What changed
Across six models and five perception benchmarks, researchers found tools were sometimes called even when returned images had no causal effect, while other runs gathered useful observations under an incoherent plan.
Why it matters
A crop, zoom, or image-inspection trace does not prove that the answer was grounded in the returned visual evidence.
What to do
Counterfactually replace or corrupt returned images during evaluation; if the answer remains unchanged, remove the call or redesign the tool-use policy.
Caveat
Results are creator-run and task-dependent; the linked MIT-licensed CauAudit repository currently contains only a README and licence, not a runnable implementation.