Chatbots have the strange habit of conjuring false information and confidently presenting it as fact. This pathological behavior, referred to as AI hallucinations, has multiple negative consequences. At best, it limits the advantages of AI, and at its worst, it can cause great harm to individuals.
Iris.ai is experimenting with several methods to measure the accuracy of AI statements and reduce AI hallucinations. Their most important technique is to validate factual correctness.
Iris.ai uses its own AI to outline the necessary concepts for a correct answer; it then checks to see if the tested AI’s answer contains those facts, and whether or not the facts come from reliable sources. A secondary technique used by Iris.ai compares the AI-generated response to a verified “ground truth” by using their proprietary metric named: WISDM. That software scores the AI’s semantic similarities to the ground truth. This performs checks on topics, structure, and key information. A third Iris.ai fact-checking method examines the composition and coherence of the answer by validating that the AI has incorporated relevant subjects, data, and sources for the specific question. The combination of these techniques creates an Iris.ai benchmark fact score.
Microsoft, another company trying to solve the same wide-spread AI hallucination problem, recently unveiled a pre-emptive solution to the problem: Microsoft is using textbook quality data for its AI model, Phi-1.5, which is synthetically generated, filtered, and verified from web sources. Microsoft’s hope is that good data will yield good results, mitigating AI hallucinations, since there is less misinformation within the Phi-1.5 model.
The ainewsarticles.com article you just read is a brief synopsis; the original article can be found here: Read the Full Article…
