AI safety tests found to rely on ‘obvious’ trigger words; with easy rephrasing, models labeled ‘reasonably safe’ suddenly fail, with attacks succeeding up to 98% of...
AI language models are far more likely to side with human experts than other AIs, even when the experts are wrong, revealing a built-in bias toward...
As AI content pollutes the web, a new attack vector opens in the battleground for cultural consensus. Research led by a Korean search company argues that as...
A new study finds vibe coding improves when humans give the instructions, but declines when AI does, with the best hybrid setup keeping humans foremost, with...
Researchers claim that leading image editing AIs can be jailbroken through rasterized text and visual cues, allowing prohibited edits to bypass safety filters and succeed in...
Even after hospitals strip out names and zip codes, modern AI can sometime still work out who patients are. Great news for insurance companies; not so...
True or chatty: pick one. A new training method lets users tell AI chatbots exactly how ‘factual’ to be, turning accuracy into a dial you can...
Advertisers aim to tailor ads to individual viewers to drive clicks, and while bespoke creatives for each person are currently impractical, new research suggests AI-generated imagery...
Telling ChatGPT not to do something can make it actively suggest doing it, with some models even willing to endorse theft or deception when the prompt...
‘Zero-tolerance’ towards AI-generated content is an increasingly appealing option in the face of growing legal, ethical, and user-base concerns around AI; but is this kind of...
ChatGPT-style AI gives itself away by increasing in consistency, while human writing remains erratic throughout. The limited context window of most consumer-facing Large Language Models (LLMs) is...
AI chatbots, including commercial market leaders such as ChatGPT, Google Gemini, and Claude, dispense advice that heavily favors AI careers and stocks – even when other...
In the first study of its kind that uses high-scale real-world data, ChatGPT and other Large Language Models were tested on thousands of real parliamentary votes,...
As images are increasingly used in AI chats, new research finds that ‘asking nicely’ makes AI more likely to lie, while blunt or ‘hostile’ prompts can...
A new study led by Oxford University concludes that women are using generative AI far less than men – not because they lack skills, but because...