Why consensus is not verification: How to build AI advisors that argue productively

2026-05-18 · Source: Thomson Reuters Institute · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Corporate Strategy & Leadership · Depth: Intermediate, short

Summary

Thomson Reuters Institute's Bryce Engelland, in a May 18, 2026 article, argues that AI functions best as an executive advisor when designed to foster disagreement rather than consensus. This approach counters the risk of correlated errors and shared blind spots among AI models, which can amplify incorrectness even across different model families. Perplexity's Model Council, launched in February, exemplifies this by using divergence between Claude, GPT, and Gemini as a signal for users to exercise caution. Research, including a 2026 paper "Consensus is Not Verification," supports that multi-agent debate improves reasoning accuracy, with three agents often being an optimal configuration. Thomson Reuters enterprise architect Zafar Khan's system, featuring two AI advisors (Adrian and Elara) built on the same model but with different analytical frameworks, demonstrates how structured disagreement can reveal insights a single advisor would miss, mirroring historical decision-making methods like Socrates' cross-examination or the Delphi Method.

Key takeaway

For AI Architects and executives evaluating AI advisory solutions, prioritize systems that explicitly design for disagreement among agents. Your organization should implement multi-agent AI frameworks that surface conflicting analyses, rather than just unified responses. This approach helps identify critical blind spots and ensures human judgment remains central, especially when AI advisors diverge, prompting necessary board-level discussions and verification.

Key insights

AI advisors are most effective when designed to preserve and leverage disagreement, not just synthesize consensus.

Principles

• Consensus among AI systems does not equate to correctness.
• AI errors can be highly correlated across models.
• Best decision systems are engineered for internal conflict.

Method

Multi-agent AI systems should be intentionally built to identify and preserve meaningful disagreement, using divergence as a signal for deeper human review.

In practice

• Implement multi-agent systems with diverse analytical frameworks.
• Use divergence as a prompt for human verification.
• Assign distinct analytical roles to different AI agents.

Topics

• AI Advisory
• Multi-agent AI Systems
• Productive Disagreement
• Correlated AI Errors
• Executive Decision-Making

Best for: AI Product Manager, Director of AI/ML, Executive, AI Architect

Related on AIssential

• AI as executive advisor: Why a single “answer machine” fails — AI advisors are most effective when designed as disagreeing persona panels, not single answer machines.
• Earn your scepticism — Holding AI views through hands-on experience and intellectual honesty is crucial, not just the view itself.
• Your AI Agent Should Disagree With You Sometimes — AI agents need calibrated disagreement to decide when to act, warn, or stop, especially for real-world actions.
• The right kind of AI sceptic — The critical factor in AI engagement is *how* one holds their view—informed by experience versus adopted as identity—not the view itself.
• Who decides what AI tells you? Campbell Brown, once Meta’s news chief, has thoughts — AI models exhibit significant accuracy and bias issues on high-stakes topics, necessitating expert-driven evaluation.
• I Gave 3 AI Agents the Same Task and Made Them Argue. The Output Beat All of Them Working Alone. — Adversarial collaboration among multiple AI agents yields superior, less predictable outputs than single-agent approaches.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Thomson Reuters Institute.