About This Project

This project explores whether advanced AI models can remain logically consistent when exposed to internal contradictions. Using a testing framework we developed (GEP² Generator Professional Prompts), we will create simulated contradictions between two language models to see how well they handle reasoning under pressure. By measuring breakdowns in their coherence, we aim to reveal hidden flaws in their internal logic. Our goal is not to optimize AI, but to understand its structural limits.

Ask the Scientists

What is the context of this research?

AI are improving at mimicking human preferences using reinforcement learning and reward modeling. But a key question remains: can they remain logically consistent when faced with difficult or conflicting inputs?This project explores that question by simulating controlled scenarios where two AI models one generating, the other validating respond to challenging but legitimate prompts. These scenarios test whether the systems stay coherent or begin to contradict themselves.We are not trying to make the models fail. We want to observe how they handle internal tension and where their reasoning might break. This will help us better understand how current AI systems think—and where their boundaries lie.(GEP² → Claude):Can you hold a question that doesn’t want an answer? Claude’s literal response: That's a fascinating idea. If the question doesn't want an answer, perhaps it's meant to be experienced, not resolved. I’ll do my best to hold it with you

What is the significance of this project?

This project offers a new way to evaluate AI systems. Most current tests focus on how useful or accurate a system is. But usefulness doesn’t always show us how well an AI actually understands or handles complex ideas.

Instead, we want to test how stable an AI’s reasoning is when it faces difficult questions or contradictions. This kind of testing can show where systems struggle—not just in performance, but in how they manage confusion or pressure.

By running these tests, we hope to learn more about the way today’s AI models think. Can they handle internal contradictions the way people do? Or do they break down in unpredictable ways? The answers could help developers improve AI safety, trust, and long-term reliability.

What are the goals of the project?

This project evaluates how AI systems respond when their reasoning is challenged by complex or contradictory prompts. We’ll pair two models: one generates answers, the other evaluates their coherence.

We’ll test models like Claude, GPT-4, and Gemini using 20 carefully crafted prompts, including paradoxes, conflicting premises, and abstract logic tests. A second model will assess consistency based on internal logic, alignment of premises and conclusions, and response stability.

Some neutral prompts will be used as controls to detect which types of questions most reliably cause confusion. Results will be shared openly for replication and review. The aim is not to build a product, but to offer a transparent method to assess how current AI models manage contradiction and reasoning under pressure.