The question of whether a piece of text is AI-generated is increasingly being joined by a second, more specific question: which AI generated it? Model attribution - determining the source model of a given text - has moved from a research curiosity to a practical capability with real implications for compliance, fraud detection, and intellectual property.
Why Different Models Leave Different Traces
Language models are trained on different datasets, with different architectures, using different alignment techniques. These differences produce subtle but consistent stylistic signatures in their outputs. GPT-4 tends toward certain structural patterns in expository writing. Claude exhibits characteristic hedging constructions and particular approaches to enumerating options. Llama-family models have identifiable tendencies in how they handle transitions and paragraph structure. These signatures are not reliably visible to human readers, but they are detectable by trained classifiers.
Fingerprinting classifiers are trained on large corpora of labeled outputs from each model, learning to identify the combination of token-level statistics, syntactic patterns, and semantic tendencies that characterize each model's generation behavior. The task is analogous to identifying a speaker from voice characteristics - the signal is there, but it requires a trained system to surface it.
Current Accuracy and Limitations
For the major commercial models - GPT-4, Claude, Gemini - fingerprinting classifiers can achieve strong accuracy on unmodified outputs. Accuracy degrades significantly for outputs that have been paraphrased, lightly edited, or passed through a second model. It also degrades for short texts, where there is insufficient signal to reliably distinguish between models.
Open-source models present a particular challenge. The Llama family, Mistral, and their derivatives have proliferated rapidly, and fine-tuned variants can produce outputs that differ substantially from their base models in ways that confound fingerprinting classifiers trained on the base model's outputs.
Applications in Compliance and Fraud Detection
Model attribution has practical value in several contexts. For publishers and platforms operating under AI disclosure requirements, knowing which model generated a piece of content helps determine which disclosure obligations apply - particularly as regulatory frameworks begin to treat different model categories differently. For fraud detection, attribution can help identify content farms and content laundering operations that use specific models at scale.
In intellectual property contexts, model fingerprinting is relevant to claims about AI-generated content that may be derivative of specific training data. If content can be reliably attributed to a model known to have been trained on particular datasets, that attribution becomes relevant to downstream IP analysis.
Integration with Detection and Proofing
Model attribution works best as one output within a broader verification report, alongside primary AI/human classification and authenticity signals. For content flagged as AI-generated, surfacing the most likely source model - along with a confidence level for that attribution - gives reviewers additional context for triage and compliance decisions. The field is evolving rapidly, and attribution classifiers require regular retraining to remain calibrated against new model releases.