Experiment finds AI mostly refuses company queries without register access, not lies

A developer behind the registry-mcp tool ran a controlled experiment on September 10, 2026, testing how accurately the Claude Sonnet 5 model answered company-related questions with and without access to official national business registers. Without tools, the model hedged or declined to answer on 34 of 45 calls, was confidently wrong on just 3, and gave correct but unsourced responses on 8. However, the study flagged two dangerous failure patterns: the model once fabricated detailed regulatory deadlines for a Norwegian company, and on two trials it appended a specific employee figure after disclaiming register access — a format that automated pipelines could easily misread as factual. With register tools attached, the model passed 18 of 27 eligible test cases under a strict three-trial agreement rule, equating to a 67% pass rate, though per-trial accuracy varied by up to 18 percentage points across identical prompts on the same day.