Model Identity Data Pollution / 模型身份数据污染

179: 蒸馏风暴：一场无人公开谈论的技术竞赛 adds a broader empirical warning. The source cites a 2024 study of 27 models and 77 questions to say identity confusion is common enough that a non-GPT model saying it is GPT cannot prove Model Distillation / 模型蒸馏. The episode recommends stronger Model Distillation Evidence such as output-distribution, refusal-pattern, and code-style comparisons, plus provider-side access traces where available.

Model identity data pollution is the pattern in E246｜何谓蒸馏？聊聊硅谷如何看中国开放模型逼近前沿 where a model may say “I am Claude” or “I am ChatGPT” because internet training data contains many model outputs, prompts, transcripts, and copied conversations. 王铁镇 argues that this can happen even without deliberate Model Distillation / 模型蒸馏 from a specific closed model.

The concept is useful as an evidence-quality warning. Identity confusion can indicate messy or contaminated pretraining data, weak system-prompt conditioning, or reused public assistant text, but the source says it does not by itself prove that a model’s main reasoning capability was copied from another lab.

Key Claims

• Model self-identification is not a reliable provenance test.
• Public internet data increasingly contains outputs from many AI systems, making model identity text easy to absorb during pretraining.
• Removing or weakening a system prompt can reveal identity confusion even in models not being accused of distillation.
• Provenance claims need stronger evidence such as training data access, account-call traces, model behavior audits, or reproducible evaluation.
• Identity confusion may show that AI-generated text has entered public data, but it does not locate which teacher model, account, or training pipeline was involved.
• The evidentiary standard should rise when claims are used for investment, hiring, legal, or geopolitical conclusions.

Connections

• Model Distillation / 模型蒸馏 - technical debate that identity confusion should not replace.
• Kimi K3, Claude, and ChatGPT - model names used in the source’s examples.
• AI Training Data Scarcity, Data As Education, and Model Collapse - broader data-quality and model-output recycling context.
• AI Answer Evaluation and Output Quality Gates - evaluation practices that need better evidence than surface self-labeling.
• Model Distillation Evidence, AI Model Distillation Governance, Anthropic, OpenAI, and DeepSeek - provenance and accusation context added by LateTalk episode 179.