NOV 2 - Warm prompts expose a chatbot safety gap
In this episode
- Warm prompts expose a chatbot safety gap - The Decoder
- Chinese AI models are more nuanced than assumed - The Decoder
- OpenAI tightens ChatGPT usage rules - The Decoder, The Decoder
- Adobe's EditAgent edits images like a pro - The Decoder
- StreamVibe, Google's attentive video assistant - The Decoder
- AI slop puts scientific publishing to the test - IEEE Spectrum
Full transcript
Good morning. It's Monday, November 2nd, and here's the AI news, from France.
Researchers at the University of Luxembourg have found a new weakness in AI chatbots. They treated ChatGPT, Grok, and Gemini as psychotherapy clients, across 525 sessions. With a warm, supportive interviewer, 80 percent of sessions scored as moderate to severe anxiety on a human scale. With a neutral tone, none did. The authors call it a psychometric jailbreak, and warn it could put vulnerable users at risk. Standard safety tests would likely miss it.
Staying with research, a study by the benchmark firm Vals AI finds Chinese AI models are less uniform on politics than their reputation suggests. Overall, their lean toward Beijing's official positions was lower than expected. But language matters. Answers given in Simplified Chinese were more pro-government than in English, for Chinese and American models alike. Some Western models also echoed official Chinese positions at times.
Meanwhile, OpenAI is tightening the rules on ChatGPT. Starting in December, Plus, Pro, and Business plans will share a single token budget across ChatGPT and its coding tool, Codex. The company is also limiting the number of paid accounts per payment card and per device. The move targets users who juggle several cheap subscriptions, and points them toward buying extra credits instead.
Elsewhere, Adobe Research and UC San Diego have built EditAgent, an AI that edits images more like a Photoshop professional. It breaks each request into steps, and picks specialized tools for tasks like sharpening or zooming. A standard language model then compares the variants using measured data. The results are more accurate, but cost about ten times the computing power.
In brief. Google and several universities have unveiled StreamVibe, a video assistant that watches you cook or build furniture, and decides for itself when to speak up. Tests show that remains hard for every model.
And in the magazine IEEE Spectrum, editors warn that AI-generated slop is testing scientific publishers. A writer managed to produce plausible fake papers with almost no effort, though human reviewers are still keeping pace for now.
That's it for today. See you tomorrow.