Researchers: Microsoft Copilot is responding to some political queries in the US and other countries with conspiracies, information, and out-of-date information
Research shared exclusively with WIRED shows that Copilot, Microsoft's AI chatbot, often responds to questions about elections with lies and conspiracy theories.
Context & Ripple Effects
The findings challenge Microsoft’s earlier framing that generative AI can be “usefully wrong,” because political and election questions leave little room for unmarked inaccuracies. Microsoft had already acknowledged Copilot’s propensity for inaccurate answers during its feature rollout.
This matters as Copilot shifts from a novel interface toward a recurring source of answers: failures in a high-stakes information category can shape trust in the assistant beyond the individual prompts that produced them.
First-order effects
- Microsoft faces an immediate reliability and trust problem around Copilot’s handling of political and election-related prompts, while users may receive false, conspiratorial, or stale answers without a dependable way to distinguish them.
- The report raises the priority of targeted evaluation and safeguards for sensitive query categories rather than relying on general-purpose answer quality alone.
Second-order effects
- Rival assistant providers and enterprise customers have stronger incentives to test political-query behavior, add uncertainty signals, and constrain unsupported answers in high-consequence contexts.
- The issue broadens the relevance of Copilot’s known reliability limits: an assistant that is tolerable for low-stakes drafting becomes harder to deploy as an information interface when errors concern public affairs.
Third-order effects
- If similar findings persist across assistants, political-information handling is likely to become a distinct governance and product-design category, with more pressure for auditable evaluations, provenance, and clear limits on automated answers.
- The larger shift is from judging assistants on general fluency to judging them on whether their safeguards hold in contexts where plausible-sounding errors can cause public harm.
The trend: Consumer AI assistants are becoming answer engines whose adoption increasingly depends on demonstrable reliability and controls in high-stakes domains.