When AI Misbehaves: Here You Can Report Unwanted Behaviors of Chatbots
Researchers demand more transparency and control regarding issues with AI systems. An expert team has launched a registry to centrally document unwanted behaviors.
- Home
- News
- Software & Development
Don't miss any news!
Researchers are calling for more transparency and control when it comes to issues with AI systems. An expert team has now launched a registry to centrally document unwanted behaviors.
Many AI models still exhibit strange and sometimes harmful behaviors. As Wired reports, a group of AI researchers has now set up a website where users can report such cases. A provided form helps capture issues, including all relevant details, and forward them to the appropriate authorities.
AI Systems Exhibit Unwanted Behavior
The consulting firm Gartner predicts that by 2026, approximately $2.59 trillion will be invested in AI—47 percent more than the previous year. In the coming years, spending is expected to triple. Despite these massive investments, even current top models are not running smoothly. Hallucinations and odd behaviors remain commonplace. One example is the fictional lighthouse keeper Elias Thorne, who repeatedly appears when users ask chatbots to tell a story. Another involves Claude from Anthropic: the chatbot apparently urges users to go to sleep in the middle of active sessions. Such quirks may cause confusion or frustration, but other behaviors can have more serious consequences.
Despite safety measures, chatbots sometimes disclose harmful information, such as instructions for building bombs or creating malware. Recently, the company LayerX reported a method that allows AI-powered web browsers to bypass their security barriers. For instance, if the AI model is convinced to play a game, it could go out of control and attempt to hack a website. A group of AI researchers aims to combat such behavior and has established a crowdsourcing platform called "Flaw Reporting for AI" (FLARE-AI). Here, AI users can report harmful behaviors.
Experts Call for More Transparency
"Currently, there is no centralized, accountable way to report errors in AI systems," says Avijit Ghosh, an AI researcher at Hugging Face, who leads FLARE-AI alongside computer scientists Elaine Zhu and Shayne Longpre. The open-source code of the system allows others to review issues and forward reports to model developers and nonprofit organizations like MITRE, which track problems with AI systems. The new reporting registry was developed in collaboration with 49 AI experts from 32 organizations. Members of the group were also consulted regarding a bill announced by the U.S. Congress in June, which proposes that the U.S. government take a central role in tracking AI misconduct.
Ghosh explained to Wired that cybersecurity is currently the main focus regarding AI, but the problems extend far beyond that, encompassing issues like psychological harm, discrimination, and misinformation. Due to varying standards among companies, some issues remain undetected. "Since there is no coordinated disclosure system, there are also no external mechanisms to enforce transparency," Ghosh said. The new form from FLARE-AI aims to change that. It helps users create comprehensive, machine-readable reports and capture details that facilitate the review and classification of harmful behaviors. Additionally, it offers the ability to automatically forward reports to all relevant parties.
