Microsoft claims its AI safety tool not only finds errors but also fixes them

Microsoft is launching a caller diagnostic called “correction” that builds connected the company’s efforts to combat AI inaccuracies. Customers utilizing Microsoft Azure to powerfulness their AI systems tin present usage the capableness to automatically observe and rewrite incorrect contented successful AI outputs.

The correction diagnostic is disposable successful preview arsenic part of the Azure AI Studio — a suite of information tools designed to observe vulnerabilities, find “hallucinations,” and artifact malicious prompts. Once enabled, the correction strategy volition scan and place inaccuracies successful AI output by comparing it with a customer’s root material.

From there, it volition item the mistake, supply accusation astir wherefore it’s incorrect, and rewrite the contented successful question — each “before the idiosyncratic is capable to see” the inaccuracy. While this seems similar a adjuvant mode to code the nonsense often espoused by AI models, it mightiness not beryllium a afloat reliable solution.

Vertex AI, Google’s unreality level for companies processing AI systems, has a feature that “grounds” AI models by checking outputs against Google Search, a company’s ain data, and (soon) third-party datasets.

In a connection to TechCrunch, a Microsoft spokesperson said the “correction” strategy uses “small connection models and ample connection models to align outputs with grounding documents,” which means it isn’t immune to making errors, either. “It is important to enactment that groundedness detection does not lick for ‘accuracy,’ but helps to align generative AI outputs with grounding documents,” Microsoft told TechCrunch.