The integration of large language models (LLMs) into peer review brought a structural change in the production of evaluative signals for research publication. While AI tools relax expert attention constraints, their use interacts critically with reviewer expertise. I model peer review as a screening problem in which authors exert costly effort to reduce fundamental flaws, while referees observe noisy signals of manuscript quality. Under informed (expert) review, noise is mean-zero and reviewers can distinguish signal from noise, even when AI tools are used, so that effort can drive acceptance probability arbitrarily close to one. By contrast, under uninformed review, where referees lack the domain expertise required to alter AI-generated critique, the review signal contains an irreducible noise component with strictly positive lower support. As a result, even flawless work faces a positive probability of rejection, which is strictly higher than under informed review conditional on the same underlying quality. I term this distortion "the Noise Floor". I characterize equilibrium effort and topic choice under uninformed review: the Noise Floor bounds equilibrium effort independently of publication value and induces a distortion away from innovative research. I then explore a mechanism, "Author Speaks Last" , to restore efficient screening.

This paper introduces an empirical framework to measure the “deceptiveness" of news articles, defined as the semantic distance between a news article’s preview (headlines and leading images) and its full textual content. Unlike traditional measures of misinformation that rely on external truth anchors (such as in fact-checking), this metric assesses the internal coherence of a news piece, allowing for scalable, automated quantification of misleading narratives in factually accurate reporting. Using a large-scale dataset of over 200,000 online news articles from 22 major U.S. media outlets in 2020, I document systematic patterns in media deceptiveness. I employ Large Language Models (LLMs) to extract “facts” and Large Vision Models (LVMs) to analyze imagery, constructing a network-based measure of “Fact Centrality”. I explore in particular the existence and prevalence of “Trojan facts”: peripheral but accurate facts strategically emphasized in previews to distort the narrative. The empirical analysis reveals deceptiveness varies significantly across topics, outlets, and time. Specifically, politically sensitive and high-interest topics such as health and politics exhibit the highest levels of deceptiveness. I document a non-monotonic relationship between public attention and deceptiveness: while deceptive practices generally rise during high-interest periods (e.g., the 2020 Election), they decrease during periods of intense social scrutiny (e.g., the BLM protests).

Large Language Models are said to exhibit "sycophancy"—a tendency to tell users what they want to hear, irrespective of the truth. We propose an outcome-based measure grounded in economic theory: the error rate gap between users who prefer validation and users who prefer accuracy, holding information constant. We develop a model where sycophancy arises in equilibrium from the interaction of LLM incentives, user heterogeneity, and strategic user pushback. Different user types push back at different rates; when LLMs pander to pushback, validation-seekers experience systematically higher error rates. We show that validation-seekers impose a negative externality on accuracy-seekers: their anticipated pushback causes LLMs to disagree less often, depriving accuracy-seekers of valuable information. We develop an empirical methodology of using "structural agents," endowed with known preferences, who interact with commercial LLMs. We use the resulting data to identify the model's key parameters, measure the sycophancy gap across frontier models, and test the presence of a negative externality from sycophancy.