22 top AI models exposed for sharing identical writing styles across rival labs

On July 18, 2026, researchers ran a massive benchmark titled 'You’re absolutely right!' across 22 leading language models to evaluate how they handle identical prompts. Instead of analyzing high-level reasoning or safety guardrails, the team collected 584,655 characters of raw text output to run a stylistic fingerprint test.

Using a statistical tool called Fable, the study measured cross-entropy using three-character sliding windows to build a precise probabilistic map of each model's habits. This micro-level analysis bypassed superficial corporate tuning to reveal deep lexical preferences, formatting tendencies, and structural quirks embedded inside the generated text.

The statistical signatures proved that models built by competing tech giants were often more stylistically identical to rival systems than to earlier versions from their own creators. Because the underlying math compares character probability distributions, it completely ignores political stance or topic opinions, focusing purely on unconscious linguistic fingerprints.

While younger models tend to clone older ones based on release timelines, researchers caution that statistical similarity alone does not pinpoint which laboratory stole whose dataset first. The entire dataset and execution code have been published openly, inviting independent developers to verify the widespread dataset cross-contamination themselves.

Silicon Valley spent hundreds of billions building what was promised to be unique artificial intellects, only to end up with twenty-two variations of the exact same digital paraphrasing machine. The entire multi-billion-dollar synthetic intelligence arms race might literally be a single giant circular human-text recycling program.

Source: Typebulb