TL;DR
Recommender systems exhibit a significant "source bias," consistently ranking AI-generated content (AIGC) higher than human-written equivalents. This creates a dangerous feedback loop: as users interact with more AIGC, the models retrain on this biased data, further amplifying the preference and eventually leading to a decline in overall recommendation quality. The authors propose a novel debiasing method using L1-loss optimization to align the embedding spaces of human and AI content, restoring balance to the digital ecosystem.
What Problem Does This Paper Address?
As Large Language Models (LLMs) flood the internet with high-quality, low-cost text, recommender systems (RS) have become the primary gatekeepers for this content. However, recent studies suggest these systems aren't neutral; they often suffer from source bias, where the underlying models (like BERT or RoBERTa) favor the stylistic patterns of LLM-generated text over human writing, even when the semantic meaning is identical.
The authors identify a critical "feedback loop" problem. If a system ranks AIGC higher, users are more likely to click it. This interaction data then becomes training data for the next version of the model, which learns to favor AIGC even more strongly. This isn't just a fairness issue for human creators; the paper warns that an ecosystem dominated by AIGC can eventually lead to "model collapse," where the recommendation performance itself degrades due to the lack of diverse, human-originated data.
Methods
The researchers explored the problem across three phases: HGC Dominate (the past), HGC-AIGC Coexist (the present), and AIGC Dominate (the future).
To study the long-term effects, they built a realistic simulator using various click models (like the Position-Based Model) to mimic how users interact with ranked lists. They then developed a black-box debiasing method that doesn't require the system to explicitly "detect" AI content. Instead, their approach involves:
- Uniform Rewriting: Prompting an LLM to rewrite all training data into AIGC copies.
- Embedding Alignment: Using an L1-loss function to minimize the distance between the embeddings of original human-generated content (HGC) and its AIGC counterpart.
- Entropy Constraint: Adding a constraint to ensure that while HGC and AIGC versions of the same item are close together, different items remain distinct in the embedding space to preserve ranking performance.
Datasets
The study utilized the Amazon product dataset, focusing on three diverse categories: Health, Beauty, and Sports.
- AIGC Generation: They used mainstream LLMs—ChatGPT (gpt-3.5-turbo), Llama 2, Mistral, and Gemini Pro—to rewrite product descriptions while maintaining semantic equivalence.
- Verification: A human evaluation study confirmed that users generally found the HGC and AIGC descriptions to be of equal quality, ensuring that any ranking preference found was due to model bias rather than a genuine difference in content utility.
Key Insights and Ideas
- The "AIGC Premium": Even when an LLM-generated description is semantically identical to a human one, popular sequential models (like SASRec and BERT4Rec) consistently rank the AI version higher.
- Bias Escalation: The bias isn't static. In the "AIGC Dominate" phase, the preference for AI content is pushed to the top of the ranking list through the iterative feedback loop.
- Performance Degradation: There is a "tipping point." While a little AIGC might not hurt, a content ecosystem dominated by it (e.g., after 20 iterations of the feedback loop) results in a measurable drop in recommendation accuracy (NDCG).
- Alignment over Detection: The most effective way to fix the bias isn't to ban AI content, but to force the model's internal representations (embeddings) of HGC and AIGC to be indistinguishable.
Key Results
- Bias Measurement: The authors used "Relative $\Delta$" to measure bias. In the Health dataset, models like GRU4Rec showed a strong initial preference for AIGC (Relative $\Delta \approx -23\%$).
- Feedback Loop Impact: Without intervention, the bias amplified significantly over 10 iterations. In some scenarios, the absolute Relative $\Delta$ increased by over 100 points.
- Debiasing Success: The proposed L1-loss method successfully reduced the Relative $\Delta$ to near zero (or even slightly positive, favoring HGC) across all datasets and models.
- Performance Boost: Interestingly, the debiasing method often improved standard recommendation metrics (like NDCG@3). By forcing the model to ignore "source noise," it became better at identifying true semantic relevance.
Limitations and Future Work
The authors acknowledge that their study primarily focuses on textual AIGC. As multimodal content (AI-generated images and videos) becomes more common, the source bias in those domains—and how they interact with text—remains an open question. Additionally, the study assumes that the LLM-rewritten content is semantically equivalent to the human original; however, in the real world, LLM "hallucinations" could introduce factual inaccuracies that this specific debiasing method might not address. Future work could explore more complex user behavior models and the impact of AIGC on non-sequential recommendation architectures.
Summary updated