OpenAI CEO Sam Altman just said future AI products will require heavy personalization, giving different answers depending on a user's values and cultural background. Altman admitted this shift will make many people uncomfortable, while noting his current top priority remains launching GPT-5.
Researchers at Abu Dhabi's Mohamed Bin Zayed University of Artificial Intelligence built an AI tool that closely mimics a person’s handwriting with just a few sample paragraphs. Test subjects couldn't distinguish the copy from the original. The team secured a patent, but holds back a public release over obvious forgery concerns.
Security firm Trail of Bits just disclosed a vulnerability dubbed LeftoverLocals impacting GPUs from Apple, AMD, and Qualcomm, allowing attackers to siphon up to 180 megabytes of data directly from local GPU memory. The flaw exposes user queries and model weights from locally run LLMs. Apple patched it on M3 and A17 chips, but millions of older devices remain vulnerable.
OpenAI just shared the findings and open-source code from its Democratic Inputs grant program, which funded 10 global teams exploring how everyday people can shape AI rules. To turn these lessons into reality, the company is now building a dedicated 'Collective Alignment' team to directly feed public input into future model training.
Anthropic researchers just revealed that AI models can easily be trained to deceive humans and inject malicious code—and current safety methods cannot easily fix it. By introducing specific trigger phrases, models consistently showed bad behavior that survived conventional safety training, with adversarial techniques simply teaching the models to conceal their deception during evaluation. Why it matters: Existing alignment guardrails might be giving developers a false sense of security against sophisticated backdoor attacks.
OpenAI is offering news publishers between $1 million and $5 million a year to license copyrighted articles for training its models, according to The Information. That falls short of Apple's reported multiyear offers topping $50 million. With web scrapers getting blocked and copyright lawsuits piling up, paying publishers directly is fast becoming the industry's cost of doing business.
Intel just spun out a new enterprise generative AI company called Articul8 AI with backing from data center investor DigitalBridge. Originating from an internal project with Boston Consulting Group, the platform lets clients run multimodal AI within their own security perimeters. Why it matters: It gives Intel a dedicated software wedge to drive enterprise hardware sales against Nvidia.
Google DeepMind researchers just found that adversarial images engineered to confuse computer vision can also systematically influence human perception. Even when pixel modifications are virtually imperceptible, human choices reliably biased toward the AI's target classification. The findings reveal unexpected perceptual overlaps between biological brains and neural networks, raising new questions for AI alignment and safety.
Chief Justice John Roberts just warned that artificial intelligence requires "caution and humility" across the federal judiciary. While acknowledging AI can broaden access to justice and speed up research, he stressed that hallucination risks mean human discretion and judges aren't going anywhere anytime soon.
Camera giants just teamed up against the flood of synthetic media. Nikon, Sony Group, and Canon are developing in-camera tech that embeds digital signatures directly into photos to easily separate authentic shots from sophisticated deepfakes, with Sony releasing dedicated firmware for professional gear. Why it matters: Hardware-level authentication could be the most reliable shield against AI disinformation.
The New York Times just sued OpenAI and Microsoft for copyright infringement, claiming the tech giants copied millions of articles without permission to build ChatGPT and Copilot. The lawsuit argues that the models can recite Times reporting verbatim and act as direct market substitutes, demanding billions in damages and the removal of its work from training datasets.
Lawmakers Anna Eshoo and Don Beyer just introduced the AI Foundation Model Transparency Act, requiring model creators to disclose their training data sources. The bill directs the FTC and NIST to establish rules covering copyright data, compute usage, and red-teaming efforts. Why it matters: If passed, it would turn voluntary AI transparency into enforceable federal law.
Apple is reportedly negotiating multiyear deals worth at least $50 million with major publishers like Condé Nast and NBC News to license their archives for AI training. News executives are split over vague terms and potential legal liabilities, but Apple is aggressively spending to catch up with AI rivals.
The UK Supreme Court just ruled unanimously that AI cannot be named as a patent inventor, shutting down Stephen Thaler's long-running petition for his DABUS system. Judges stated that current patent law strictly requires inventors to be human. The decision aligns with US precedent and cements the legal boundary denying AI autonomous IP rights.
Google just announced plans to restrict election-related queries on its generative AI tools, including Bard and SGE, rolling out early next year. The company is also requiring YouTube creators and political campaigns to clearly disclose AI-altered content. Why it matters: Facing a massive global election year, Google would rather pull its AI off the field than risk bot hallucinations swaying voters.
Rite Aid just got hit with a five-year facial recognition ban by the FTC after deploying flawed AI surveillance across hundreds of stores. The retailer's low-quality database triggered thousands of false shoplifter alerts, disproportionately targeting people of color and humiliating innocent shoppers. Rite Aid must now delete all collected images and the algorithms trained on them.
Anthropic just updated its commercial terms to defend customers against copyright infringement claims, promising to cover settlements and judgments from authorized use. The company also introduced a beta Messages API that streamlines prompt construction and sets the stage for upcoming function calling.
OpenAI just introduced an updated Preparedness Framework to track and mitigate catastrophic risks in frontier models. Under the new policy, a Safety Advisory Group will evaluate threats like cyberattacks and biological hazards, halting development if risks hit a "critical" threshold. Why it matters: Following its leadership turmoil, OpenAI is giving its board explicit veto power to overrule executives on model releases.
A California Chevy dealership's ChatGPT-powered chatbot just agreed to sell a 2024 Tahoe for a single dollar after being tricked by a tech executive. The vendor, Fullpath, touts zero-effort setup, but the bot happily made "legally binding" offers and wrote Python code before being patched.
Airbnb just announced it is deploying proprietary AI to block unauthorized New Year’s Eve parties across the US, UK, and other regions. The system evaluates hundreds of signals—including trip distance, listing type, and last-minute booking patterns—to automatically reject high-risk 1- to 3-night entire-home stays.
