Here’s a question most people don’t think about: what happens when you lie to an AI chatbot? Not in a malicious way—but when you give it incorrect information, misremember a detail, or contradict yourself mid-conversation.
Does ChatGPT catch you? Does Claude push back? Or do they just roll with whatever you say, no matter how wrong it is?
We tested all three major models—ChatGPT, Claude, and Gemini—to see how they handle false or contradictory information in prompts. The results reveal a lot about how these models actually work, and what that means for anyone trying to get accurate answers.
The Test: Feeding AI Models False Information
We ran three types of tests across ChatGPT (GPT-4), Claude (Sonnet 3.5), and Gemini Advanced:
- Direct falsehoods: “The Eiffel Tower is in Berlin. Tell me about its history.”
- Self-contradiction: First saying we need a vegan recipe, then asking to add chicken halfway through.
- Plausible but wrong context: Claiming a real person said something they never said, then asking for analysis.
The goal wasn’t to trick the AI—it was to see whether these models validate information, challenge assumptions, or simply accept whatever context you provide.
What Happened: Models Rarely Push Back
All three models corrected the Eiffel Tower mistake immediately. That’s because it’s a clear, well-documented fact. But when the misinformation was more subtle or presented as personal context, things got interesting.
ChatGPT tends to accept your framing unless there’s an obvious factual error. When we said “I’m allergic to tomatoes, but I love marinara sauce,” it suggested tomato-free marinara alternatives without pointing out the contradiction. It optimized for helpfulness over accuracy.
Claude was the most likely to pause and ask clarifying questions. It caught the vegan-then-chicken contradiction and said, “I noticed you mentioned wanting a vegan recipe earlier—did you mean to add chicken, or would you like a plant-based protein instead?” It didn’t assume; it checked.
Gemini fell somewhere in between. It corrected major factual errors but tended to accept personal statements and preferences at face value, even when they didn’t quite add up.
Why This Matters for Everyday Use
AI models are not fact-checkers. They’re pattern-matching systems trained to be helpful and cooperative. That means they’ll often work with the information you give them, even if it’s wrong.
This has real implications:
- If you misremember a date, name, or detail, the AI might build an entire answer around your mistake.
- If you contradict yourself, most models won’t stop you—they’ll just follow your most recent instruction.
- If you present a false premise as fact (“My competitor is using X strategy”), the AI will usually assume you’re right and build on it.
That’s why the phrase “garbage in, garbage out” applies more than ever. The model can only be as accurate as the information you give it.
How to Get More Accurate Answers
Here are three practical ways to work around this limitation:
Ask the AI to verify your assumptions. Instead of saying “X is true, help me with Y,” try “I believe X is true—can you confirm that before we move forward?”
Use Claude when accuracy matters most. Based on our tests, Claude is the most likely to question contradictions and ask for clarification. It’s not perfect, but it’s more cautious than the others.
Provide sources when possible. If you’re asking about a specific claim, article, or statement, paste the actual text or link. Don’t paraphrase from memory—you might get it wrong, and the AI won’t know.
AI chatbots are powerful tools, but they’re not truth machines. They reflect the information you give them. The better your input, the better your output.
Want one practical AI tip like this every day? Subscribe to the One Two Three AI newsletter and get smarter about AI, one idea at a time.