Experts find flaws in hundreds of tests that check AI safety and effectiveness | Artificial intelligence (AI)

Experts have found weaknesses, some serious, in hundreds of tests used to check the safety and effectiveness of new artificial intelligence models introduced to the world.

British government computer scientists AI Security Instituteand experts at universities including Stanford, Berkeley, and Oxford reviewed more than 440 criteria that provide an important safety net.

They found flaws that “undermine the validity of the resulting claims,” that “nearly all…have weaknesses in at least one area,” and that the resulting scores may be “irrelevant or even misleading.”

The study’s lead author, Andrew Bean, a researcher at the Oxford Internet Institute, said many of the criteria are used to evaluate the latest AI models released by big tech companies.

Due to the lack of nationwide AI regulation in the UK and US, benchmarks are used to check whether new AIs are safe, in line with human interests and achieve claimed abilities in reasoning, mathematics and coding.

The investigation into the tests comes at a time when concerns are growing about the safety and effectiveness of artificial intelligence being released at high speed by rival technology companies. Some have recently been forced to roll back or tighten restrictions on AIs that contribute to harm ranging from insults to character to suicide.

“Benchmarks support almost all claims about advances in artificial intelligence,” Bean said. “But without common definitions and robust measurements, it becomes difficult to know whether the models are actually improving or continuing to improve.”

On Google this weekend Withdrew Gemma, one of its latest AIsAfter making false claims, including fake links to reports that a U.S. senator had non-consensual sexual intercourse with a state trooper.

“There has never been an accusation like this, there is no person like this, and there are no new stories like this,” Marsha Blackburn, a Republican senator from Tennessee, wrote in a letter to Google CEO Sundar Pichai.

“This is not a harmless hallucination. This is an act of defamation generated and distributed by a Google-owned AI model. A public tool that invents false allegations of wrongdoing about a sitting U.S. senator represents a disastrous failure of oversight and ethical responsibility.”

Google said Gemma models are designed for AI developers and researchers, not for fact-based help or consumers. It withdrew them from the AI ​​Studio platform after what it described as “reports of non-developers trying to use them.”

“Hallusions, where models simply make things up about all kinds of things, and flattery, where models tell users what they want to hear, are challenges in the AI ​​industry, especially smaller open models like Gemma,” he said. “We are committed to minimizing hallucinations and continually improving all of our models.”

Last week, popular chatbot startup Character.ai banned teens from having open-ended conversations with its AI chatbots. This follows a series of controversies, including a 14-year-old boy in Florida who killed himself after becoming obsessed with an AI-powered chatbot that his mother claimed had manipulated him into taking his own life. US lawsuit from a teenager’s family A person who claims that a chatbot manipulated him into self-harm and encouraged him to kill his family.

The research examined commonly available criteria, but leading AI companies also have their own internal criteria that were not examined.

The report concluded that there is “an urgent need for common standards and best practices.”

Bean said a “shocking” finding was that only a small minority of criteria (16%) used uncertainty estimates or statistical tests to indicate the likelihood of a criterion being correct. In other cases where benchmarks set out to evaluate the properties of an AI (e.g. “harmlessness”), the definition of the concept under study was controversial or ill-defined, making the benchmark less useful.