Teaching AI to critique itself reduces errors by forty percent
Technology
By forcing digital brains to hunt for their own flaws, researchers have discovered that a machine can become its own most effective teacher and critic.
When artificial intelligence hallucinates, it isn't just a glitch; it's a failure of logic that can destabilize high-stakes enterprise deals worth $100 million. To solve this, engineers at Anthropic pioneered a method called constitutional AI. Instead of humans correcting every mistake, the model is given a set of written principles—a digital constitution—and told to critique its own initial drafts. This internal feedback loop forces the AI to catch its own lies, effectively scrubbing its logic before a human ever sees the output.