AI model finally learns to say ‘I don’t know’ in breakthrough to curb chatbot overconfidence
Sign up to our free weekly IndyTech newsletter delivered straight to your inboxSign up to our free IndyTech newsletterSign up to our free IndyTech newsletterSouth Korean researchers have developed a new way to finally make AI models acknowledge their unfamiliarity with topics – similar to human behaviour.The breakthrough could improve the reliability of AI models used in fields like autonomous driving and medicine, researchers from the Korea Advanced Institute of Science and Technology say.Previous research has exposed AI “overconfidence” as one of the major risks in the use of such tools to make decisions, especially in fields like medical diagnosis.Commonly used AI models like OpenAI’s ChatGPT have been shown to “hallucinate”, or make up facts, as they are incentivised to make guesses rather than admit their lack of knowledge.Now, researchers have developed a method that enables AI to recognise situations with unfamiliar or unseen knowledge, helping improve the overall reliability of chatbots.They say a fundamental cause of overconfidence in AI is the way they learn from initial data using artificial neural networks, which form their backbone infrastructure.Small errors that creep up at this stage can propagate and cause significant errors during subsequent training if they are not corrected.Researchers found that when random data was input into a neural network during the initialisation phase, the model exhibited high confidence despite not having learned anything.This led to “hallucination”.Grok, DeepSeek and ChatGPT apps displayed on a phone screen (AFP via Getty)To address this, researchers say they used clues from the way the human brain solves the issue.In humans, brain signals are generated without external input even before birth, which helps deal with the issue.Mimicking this, scientists developed a system in which the neural network backbone of an AI model underwent brief pre-training with random noise inputs before actual learning.This process, according to researchers, helps AI set a baseline for itself by adjusting its own uncertainty before starting data learning.The warm-up process can help an AI model set its initial confidence to a low level close to chance, and significantly reduce its overconfidence bias.In other words, researchers say, the method helps models first learn the state of "I don't know anything yet”.“While conventional models tend to give incorrect answers with high confidence even for data they have not encountered during training, models with warm-up training showed a clear improvement in their ability to lower confidence and recognise that they ‘do not know’,” researchers explained.This can help AI develop the ability to distinguish “what it knows" from "what it does not know"."This study demonstrates that by incorporating key principles of brain development, AI can recognise its own knowledge state in a way that is more similar to humans," Se-Bum Paik, an author of the study published in the journal Nature Machine Intelligence, said."This is important because it helps AI understand when it is uncertain or might be mistaken, not just improve how often it gives the right answer.”
Comments (0)