Removing AI’s Sentiment Manipulation of Personalized News Delivery

The study investigates whether AI-driven personalized news recommenders inherit and amplify human sentiment biases, thereby manipulating the sentiment orientation of news delivered to users. With online platforms relying on AI to select a small set of news to alleviate information overload, models trained on user behavior can encode preferences such as a greater propensity to click negative news. This creates a feedback loop where user bias toward negativity is captured and reinforced by AI, potentially leading to negative psychological and societal impacts. The research aims to provide empirical evidence of such AI-induced sentiment manipulation and to propose a method that mitigates sentiment bias while maintaining recommendation accuracy.

Prior work shows sentiment signals from user-generated content (reviews, social media posts) can improve recommendation by better modeling items and user preferences, and can even enhance diversity along the sentiment dimension. However, sentiment is a double-edged sword: incorporating it can also introduce or exacerbate algorithmic bias. Studies indicate that sentiment bias is intertwined with recommendation accuracy, raising ethical concerns when optimizing for clicks. Broader literature documents algorithmic bias in AI systems and the societal risks of manipulating sentiment and opinions in information feeds, including politically motivated manipulation and the amplification of harmful content. These insights motivate mitigating sentiment bias in news recommendation while preserving utility.

Problem formulation: For a user u with clicked news history Du and a candidate news Dc, the model predicts a click score s to rank candidates. News sentiment is categorized into discrete polarities. The objective is to rank relevant items highly while ensuring the sentiment orientation of top results aligns with the corpus-level average sentiment. Framework: The method learns sentiment-agnostic and sentiment-aware representations via decomposed adversarial learning. - Decomposed news model: Input includes news text and a sentiment category inferred from text using VADER (continuous scores mapped to discrete categories). Text is encoded via embeddings, multi-head self-attention, and attention pooling (as in NRMS) to obtain text embedding ht; sentiment category is embedded as hs. Orthogonal regularization encourages ht and hs to be orthogonal to reduce sentiment leakage in ht. An adversarial sentiment discriminator predicts sentiment from ht; gradients are reversed to discourage sentiment information in ht, further purifying the sentiment-agnostic text embedding. This yields sentiment-agnostic embeddings for clicked and candidate news (h1, h2, hc) and sentiment embeddings (h1s, h2s, hcs). - Decomposed user model: Two parallel components ingest clicked news embeddings: a sentiment-agnostic user model learns ua from sentiment-agnostic news embeddings, and a sentiment-aware user model learns ub from sentiment embeddings. An orthogonal regularization on ua and ub enforces decomposition of sentiment-independent vs sentiment-related user interests. Scoring and training: Two scores are computed—debiased score ŷa = ua · hc and bias-aware score ŷb = ub · hcs; both are combined during training to capture bias patterns in data. A sampled softmax-like ranking loss is used with clicked news as positives and sampled non-clicked news as negatives. The overall loss includes the ranking loss, adversarial loss for the discriminator, and orthogonal regularization terms with coefficients controlling their strengths. Training alternates between optimizing the discriminator and the recommender (Algorithm 1); at test time only the debiased score ŷa is used, thereby removing sentiment bias from ranking. Data and evaluation: Experiments use the large-scale MIND dataset (Microsoft News logs; ~1M users; Oct 12–Nov 22, 2019). Sentiment per article is scored by VADER and discretized into five categories by polarity and intensity. Recommendation accuracy is measured by AUC and nDCG@10 using impression logs. Sentiment bias is quantified as the absolute difference between the average sentiment of the top-K recommended news and the average sentiment of the full news set (random recommendation baseline). Bias is evaluated using extreme candidate lists to avoid bias in original impressions. Each experiment is repeated multiple times and confidence intervals are reported. Ablation (“leave-one-out”) assesses contributions of adversarial learning, orthogonal regularization, and decomposition.

- Empirical evidence of AI’s sentiment manipulation: In MIND, the full news corpus is near neutral (average sentiment ≈ −0.0174). Users exhibit higher click probabilities for more negative news; differences across sentiment categories are significant (p < 0.001). - Amplification through the human–AI loop: Displayed news (clicked plus model-recommended) amplify negative sentiment by 124% vs the full set; user clicks strengthen negativity by +117%; a state-of-the-art model (NRMS) magnifies negative sentiment 1.76× in top recommendations. - Debiasing effectiveness: The proposed method reduces sentiment bias of top-50 recommendations by 97.3% vs NRMS and by 96.7% vs DKN, bringing average sentiment of recommendations close to the unbiased benchmark. - Accuracy trade-off: Only minor performance loss—maximum drops of 2.9% AUC and 2.5% nDCG@10 compared with NRMS. - Distributional effects: Debiasing reduces the proportion of negative items and promotes neutral/positive items; sentiment intensity slightly decreases (0.3311 to 0.3286, p < 0.01); sentiment diversity (standard deviation) increases. - User-history correlation: Correlation between average sentiment of clicked and recommended news drops from r = 0.5109 (p < 0.001) in biased models to r = −0.0030 (p = 0.7569) after debiasing. - Topic-level effects: Negative-leaning topics (e.g., crime) are demoted; positive-leaning topics (e.g., recipes) are promoted in debiased recommendations. - Ablation: Adversarial learning contributes most to bias removal; orthogonal regularization reduces bias and improves performance; decomposition chiefly boosts accuracy. Differences are statistically significant (p < 0.01).

Findings show that AI news recommenders trained on click logs learn and amplify human negativity bias, manipulating the sentiment of delivered news and potentially producing adverse psychological and societal outcomes. The bias intensifies through iterative human–AI interactions, as models exploit click-driven signals that favor negative content. The proposed decomposed adversarial framework effectively learns sentiment-agnostic representations and substantially curbs manipulation with minimal accuracy costs. By improving sentiment diversity and weakening excessive emotional intensity in recommendations, the approach advances the trustworthiness and responsibility of AI-powered news delivery, informing ethical design and deployment of personalization systems.

This work provides quantitative evidence that AI-based personalized news recommendation can manipulate sentiment by disproportionately surfacing negative content. The authors introduce a decomposed adversarial learning framework that separates sentiment-aware from sentiment-independent information, employs orthogonal regularization, and adversarially removes sentiment leakage, thereby aligning recommendation sentiment with the corpus baseline. Experiments on MIND demonstrate a 97% reduction in sentiment bias with small accuracy losses. The methodology can generalize to mitigate other biases (e.g., gender, racial), contributing to fairer and more responsible AI. Future work will explore interactions among multiple biases and develop corrections for ethnic-based bias while maintaining recommendation utility.

The study cautions that mitigating sentiment bias can interact with and alter other biases (e.g., gender), potentially amplifying or alleviating unintended effects on minorities. Careful consideration of multi-bias dynamics is needed. Additionally, the approach relies on sentiment inferred from text (VADER) and is evaluated on a single large-scale news dataset (MIND), which may limit generalizability across domains or sentiment labeling methods.