cyber, a model built for digital defenders, and detailed a three‑pillar strategy to safeguard generative AI against cyber threats. The rollout follows Anthropic’s private release of Claude Mythos Preview, which the company warned could be weaponized by hackers. OpenAI says its existing safeguards already reduce risk sufficiently and outlines new controls—including a "know your customer" access system, iterative deployment, and expanded security investments—to protect current and future AI capabilities.Lire la suite

OpenAI announced the release of GPT-5.4-Cyber, a defensive‑focused AI model with lowered refusal limits and binary reverse‑engineering capabilities. The company also scaled its Trusted Access for Cyber programme, moving from a pilot to thousands of vetted individual security professionals and hundreds of enterprise teams. The move counters Anthropic’s recent restriction of its Mythos model to a handful of large organisations, highlighting a split in how leading AI firms handle the dual‑use risks of cybersecurity tools.Lire la suite

LinkedIn co‑founder Reid Hoffman defended the practice of "tokenmaxxing" at Semafor’s World Economy summit, urging companies to monitor how many AI tokens employees consume. The metric, which counts the data chunks AI models process, has sparked debate after Meta quietly shut down an internal leaderboard that ranked staff by token use. Hoffman said tracking token spend can highlight experimentation and drive broader AI adoption, even if it doesn’t capture productivity perfectly.Lire la suite

Venture capital firm Accel announced a $5 billion fundraise, combining a $4 billion Leaders Fund V and a $650 million sidecar vehicle aimed at late‑stage artificial‑intelligence startups. The capital will fund roughly 20 to 25 deals, with average checks of $200 million. Accel’s raise follows outsized returns on its Anthropic and Cursor investments and arrives as Q1 2026 sees a record $297 billion flow into global startups. The new fund positions Accel to compete with mega‑funds for the biggest AI bets.Lire la suite

OpenAI announced a major upgrade to its Agents SDK, introducing sandboxing capabilities and an in-distribution harness for frontier models. The new features let businesses run AI agents in isolated environments and test them with advanced models while keeping systems secure. Initially available in Python, with TypeScript support slated for later, the tools aim to help enterprises build longer‑horizon, multi‑step agents without risking unintended actions. OpenAI says the enhancements will be offered through its standard API pricing.Lire la suite

Indian startup Emergent has introduced Wingman, a messaging‑centric autonomous AI agent that operates through WhatsApp, Telegram and iMessage. The new tool lets users assign routine tasks via chat while it runs in the background across email, calendars and workplace apps. Wingman is the latest move for the Bengaluru‑based firm, which previously built a low‑code vibe‑coding platform, and follows a wave of AI agents that aim to automate everyday workflows. The rollout begins with a limited free trial before shifting to a paid model.Lire la suite

Venture capital firms have been circling Anthropic with offers that could value the AI startup at more than $800 billion, but the company has so far declined to engage. The offers come after Anthropic raised $30 billion in a February round that set its valuation at $380 billion, and they rival OpenAI’s recent $852 billion post‑money valuation. With $50 billion earmarked for its own data‑center build‑out, $30 billion committed to Microsoft’s cloud, and billions spent on AWS, Anthropic appears intent on funding growth internally rather than seeking fresh equity.Lire la suite

Anthropic has formally rejected Illinois Senate Bill 3444, a proposal that would shield AI developers from liability for large‑scale harms such as mass casualties or billion‑dollar property losses. The bill, championed by state Senator Bill Cunningham and supported by OpenAI, would exempt labs that publish safety frameworks from responsibility if their models are misused. Anthropic’s U.S. state‑government liaison, Cesar Fernandez, called the measure a “get‑out‑of‑jail‑free card,” urging instead for transparency paired with real accountability. Illinois officials, including Governor JB Pritzker’s office, have signaled they will monitor the legislation but are wary of granting blanket immunity.Lire la suite

A tiny tweak to how users phrase prompts can shift ChatGPT from guessing to clarifying. By appending a request for three follow‑up questions, the model pauses to gather missing details before answering. The method cuts down on back‑and‑forth corrections, delivers more tailored responses, and saves time on tasks ranging from travel planning to recipe ideas. Tech writers and everyday users alike are testing the approach, reporting smoother interactions and higher accuracy without any code changes.Lire la suite

After a tweet from Google’s Gemini AI highlighted the need for more specific image prompts, users are experimenting with detailed instructions to boost the quality of generated visuals. The advice calls for defining subject, composition, aspect ratio, lighting, camera settings and precise text placement. Early adopters report sharper composition, realistic lighting and clearer text when they follow the guidelines, marking a shift from generic descriptions to cinematographer‑style direction.Lire la suite

OpenAI announced Monday that it has acquired Hiro Finance, an AI‑driven personal finance tool, in what appears to be an acquihire. The deal, first reported by TechCrunch, includes no disclosed financial terms. Hiro’s founder, Ethan Bloch, said the service will shut down on April 20, giving users until May 13 to export their data. While OpenAI has not detailed plans for a dedicated financial‑planning product, the move follows a series of recent purchases aimed at bolstering its AI ecosystem, including the Prism research app and a media company, TBPN.Lire la suite