As organizations increasingly use AI, the real challenge is knowing when its answers can be trusted. The tweedledumming technique uses two AI systems to check each other's work before relying on the results. It requires only two browser tabs and a repeatable process, making a proven approach for reducing AI hallucinations practical for anyone.

Artificial intelligence (AI) can summarize reports, draft policies, analyze research, and generate ideas in seconds. AI has a well-known weakness, however; it often presents guesses as facts. AI researchers call this a hallucination, that is when an AI generates information that sounds authoritative but is unsupported or incorrect.

Hallucination is not a rare glitch (Du et al., 2024). It is a predictable byproduct of how these systems work. Most advice for reducing hallucinations focuses on writing better prompts or providing more context. Those techniques help - better prompts improve the input - but there is another practical approach that plays one AI against another in a second, independent review.

We call it “Tweedledumming,” playing one AI against another. This method is a manual version of what researchers call Multi-Agent Debate (MAD) (Liang et al., 2024). Unlike most research implementations, it requires no programming, APIs, or technical expertise. Anyone with two browser tabs can do it.

The Problem: AI Fills in the Blanks

A single example shows how quickly AI substitutes assumptions for facts. Suppose you ask an AI:

How do we make sure my project follows applicable rules and laws, delivers data for my organization's performance metrics, and achieves my organization's goals?

The AI may return a polished answer covering legal compliance, accessibility, governance, and performance measurement. It sounds authoritative, but, unless you identify your organization, the AI has no way of knowing what country or jurisdiction applies, whether you work in government, a nonprofit, or the private sector, which laws are relevant, or what your organization's goals actually are. Instead of admitting what it doesn't know, the AI often fills in the blanks with plausible assumptions.

You have two bad options at this point: accept the risk of hallucination, or fact-check every sentence yourself. Tweedledumming offers a third option; recruit a second AI to fact-check the answer and generate corrective prompts that you copy-paste to the first AI so that it provides a more factual response.

A Different Way to Use AI

Instead of accepting the answer or checking every statement yourself, ask a second AI to review it. Provide AI #2 with your original prompt and AI #1's response. Then, ask it to identify unsupported assumptions or hallucinations, explain where the reasoning overreaches, and generate a corrective prompt addressed directly to AI #1. The critical step is asking AI #2 to write to the first AI, not to you. The result is a correction that can be copied directly back into AI #1. We outline how to do this below.

What Happens Next

The review itself is only half the loop. The payoff comes when AI #2's correction lands back in AI #1's conversation.

When the corrective prompt is pasted into AI #1, the response often changes dramatically. Instead of making confident assertions, the AI begins asking for missing context. It will state: "The answer depends on your jurisdiction, organization, funding source, users, and applicable policies." The response becomes narrower, more transparent about uncertainty, and ultimately more useful. You can then send the revised answer back to AI #2 for another review.

One round exposes the obvious knowledge gaps. Pasting the responses back and first between the two AIs drives out the subtler ones.

After several rounds, the two systems often converge on recommendations that are better supported and contain far fewer unsupported assumptions. Neither AI becomes smarter. Rather, each simply helps expose weaknesses in the other's reasoning.

Why It's Called Tweedledumming

Convergence through mutual challenge is not a new idea; researchers have formalized it, and Lewis Carroll anticipated it.

MAD pits multiple AI systems to critique one another before producing a final answer. Tweedledumming applies the same principle manually; instead of software coordinating multiple AI agents, you simply act as the intermediary between two conversational AI systems. Our approach to this manual workflow was inspired by Tweedledum and Tweedledee from Lewis Carroll's Through the Looking-Glass. Although Carroll never portrayed them as truth-tellers and liars, the characters have long appeared in logic puzzles as mirrored characters who challenge each other's claims. The name reflects the method: Neither AI is trusted by default and each becomes the other's skeptical reviewer.

The Tweedledumming Method

  1. Ask your question to AI #1.
  2. Give AI #2 your original prompt and AI #1's response.
  3. Ask AI #2 to identify unsupported assumptions and generate a corrective prompt addressed directly to AI #1.
  4. Paste that correction into AI #1.
  5. Send AI #1's revised response back to AI #2.
  6. Repeat until both systems largely agree.
  7. Ask AI #2 to answer the original question independently.
  8. Compare both final responses and combine the strongest validated elements.

Why It Works

Different language models make different assumptions and notice different weaknesses. When you ask one AI to critique another rather than simply answer the same question, it becomes easier to identify unsupported assumptions, lies, logical fallacies, incomplete context, and “overconfident” conclusions. This process encourages both models to replace speculation with explicit uncertainty and to request the information they actually need before making recommendations.

A caveat, however: The Tweedledumming method raises the floor on quality. It does not remove the ceiling on risk. It does not guarantee correctness, and a human in the loop is required since judgment remains essential, particularly for legal, regulatory, medical, scientific, or policy decisions. You may need to consult an expert resource as a result, but this first pass provides an effective quality-control step before AI-generated work reaches a human decision-maker.

References Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., & Mordatch, I. (2024). Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning (ICML'24), Vol. 235. JMLR.org, Article 467, 11733–11763. https://dl.acm.org/doi/10.5555/3692070.3692537

Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Tu, Z., & Shi, S. (2024, November). Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 17889–17904. https://aclanthology.org/2024.emnlp-main.992/