The Missing Colleague: AI and Informal Intellectual Collaboration

A scientific paper has two teams: its named authors and the invisible network of colleagueswho provide feedback, criticism, and advice. Generative AI can increasingly provide similarintellectual assistance without requiring human interaction. Does AI therefore make the invisibleteam smaller? Combining AI-text classifications with acknowledgments from 14,164 NBERand arXiv economics working-paper records, we find that AI-flagged manuscripts acknowledgeapproximately 21% fewer informal collaborators in specifications with researcher fixed effectsand extensive controls - equivalent to roughly one fewer intellectual contributor per paper. Thenegative association grows in magnitude with AI-intensity. We develop a theoretical frameworkidentifying two channels through which AI may substitute for human intellectual input throughtwo channels. First, an expertise-substitution channel: AI may replace domain-specific feedbacktraditionally obtained from colleagues working on similar research topics. Second, an access-costchannel: AI may reduce the need to seek feedback from colleagues who are more costly toreach. We find suggestive evidence for both channels. Consistent with expertise substitution,AI-flagged manuscripts acknowledge fewer same-topic contributors. Consistent with reduced costsof accessing expertise, they acknowledge fewer first-observed contributors and fewer contributorsfrom other institutions. Our findings suggest a potentially important organizational trade-off: AImay substitute for the intellectual contributions of colleagues without reproducing the professionalrelationships through which knowledge is exchanged, ideas circulate, and future collaborationsemerge.