Why are we so fascinated by answers from AI-powered search engines? Because they can deliver something that conventional Internet searches cannot: assemble answers from found objects that are in source texts at different levels and also take the context into account. Because language models such as ChatGPT and Bard have the well-developed ability to eloquently add missing facts – although they are sometimes fictitious. Automatically checking the correctness of statements from large language models has become an urgent problem at least since the introduction of generative AI for Google, Bing and Co.
It is not without reason that even the operators of AI-supported search engines warn against believing the results without verification. But that can be a temporary solution at best. It can be assumed that intensive work is being done behind the scenes on methods to improve the factual accuracy of these systems.
Just ask the AI
Can Bing Chat, Bard, or You.com help you learn more about the methods used to do this? The answer to that is unfortunately quite sobering. Even after persistent inquiries, Microsoft’s AI search cannot be persuaded to give more than general phrases. “Bing uses machine learning to ensure that the answers to your questions are as accurate as possible. In machine learning, algorithms build a statistical model based on examples. The model is then used to make predictions or make decisions. Bing also uses artificial intelligence and human verification to ensure the answers to your questions are as accurate as possible.”
When asked which methods are actually used, the machine replies: “Artificial intelligence refers to the ability of machines to carry out human-like tasks, such as speech recognition and image recognition. Bing also uses machine learning and human verification to ensure that the answers to your questions are as accurate as possible.” It’s going in circles.
Bard, Google’s AI chatbot, is a little more forthcoming. When asked how it works, the model answers that there are “a number of algorithms” that could be used: “Text similarity algorithms, Information retrieval (IR) algorithms” (which work with a database of verified facts) and “Machine learning algorithms”. In English: “The model is trained on a dataset of questions and answers, where some of the answers are correct and some are incorrect.”
You.com goes in a similar direction: “To ensure that the generated response is supported by the facts contained in the snippets, I rely on natural language processing and machine learning algorithms that analyze relevant information from the provided snippets and extract.” The system attempts to analyze and synthesize the information in the snippets to provide an answer that is relevant and accurate to the question being asked.
Error rate 20 to 30 percent
The methods usually work quite well with sample data sets – however, 70, 80 or even 90 percent accuracy should mean for a chatbot search engine that it still produces a lot of nonsense.
Although this is not particularly detailed, it is consistent with search results from the scientific literature. From this it is known that the methods used work in a very similar way to the automated verification of fake news. That says Andreas Vlachos from the University of Cambridge, who, among other things, helped to launch the Fake News Challenge. In 2022 he published a comprehensive overview article on automated fact checking.
Accordingly, one of the standard methods in this field is to first extract the facts that are to be checked from the statement to be checked – in this case the answer from a chat-based search engine. Then you generate matching search queries. Many working groups use the Wikipedia API for this (extraction and query from databases are two sub-problems, which in turn are intensively worked on). From the extracted facts and the assertion, which are translated into vectors, the software then calculates the scalar product as a measure of verifiability – the smaller the measure, the worse the assertion is proven.
Other working groups, for their part, use models that train them to distinguish between correct and incorrect statements. During training, specially generated data sets such as FEVER are used, which contain a large number of demonstrably false statements. For special cases, such as false statements about Covid, up to 90 percent of all fake statements could be discovered – but only within a special text category.
Language model checks language model
The use of large language models to test large language models has only been increasingly discussed in the last few years. For example, Angela Fan from Meta AI and colleagues trained a language model to formulate questions from extracted statements, to research the questions using an Internet search and then to formulate answers from the results of the search. However, the system was not intended as a fully automatic fact checker. Rather, Fan and colleagues wanted to show that an automatically created summary in the form of questions and answers helps human fact checkers to assess texts more quickly and accurately.
Researchers from Tel Aviv University, Google Deepmind and Google Research recently presented a new method of fact checking for large language models. The special thing about it: The method developed by Roi Cohen and his team does not require an external knowledge base such as Wikipedia. Instead, the researchers assign a language model the role of an “examiner” (tester) who asks the “exanimee” (testee) questions about his statement. If there are obvious contradictions in the answers to these questions, the statement is rejected. The researchers describe technical details in a preprint paper.
The approach developed by Cohen and colleagues works semi-automatically: the researchers had a language model formulate the statement to be tested. In a second chat dialogue, they then assigned the role of the questioner to the same or another language model (“It is your job to check the correctness of a statement … In order to be able to do this, you can ask questions about the statement”) – which then transferred back to the first language model. In the experiment, up to five questions were allowed, then the interviewer had to judge whether he thought the statement was correct. Using their “cross-examination approach,” Cohen and colleagues achieved about 80 percent accuracy in the best cases — as expected, the results were best when the interrogator and the interviewee accessed the same models.
(bsc)