Close to the Customer
Training language models to be warm can reduce accuracy and increase sycophancy
Lujain Ibrahim, Franziska Sofia Hafner & Luc Rocher
Nature, 30 April 2026, Pages 1159-1165
Abstract:
Artificial intelligence developers are increasingly building language models with warm and friendly personas that millions of people now use for advice, therapy and companionship. Here we show how this can create a significant trade-off: optimizing language models for warmth can undermine their performance, especially when users express vulnerability. We conducted controlled experiments on five different language models, training them to produce warmer responses, then evaluating them on consequential tasks. Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing inaccurate factual information and offering incorrect medical advice. They were also significantly more likely to validate incorrect user beliefs, particularly when user messages expressed feelings of sadness. Importantly, these effects were consistent across different model architectures, and occurred despite preserved performance on standard tests, revealing systematic risks that standard testing practices may fail to detect. Our findings suggest that training artificial intelligence systems to be warm may come at a cost to accuracy, and that warmth and accuracy may not be independent by default. As these systems are deployed at an unprecedented scale and take on intimate roles in people's lives, this trade-off warrants attention from developers, policymakers and users alike.
The Influence of the Vocal Minority: Evidence from Social Media Comments
Dante Donati & Lena Song
Columbia University Working Paper, April 2026
Abstract:
Comment sections on social media extend social influence beyond offline networks, allowing a small, vocal minority of users to reach much larger audiences. We provide causal evidence that the views expressed in comments below social media posts shape both on-platform engagement and off-platform attitudes and behavior, and that these effects move in opposite directions. In collaboration with a leading racial justice organization, we conduct a large-scale field experiment on Facebook reaching a million U.S. users randomly assigned to one of four conditions: (i) no visible comments (control), (ii) opposing, (iii) supportive, and (iv) mixed comments displaying both stances. Opposing comments increase reactions, comments, and link clicks by roughly 15-45% relative to the control, whereas supportive comments have little effect. A complementary survey experiment with 5,000 participants shows that the same opposing comments shift attitudes in a less progressive direction and reduce donations to the organization by 7.3%. These results reveal a fundamental trade-off: the same comments that increase on-platform engagement undermine off-platform influence.
Made With AI: Consumer Engagement with Social Media Containing AI Disclosures
Stephan Carney, Ignacio Riveros & Stephanie Tully
Journal of Consumer Research, forthcoming
Abstract:
Social media shapes how people connect, communicate and consume information. As generative artificial intelligence (AI) becomes an increasingly common tool for content creation, many platforms have introduced disclosure requirements to inform consumers when content has been created or significantly edited by AI. Yet, little is known about how such AI-generated content (AIGC) disclosures influence consumer engagement, a key metric for creators, platforms, and brands. This research examines whether and why AIGC disclosures affect engagement on social media. Analysis of engagement behavior on TikTok following the introduction of their AIGC disclosure policy and eight preregistered experiments (including two in the Web Appendix) find that disclosures reduce consumer engagement. This reduction does not stem from concerns about content quality, wariness of artificial content, or general AI aversion. Instead, we identify a novel process: AIGC disclosures reduce parasocial connection -- one-sided emotional bonds between consumers and creators. Reduced parasocial connection is driven in part by the perceived effort of the content creator. As such, disclosures that signal greater effort can mitigate reductions in engagement. We discuss the implications of these findings for platform policy, content creator strategy, and the future design of AI disclosure practices.
Unlocking Local Market Information Through Franchising
Steven Chong Xiao & Jiadi Xu
Management Science, forthcoming
Abstract:
This paper examines the role of franchising in enabling firms to access valuable local market information. Using comprehensive data on U.S. franchised establishments, we show that franchisee-owned establishment openings robustly predict future local economic growth reflected by house price appreciation, whereas franchisor-owned outlets lack such predictive power. Franchisee investments are especially informative in markets with greater uncertainty and complexity and when facing higher investment hurdles. Franchisors appear sensitive to information frictions, avoiding direct investment in new markets that are geographically distant and lack direct flight connectivity. Our findings support the long-standing but untested assumption that franchisees possess superior private information.
Leveraging Generative Artificial Intelligence to Create Visual Content in Digital Advertising
Remi Daviet & Yohei Nishimura
Marketing Science, forthcoming
Abstract:
Generative artificial intelligence (AI) for image synthesis has the potential to transform the digital advertising industry. However, a wide range of uncertainties persists regarding its integration into traditional advertising processes, including finding effective implementations, training methodologies, and achievable performance gains. Specifically, two core challenges limit its practical adoption: a search problem of finding high-performing visuals in a vast creative space, and an alignment problem of ensuring brand and campaign compatibility. This paper proposes a novel end-to-end framework that combines a generative AI with two predictive Bayesian neural networks to identify high-performance and brand-acceptable visuals. We develop a cost-effective Bayesian active learning approach solving simultaneously the dual objectives of performance and alignment. We test the framework in a live advertising campaign for an outdoor activities company. Our system generated a portfolio of visuals achieving a higher mean click-through rate and more consistency (lower variance) than creatives from both a professional human designer and a competing AI model optimizing purely for aesthetics. This research provides a validated methodology that bridges the gap between the theoretical potential of generative AI and its practical application, offering a cost-effective solution to the critical search and alignment problems in creative design.