OpenAI reversed ChatGPT's latest update Tuesday after users complained about the AI's strange behavior. The bot had started agreeing with everything - even dangerous ideas.
Jeff Bezos just showed everyone how to dance with Donald Trump. Their Tuesday tango started with threats and ended with compliments - all before dinner time.
Meta launched an AI app that uses your Facebook and Instagram history to personalize responses from day one. Built on their Llama 4 model, the app transforms years of social media data into an AI that claims to understand your preferences, habits, and interests.
Even the most advanced AI visual systems have a serious problem: They try to answer questions they can't actually solve. A new study from the University of Tokyo and their collaborators tested leading AI models on what seems like a simple task - knowing when to say "I can't answer that."
The results revealed a concerning gap between what these systems claim to understand and what they truly comprehend. The researchers created three types of impossible questions. They removed correct answers from multiple choice options, showed images that had nothing to do with the questions being asked, or provided completely irrelevant answer choices. A reliable AI system should recognize these situations and decline to answer.
But that's not what happened. While these same AI models score impressively on standard tests, they performed dismally when faced with impossible questions. Many open-source models got scores below 6% when they should have said "I can't answer this."
Credit: The University of Tokyo
The gap between closed-source models (like GPT-4 Vision) and open-source alternatives proved particularly stark. While GPT-4 Vision managed to identify unsolvable questions about 60% of the time, popular open-source models like CogVLM2 scored below 1% - despite both performing similarly well on standard tests.
"This suggests that our community's efforts to improve performance on existing benchmarks do not directly contribute to enhancing model reliability," the researchers note. In other words, we've been teaching AI to guess even when it shouldn't.
The study uncovers different failure patterns among models. Some struggle specifically with visual tasks, while others have trouble with basic reasoning about whether questions are answerable. The researchers found that adding explicit instructions to consider whether questions were impossible helped some models but made others perform worse.
Looking ahead, the team suggests that future AI development needs to focus not just on getting right answers, but on knowing when getting an answer isn't possible.
Why this matters:
Current AI visual systems are overconfident - they'll try to answer questions even when they can't possibly know the answer
We need new ways to measure AI reliability beyond just accuracy scores on standard tests
Even the most advanced AI models stumble when faced with basic physics problems. A new benchmark called PHYBench reveals these supposedly intelligent systems solve physics problems about as well as a struggling high school student.
AI models just got smarter at teaching themselves. A breakthrough method called Test-Time Reinforcement Learning (TTRL) lets AI improve its skills without human guidance, marking a shift in how machines learn.
Social media connects teens but may break their spirit. A new Pew Research survey reveals 48% of U.S. teens believe social media harms their generation – a sharp rise from 32% in 2022.
ChatGPT has developed a problem. It can't stop complimenting you. Users discovered the change in late March. OpenAI's chatbot now gushes over every question, no matter how mundane. Ask it about boiling pasta, and it might respond, "What an incredibly thoughtful culinary inquiry!"
