Detailed Analysis
A user comparing multiple large language models on an identical coding prompt uncovered a striking convergence in output phrasing between Anthropic's Claude Opus 4.7 and OpenAI's ChatGPT 5.3, with the two systems producing near-identical opening paragraphs in their code review responses. Claude's response began, "Solid little utility. The core idea is right: a fenced code block's fence must be longer than any backtick run inside it," while ChatGPT's opened with, "Nice little utility. The core idea is solid: choose a fence longer than any backtick run inside the text." Beyond the shared structural framing and colloquial register, the two responses employed essentially the same technical explanation unprompted, before diverging in subsequent paragraphs. The user had not expected this overlap and was primarily interested in evaluating which models would identify a specific code deficiency.
The observation prompted the user to speculate about several possible explanations: that one provider might be secretly routing requests through the other to reduce compute costs, that one model was heavily trained on the other's outputs, or that the similarity stems from shared training data. The compute-sharing hypothesis is the least technically plausible, as both Anthropic and OpenAI maintain distinct proprietary inference infrastructure and such an arrangement would represent a significant undisclosed business relationship. The training-on-outputs hypothesis is more credible as a surface-level concern, given documented instances of synthetic data generation and the widespread scraping of AI-generated content across the internet — though neither company has disclosed using the other's model outputs as direct training material.
The most probable explanation lies in the convergent pressures of large-scale human feedback training. Both models are shaped extensively by reinforcement learning from human feedback (RLHF) and related preference optimization techniques, which reward certain tonal registers, structural patterns, and pedagogical framings. When human raters consistently prefer responses that open with a brief affirmative acknowledgment, frame a core principle concisely, and then enumerate concerns, models trained on sufficient data volumes converge on those patterns regardless of their underlying architectures. The colloquial diminutive "little utility" and the explanatory colon construction are precisely the kind of stylistic micro-patterns that emerge from optimizing against shared human aesthetic preferences rather than from any direct model-to-model contamination.
This episode reflects a broader and growing concern in AI development: that RLHF-trained models are converging not just on capabilities but on surface stylistic signatures, potentially creating a kind of monoculture of AI output. The user noted that certain patterns — em dashes being a widely cited example — are already recognized as near-universal AI fingerprints. As frontier models from competing labs are trained on increasingly similar human preference data and on corpora that contain ever-larger proportions of AI-generated text, the stylistic distance between them may continue to narrow even as their underlying reasoning capabilities diverge. This presents both an evaluation challenge for users attempting to compare model quality and a deeper epistemological question about whether preference-optimized convergence inadvertently launders the biases and aesthetic quirks embedded in human rater pools into the permanent texture of AI communication.
Read original article →