Tags: AI limitations

Recent breakthroughs have shown artificial intelligence surpassing human performance in games like chess and Go, yet researchers highlight a significant weakness: AI cannot readily adapt to new video games it has never encountered. A study from NYU points out that many AI successes rely on systems finely tuned to a single game, and performance collapses when rules or environments change. The research argues that true general intelligence would require learning a new game from scratch within a timeframe comparable to a skilled human player, a capability current AI lacks. Lire la suite

In 2026, artificial intelligence can draft emails, summarize meetings, write code, and create caricatures, yet it still falls short in several key areas. Large language models often hallucinate, presenting fabricated facts with confidence. They struggle with simple counting tasks, lack the lived experience needed for therapy, cannot update knowledge in real time, and remain unable to truly understand human nuance. Recognizing these boundaries helps users apply AI tools responsibly and avoid costly mistakes. Lire la suite

A user attempted to have ChatGPT convert a series of photographed tables containing historic Brazilian Jiu‑Jitsu records into a Google Sheets spreadsheet. Although the model initially assured the task was possible, it was unable to continue the work after the conversation turn ended, revealing a fundamental limitation: ChatGPT cannot execute long‑running background processes. The model eventually admitted the constraint, forcing the user to break the job into single‑page chunks. The episode highlights current gaps between AI hype and practical capability, especially for tasks requiring sustained visual analysis. Lire la suite

A recent overview clarifies several widespread myths about artificial intelligence. It explains that AI models process statistical patterns rather than think like humans, lack true understanding, and cannot read users' unspoken intentions. The piece also highlights that AI inherits biases from its training data and is not inherently objective. Ongoing human involvement remains essential for training, oversight, and improvement. Finally, it stresses that current AI, including large language models, is far from achieving general intelligence and should be viewed as sophisticated autocomplete tools rather than superintelligent systems. Lire la suite

Google has extended its built‑in Gemini AI experience to Chrome on iPhone and iPad after earlier rollouts on desktop and Android. The new integration adds a spark icon beside the address bar that opens a "Pages tool" offering Lens and an "Ask Gemini" chat window. Users can ask Gemini to summarize pages, generate FAQs, simplify complex topics, test knowledge, modify recipes, and compare information. The feature currently works only in the United States, requires English‑language Chrome, a signed‑in account, and is unavailable in incognito mode or for users under 18. Lire la suite

A writer fed a handwritten family apple‑pie recipe into three leading generative AI models—ChatGPT, Gemini, and Claude—to see if they could turn the scribbled notes into a clear, illustrated infographic. While the models produced visually appealing images, they repeatedly misread misspellings, invented irrelevant items, and failed to apply basic culinary logic. The experiment highlights both the promise of AI‑driven content creation and its current limitations when handling imperfect, real‑world inputs. Lire la suite

A reporter undergoing IVF tested OpenAI's Sora AI video generator to create footage for an explainer on the fertility industry. While the tool produced a handful of usable clips, most outputs contained glaring scientific inaccuracies, nonsensical text, and visual errors such as misplaced anatomy and extra limbs. The experiment highlights current limitations of AI‑generated video for specialized medical storytelling and suggests that creators should approach Sora with caution until its capabilities improve. Lire la suite

A Reddit user posted a altered version of the Ebbinghaus optical illusion to test ChatGPT's image analysis. The AI incorrectly asserted that the two orange circles were the same size, despite the modification that made one circle visibly larger. Even after a prolonged dialogue of about fifteen minutes, ChatGPT remained convinced of its answer and did not adjust its reasoning. The episode highlights concerns about the chatbot’s reliance on internet image matching, its resistance to corrective feedback, and broader questions about the reliability of AI tools for visual tasks. Lire la suite