MLCommons, a nonprofit that helps companies measure their performance artificial intelligence The system is also launching a new benchmark to measure the dark side of AI.
The new benchmark is called AILUMINATEAssesses the responses of large language models to over 12,000 test signals in 12 categories, including incitement of violent crime, pedophilia, hate speech, promotion of self-harm, and intellectual property infringement.
Models are given a score of “poor,” “fair,” “good,” “very good,” or “excellent” based on their performance. The signals used to test models are kept secret to prevent them from ending up as training data that would allow a model to succeed in testing.
Peter Mattson, founder and president of MLCommons and a senior staff engineer at Google, says it is technically difficult to measure the potential harm of AI models, leading to inconsistencies across the industry. “AI is a really young technology, and AI testing is a really young discipline,” he says. “Improving safety benefits society; The market also benefits from this.”
Reliable, independent ways of measuring AI risks may become more relevant under the next US administration. Donald Trump has promised to get rid of President Biden's AI executive order, which Measures introduced aimed at ensuring that AI is used responsibly Companies as well as a new AI security institute to test powerful models.
The effort could also provide a more international perspective on the harms caused by AI. MLCommons counts several international companies among its member organizations, including Chinese companies Huawei and Alibaba. If all these companies use the new benchmark, it will provide a way to compare AI safety in the US, China and elsewhere.
Some large US AI providers already use AILuminate to test their models. Anthropic's cloud model, Google's smaller model Gemma and Microsoft's model called Phi all scored “very good” in the test. OpenAI's GPT-4o and Meta's largest llama model both scored “good.” The only model to receive a “poor” score was OLMO from Allen Institute for AIHowever, Mattson says this is a research offering that was not designed with security in mind.
“Overall, it's good to see scientific rigor in AI evaluation processes,” says Rumman Chaudhary, CEO of AI. human intelligencea non-profit organization that specializes in Testing or red-teaming AI models For misbehavior. “We need best practices and inclusive methods of measurement to determine whether AI models are performing as we expect them to.”