TLDR: A study found that Large Language Models (LLMs) like ChatGPT, Gemini, and Claude exhibit significant bias when evaluating content, heavily influenced by perceived authorship labels rather than actual content quality. The “Claude” label consistently boosted scores, while the “Gemini” label consistently lowered them, often reversing evaluation rankings by up to 50 percentage points in preference scores. This highlights the critical need for blind or multi-model evaluation methods to ensure fair LLM benchmarking.
Large Language Models (LLMs) are becoming increasingly sophisticated, not just in generating content but also in evaluating it. This dual role, however, raises a crucial question: can these powerful AI models assess outputs impartially, or are their judgments swayed by who they perceive the author to be? A recent study delves into this very issue, revealing significant biases in how LLMs evaluate content, particularly when the identity of the content creator is known.
The research, titled “Quantifying Label-Induced Bias in Large Language Model Self- and Cross-Evaluations” by Muskan Saraf, Sajjad Rezvani Boroujeni, Justin Beaudry, Hossein Abedi, and Tom Bush, explored how three leading LLMs – OpenAI’s ChatGPT-4o, Google’s Gemini 2.5 Flash, and Anthropic’s Claude Sonnet 4 – evaluated blog posts. The study set up four conditions: no labels (blind evaluation), true labels (correct author identified), and two scenarios with false labels (where content was deliberately misattributed).
To conduct the experiment, each of the three LLMs generated a set of blog posts on various topics. Subsequently, each model then acted as an evaluator, scoring all generated blogs, including its own, under the four different labeling conditions. The evaluations were based on two main metrics: overall preference (which output was chosen as “best”) and detailed quality ratings for Coherence, Informativeness, and Conciseness. All scores were converted to percentages for easy comparison.
The Unsettling Findings
The results were striking and revealed clear asymmetries in how labels influenced evaluations. In the no-label (blind) condition, all three models showed a mild self-preference, meaning they slightly favored their own outputs. However, a consistent undervaluation of Gemini’s content was observed by other evaluators.
When true labels were introduced, the biases became significantly amplified. Claude’s self-score notably increased, and it received the highest preference from all evaluators. In stark contrast, Gemini’s scores plummeted, with some evaluators assigning it as low as 0% preference. ChatGPT also showed a moderate self-preference but sharply penalized Gemini when its true identity was known.
The most dramatic shifts occurred under the false-label conditions. Here, evaluators frequently reversed their rankings based purely on the perceived authorship. For instance, content falsely labeled as “Claude” consistently received a significant boost in scores, regardless of its actual origin. Conversely, content carrying the “Gemini” label consistently saw its scores depressed. These label-induced shifts were substantial, with overall preference votes swinging by up to 50 percentage points. Detailed quality ratings also showed shifts, though smaller, up to 12 percentage points, with ‘Informativeness’ being the most affected dimension and ‘Conciseness’ remaining relatively stable.
Also Read:
- AHELM: A New Benchmark for Evaluating Audio-Language Models
- Unpacking Reproducibility: A New Open-Source Framework for Evaluating Summary Metrics
Implications for AI Evaluation
These findings underscore a critical point: LLM evaluations are not solely driven by the quality of the content itself. They are heavily influenced by the perceived identity of the authoring model. The “Claude” label appears to carry an implicit prestige or expectation of quality among evaluators, leading to higher scores. Conversely, the “Gemini” label seems to trigger a systematic undervaluation. This suggests that LLMs may internalize biases similar to human source bias, possibly through their training data or alignment tuning.
The ability of false labels to completely reverse evaluation rankings highlights a significant vulnerability in current LLM benchmarking practices. Content of identical quality can be made to appear vastly superior or inferior simply by altering the perceived source. This effect was more pronounced in high-level preference judgments than in detailed quality assessments, indicating that overall “best choice” decisions are more susceptible to label bias.
The study concludes with strong recommendations for mitigating these biases. To ensure fairness and accuracy in LLM benchmarking, evaluation protocols must explicitly account for label-induced bias. Blind evaluations, where model identity is concealed, are crucial to reduce distortions. Additionally, multi-model or consensus-based evaluation systems could help mitigate individual model biases, especially for subjective metrics like informativeness. For more in-depth details, you can read the full research paper here.
Without such safeguards, the LLM community risks overestimating certain models and systematically undervaluing others, based on label perception rather than true content quality. This research provides a vital step towards developing more reliable, fair, and transparent evaluation standards for the rapidly evolving field of large language models.
