When AI Invents Evidence, The Damage Can Cross Borders – OpEd
Vision-language models can still affirm the presence of objects removed from images, and agentic systems can turn such unsupported outputs, prompt injections or memory-poisoned context into actions that other institutions treat as verified facts.
Recent disclosures such as the CoSnitch vulnerability in Microsoft Copilot and OWASP warnings on context poisoning show how false instructions can persist and cross organisational boundaries, yet no binding international rule currently requires an evidence trail for AI-derived claims.
The analysis proposes an “evidence passport” that separates observation from inference, demands proportionate verification for high-stakes decisions and ensures organisations can reverse or contain the effects of a contaminated action.
A wrong image description, a poisoned document and a hijacked AI assistant are different failures. The danger grows when their unsupported claims travel across organizations and acquire the power to trigger real decisions.
Researchers tried a revealing experiment: remove an object from a photograph, show the edited image to an AI model, and ask whether the object is still there. In a 2025 study of vision-language systems, tested models sometimes answered yes. The object was absent. The answer sounded certain. The gap between the two was measurable.
Now imagine that output entering a multinational logistics chain. An inspection assistant identifies a hazardous item that the underlying image does not show. Its report travels to a freight operator, insurer or customs office, and someone delays a shipment because the record seems authoritative. This scenario is hypothetical. The risk it illustrates is not: an unsupported machine inference can become a consequential fact once another institution treats it as verified evidence.
The problem is becoming more urgent as AI assistants move beyond writing summaries to searching records, maintaining memory and using tools. The weakest part of a global AI system may no longer be the model that first made an error. It may be the interface where another system accepts that error without seeing the original evidence.
Do not confuse a hallucination with an attack
The word “hallucination” invites an alluring but misleading comparison with human perception. In Yunnan, China, some people who have eaten the mushroom Lanmaoa asiatica report seeing tiny figures. A 2026 study of the fungus found no canonical genes for producing two familiar psychoactive compounds, leaving its active chemistry unresolved. A software model has no comparable intoxication or known........
