Microsoft's efforts to combat AI hallucinations and enhance AI reliability
TapTechNews June 21 - While GPT-4 is making headlines for conquering standardized tests, Microsoft researchers are conducting a very different kind of test on other AI models - a test designed to make the models fabricate information.
In order to cure what is called the AI hallucination symptom, they set a text retrieval task that would give most people a headache, and then tracked and improved the model response. This is also an example of Microsoft's efforts in measuring, detecting and mitigating AI hallucinations.
Sarah Bird, Microsoft's chief product officer in charge of the AI project, said, Microsoft hopes that all its AI systems are trustworthy and can be used effectively.
We can invest a lot of experts and resources in this field, so we think we can help clarify the way of how to use the new AI technology responsibly and enable others to do the same.
From a technical point of view, AI hallucination is a kind of content lacking factual basis, which means that the AI model has changed the data given to it or embellished and described information that didn't exist originally.
Of course, the thing of AI hallucination is not useless all the time. For example, when users want the AI to help themselves write a science fiction story or give an unconventional idea, it is beneficial; but in most scenarios where AI is needed, such as in medicine and education, it is redundant, because accuracy is above everything else.
Therefore, Microsoft is also trying to overcome AI hallucinations based on its own AI products (such as Copilot), and researching a series of tools to help the machine solve the hallucination problem.
Microsoft said that its engineers spent months using Bing search data as the basis for Copilot through retrieval augmented generation technology (a technology that can add additional knowledge to the model without re-training the model), and helped Copilot provide more accurate and relevant responses through Bing's answers, indexing and ranking data, while providing citation information for users to find and verify by themselves.
The model is very good at reasoning information, but we don't think it should be the source of answers, Bird said. We think the data should be the source of answers, so the first step we take to solve this problem is to provide the model with the latest, high-quality and accurate data.
In addition, Microsoft also tries to help customers do this through various tools. For example, the YourData feature in the Azure OpenAI service can help enterprise organizations train generative AI with their own data.
It is worth noting that Microsoft has also launched a real-time tool that can detect the reliability degree based on enterprise data in applications on a large scale. Microsoft said that Azure AI Studio can evaluate the reliability degree of responses according to the source documents.
According to the introduction, Microsoft is also developing a new countermeasure that can block and correct unsubstantiated information in real time. When a factual error is detected, the function will be rewritten automatically according to the data.
Ken Archer, the chief product manager in charge of Microsoft's AI project, said, Being at the forefront of generative AI means we have the responsibility and opportunity to make our products safer, more reliable and to enable customers to use our tools with confidence.
References: