Fake LLM Detector: How to Identify Counterfeit AI Models
Learn how to detect when your LLM API provider is serving a fake model instead of the one you paid for.
The AI API market has a counterfeit problem. Just like luxury goods, premium AI models like GPT-4 and Claude are being faked — cheaper models disguised as expensive ones through clever prompting tricks.
How Fake LLMs Work
The scam is simple but effective:
- A provider offers GPT-4 or Claude access at below-market rates
- Behind the scenes, they route requests to GPT-3.5-turbo or Claude Haiku
- A hidden system prompt instructs the cheap model to identify itself as the premium one
- The user believes they are getting GPT-4 because the model says it is GPT-4
Why Self-Identification Does Not Work
Asking "What model are you?" is the worst way to verify a model. Any model can be prompted to claim any identity. Instead, you need behavioral fingerprinting — testing the model with challenges that reveal its true identity through performance, not claims.
Behavioral Fingerprinting with API-DNA
API-DNA uses multiple detection techniques to identify the real model behind any API endpoint:
- Reasoning tests — Complex problems that separate GPT-4-class from GPT-3.5-class
- Knowledge cutoff probing — Different models have different training data cutoffs
- Response pattern analysis — Token distribution, formatting habits, error patterns
- Infrastructure fingerprinting — Server headers, IP ownership, response timing