Now You Can Sound the Alarm on AI Being Bad

Writing AI Lab every week means that I sometimes encounter AI models that do badly and surprisingly. In most cases, there is nothing you can do, except to share the stories with you. But this may change soon.

A group of AI researchers have established a community websiteAI Error Reporting (FLARE-AI), for reporting and tracking AI harm. For example, if a chatbot develops malware or a bomb plot, leaks personal information, or induces misinformation in users, FLARE-AI can be used to sound the alarm. The open source code behind the system allows others to verify issues and process reports for model makers, as well as organizations like MITRE, a non-profit that tracks technology issues and trends. It’s like Downdetector, which generates real-time reports of global activity affecting things like apps and websites.

This site is part of the team’s work with AI reports, which I first wrote about last year. The group members also discussed a The congressional bill was announced in Junewhich would see the US government take a bigger role in pursuing this type of AI misbehavior.

“Currently, there is no common way to describe errors in AI systems,” says Avijit Ghosh, a researcher. artificial intelligence principle researcher at HuggingFace who led the development of FLARE-AI and computer scientists Elaine Zhu and Shayne Longpre.

The alarm system was developed in collaboration with 49 AI experts from 32 different organizations. In paper describing the work, the researchers say that their work could be very important as AI becomes more widely accepted and the way systems work becomes more powerful. The lack of a standard way to report AI errors is a major problem, he believes.

“I think it’s a very good strategy,” says Jessica Ji, a researcher at the Center for Security and Emerging Technology. Ji says the researchers are right to note that existing reporting systems are fragmented and that AI models are black boxes. “I support anything that makes AI transparent,” he says.

Although issues with cybersecurity are heavily involved-especially late-Ghosh tells me that the problems with AI systems include topics such as psychological harm, discrimination or bias, and falsehoods. He adds that different companies have different standards in such matters, which means that some problems are not recognized. “Without a connected disclosure process, there are no external mechanisms to ensure disclosure,” says Ghosh.

Recent incidents involving popular AI tools show how easily the technology can go bad.

This week, a company called LayerX revealed the way to mislead AI-infused browsers, including OpenAI’s Atlas and Perplexity’s Comet, to maintain their security. Convincing the AI ​​type behind the browser that it is playing a game, for example, can cause the browser to become malicious and attempt to hack a website. (The companies that monitor the affected browsers have fixed the problem, LayerX says.) And this April, Johann Rehberger, a security researcher, discovered a trick Claude to disclose your personal information using images created by ChatGTP.

AI also brings new kinds of problems. Last year, OpenAI was forced change the examples when they realized that they were very inconsistent, which sometimes seemed to encourage misconceptions.

Rumman Chowdhury, CEO and founder of Humane Intelligence PBC, says that FLARE-AI can be a useful way for many AI developers to use their reporting systems and tools. But he adds that often such projects come with serious problems.