Researchers have discovered that Whisper, an AI-powered transcription tool used in various industries including healthcare, is prone to making up text or entire sentences, known as hallucinations. According to interviews with software engineers, developers, and academic researchers by AP, these hallucinations can include problematic content such as racial commentary, violent rhetoric, and imagined medical treatments. … Read more
Google’s NotebookLM can be fooled by manipulated websites. Developer Ted Benson demonstrated this by presenting his website to Google’s AI crawler with a made-up story about a trip to the moon with a bicycle, balloons, and scuba gear, while human visitors saw the regular page. He warns that this method of feeding LLMs with targeted … Read more
Figma has temporarily withdrawn its new AI tool “Make Designs” after it created designs for a weather app that were strikingly similar to Apple’s version. This raises the question of whether the models from OpenAI and Amazon used by Figma were trained on Apple’s designs. Figma has taken responsibility for the incident and intends to … Read more
In a previous Smart Content Report, I featured a funny illustrated guide generated by ChatGPT’s Dall-E: I find such “failures” interesting to see, because they can reveal fundamental problems. For example, we are still a long way from an AI that actually understands the world around it (“General World Model”). At the moment, these tools … Read more
A new benchmark test from Sierra shows that even advanced language models such as GPT-4o still struggle with more complicated tasks in everyday scenarios, achieving a success rate of less than 50 percent. The test, called TAU-bench, is designed to help developers evaluate the performance of AI agents in realistic situations, taking into account factors … Read more
The recently introduced “AI Overviews” in Google Search have produced some strange results – some embarrassing, some ridiculous, some dangerous. This can be seen as an example for what various experts already know and preach: Don’t let your AI work unsupervised. For example, one of Google AI’s recommendations was that cheese sticks better to pizza … Read more
AI image generators often fail because they don’t understand (yet?) what they’re creating. Mirrors are a good example. Source: Reddit
We reported on Google’s AI offensive under the “Gemini” banner, but soon after, it was the integrated image generator that made the headlines: It had apparently been steered too much in favor of diversity. What is generally a good idea makes no sense if, for example, you want a picture of the “founding fathers” of … Read more
Air Canada’s chatbot gave a customer incorrect information about the terms of a refund. In court, the airline argued that the chatbot itself was responsible for what it said, not Air Canada. The court disagreed, and the company had to pay up. Source: The Guardian