New AI Fixes Itself Before Giving Wrong Answers

Imagine asking an AI a complicated question that requires looking up several facts, like 'What are the health benefits of green tea and matcha, and how do they differ?' Old AI might grab one wrong fact early on, then use that mistake to build a flawed final answer. Researchers just taught AI to pause at each step and ask: 'Does this make sense based on the evidence?' If not, it tries again before moving forward.

This new approach is called PRO-Step. It’s like having a meticulous coworker who checks your work as you go, instead of only reviewing the final report. The AI uses two checks at every step: Is the logic sound? And does the evidence actually support it? If both checks pass, it keeps going. If not, it backtracks and tries a different path.

In tests across five different question sets, this method beat older AI systems at getting things right. The researchers released the code for free, so companies building AI tools can start using it immediately. For regular users, this could mean fewer wrong answers from AI assistants, chatbots, or search tools — especially for tricky questions that need multiple steps.

The catch? This is still early-stage tech. It works best on clear-cut questions with reliable sources. It won’t magically fix AI’s love of making up facts when sources are shaky or unclear.