OpenAI, the company behind generative artificial intelligence (AI) ChatGPT, has reached an agreement with the US government that provides access to its new language models before they are launched, in part to identify potential risks.
Language models are software programs that can generate, depending on the case, text, images or sound on request in everyday language, like ChatGPT.
The AI Safety Institute, a body created in 2023 by the Joe Biden administration, has entered into the same agreement with one of OpenAI’s major competitors, Anthropic, according to a press release published Thursday.
The two start-ups have committed to working with the institute to “assess the capacities [des modèles] and security risks,” as well as working on “methods to manage these risks,” the AI Safety Institute said.
“These agreements are just a start, but they mark an important step in helping to guide the future of AI in a responsible manner,” said Elizabeth Kelly, the Institute’s director, in the statement.
The progression of large language models (LLM) constantly increases their capabilities but also the associated risks, particularly in the event of malicious use.
“Safe and reliable AI is critical for this technology to have a positive impact,” Jack Clark, co-founder of Anthropic, told AFP.
“Our collaboration with the Institute will draw on their experience to rigorously test our models before wide deployment,” he added.