Just as GPT-4 was making headlines for conquering standardized tests.MicrosoftThe researchers are running a very different kind of test on other AI models -- one designed to get the models to fabricate information.
To cure the disease known as "AI Hallucinations"Symptoms, they set up a text retrieval task that would give most people a headache, and then tracked and improved the model response, an example of Microsoft's approach to measuring, detecting, and mitigating AI hallucinations.
"Microsoft wants all of its AI systems to be trustworthy and effective to use", said Sarah Bird, Microsoft's chief product officer for AI responsible for the program.
We have a lot of expertise and resources to invest in this area, so we think we can help articulate how new AI technologies should be used responsibly and enable others to do the same.
Technically speaking, an AI illusion is something that is "unsubstantiated," meaning that the AI model has altered the data it was given or added to it to describe information that wasn't there in the first place.
Of course, AI illusions are not useless all the time, for example, when users want AI to help them write a sci-fi story or give a non-traditional idea, it is beneficial; but in most scenarios where AI is needed, such as medical and educational scenarios are redundant because accuracy is more important than everything else.
As a result, Microsoft is also working to try to attack AI hallucinations based on its own AI products (e.g., Copilot), looking at a range of tools to help machines solve the hallucination problem.
Microsoft said its engineers spent months using Bing search data as the basis for Copilot through Retrieval Enhanced Generation, a technique that adds additional knowledge to a model without re-training it, to help Copilot provide more accurate and relevant responses through Bing's answer, indexing, and ranking data, as well as to provide citation information that users can find and verify on their own. The
"The model is very good at reasoning about information, but we don't think it should be the source of the answers," says Bird, "We think the data should be the source of the answers, so our first step in solving this problem was to provide the model with up-to-date, high-quality and accurate data. "
In addition, Microsoft is trying to help customers do this through various tools, such as the "Your Data" feature in the Azure OpenAI service, which helps organizations train generative AI with their own data.
It's worth noting that Microsoft has also introduced a real-time tool that can detect at scale how reliable an application is based on enterprise data. Microsoft says Azure AI Studio can assess how reliable a response is based on source documents.
According to the report, Microsoft is also developing a new countermeasure that can block and correct "unfounded" information in real time. When a factual error is detected, the feature will be automatically rewritten based on the data.
"Being at the forefront of generative AI means we have the responsibility and opportunity to make our products safer and more reliable, and to give our customers the peace of mind that they can use our tools as well," said Ken Archer, Microsoft's lead product manager for AI responsible for the program.