Publications
Recent preprints (2) · Published (9) · Under review (5) · In progress (7)
Recent preprints
-
Human-AI Mutualism Is Unstable under Capability Asymmetry
Abstract
Hoping that humans and AI will need each other, and so keep the relationship good for both, is no substitute for alignment. In a game-theoretic model, humans stay as well off as before AI only while they have leverage over it; once they have none, only alignment protects them, and against an AI with insatiable goals the alignment required approaches perfect alignment as capability grows.
Cite
@unpublished{mutualism-asymmetry, author = {Mohseni, Aydin}, title = {Human-AI Mutualism Is Unstable under Capability Asymmetry}, year = {2026}, note = {arXiv:2611.23322}, } -
Why and How to Do Science When AI Is Better at It: A Framework Grounded in Human Understanding and Direction
With Hoda Heidari.
Abstract
Science should keep aiming at shared human understanding, and that aim decides which research tasks AI can take over. From this aim the paper derives a two-part test, human understanding and human direction, for when a research task can be delegated to AI, and applies it to each phase of research.
Cite
@unpublished{science-ai, author = {Heidari, Hoda and Mohseni, Aydin}, title = {Why and How to Do Science When AI Is Better at It: A Framework Grounded in Human Understanding and Direction}, year = {2026}, note = {arXiv:2610.36116}, }
Published
-
Why are the Social Sciences Hard? Two New Hypotheses
Studies in History and Philosophy of Science, with Daniel A. Herrmann and Gabe Avakian Orona.
Abstract
Proposes two reasons the social sciences are hard: researchers are committed to a fixed set of prediction targets and cannot switch to more tractable ones, and human baseline performance on easy behavioral predictions is already high. Offers a formal framework for judging how hard a discipline is and how much its achievements mean.
Cite
@article{social-sciences-hard, author = {Mohseni, Aydin and Herrmann, Daniel A. and Orona, Gabe Avakian}, title = {Why are the Social Sciences Hard? Two New Hypotheses}, journal = {Studies in History and Philosophy of Science}, volume = {119}, pages = {102197}, year = {2026}, doi = {10.1016/j.shpsa.2026.102197}, } -
The Role of Modeling in Improving Science
Philosophy Compass, forthcoming (commissioned).
Abstract
A critical survey of formal models of the social structure and epistemic properties of science, and of the assumption that such models support meaningful inferences about them: what this research has shown, what it faces, and the pressing open questions.
Cite
@article{role-of-modeling, author = {Mohseni, Aydin}, title = {The Role of Modeling in Improving Science}, journal = {Philosophy Compass}, note = {Forthcoming}, } -
Methods for Modelers of Science
In Methods in the Philosophy of Science: A User’s Guide.
Abstract
A guide for philosophers and metascientists who model science: the implicit know-how of modeling, and the challenges and good practices of studying the social and epistemic structure of research.
Cite
@incollection{methods-for-modelers, author = {Mohseni, Aydin}, title = {Methods for Modelers of Science}, booktitle = {Methods in the Philosophy of Science: A User's Guide}, editor = {Veigl, Sophie J. and Currie, Adrian}, publisher = {The MIT Press}, pages = {485--506}, year = {2025}, doi = {10.7551/mitpress/15452.003.0036}, } -
The Best Paper You’ll Read Today: Media Biases and the Public Understanding of Science
Philosophical Topics, with Cailin O’Connor and James Owen Weatherall.
Abstract
Computational models of how journalists’ selection and sharing of scientific evidence shape public belief. Standard practices can distort belief substantially, existing errors can compound, some practices that look harmless do real epistemic harm, and some drive polarization.
Cite
@article{best-paper, author = {Mohseni, Aydin and O'Connor, Cailin and Weatherall, James Owen}, title = {The Best Paper You'll Read Today: Media Biases and the Public Understanding of Science}, journal = {Philosophical Topics}, volume = {50}, number = {2}, pages = {127--153}, year = {2022}, doi = {10.5840/philtopics202250220}, } -
The Tragedy of the AI Commons
Synthese, with Travis LaCroix.
Abstract
Compliance with AI ethics guidelines as a social dilemma, modeled with stochastic evolutionary game dynamics, and the conditions under which cooperation among developers can hold. Stochastic effects can make cooperation viable, which suggests coordinating in smaller groups where cooperation is cheap and the risk of failure is high.
Cite
@article{ai-commons, author = {LaCroix, Travis and Mohseni, Aydin}, title = {The Tragedy of the AI Commons}, journal = {Synthese}, volume = {200}, number = {4}, pages = {289}, year = {2022}, doi = {10.1007/s11229-022-03763-2}, } -
Truth and Conformity on Networks
Erkenntnis, with Cole Randall Williams.
Abstract
Agents learn by Bayesian conditionalization from private evidence and from what their neighbors say, while trading accuracy against conformity. Networks that sustain a diversity of opinion support honest communication and reliable true belief; less centralized and less connected networks do best.
Cite
@article{truth-and-conformity, author = {Mohseni, Aydin and Williams, Cole Randall}, title = {Truth and Conformity on Networks}, journal = {Erkenntnis}, volume = {86}, number = {6}, pages = {1509--1530}, year = {2021}, doi = {10.1007/s10670-019-00167-6}, } -
On the Emergence of Minority Disadvantage: Testing the Cultural Red King Hypothesis
Synthese, with Cailin O’Connor and Hannah Rubin.
Abstract
Tests whether a difference in group size alone can produce inequitable outcomes for a minority, without explicit or implicit bias. In experiments with 112 participants playing a repeated Nash demand game, members of the minority group made low demands more often and received lower payoffs.
Cite
@article{red-king, author = {Mohseni, Aydin and O'Connor, Cailin and Rubin, Hannah}, title = {On the Emergence of Minority Disadvantage: Testing the Cultural Red King Hypothesis}, journal = {Synthese}, volume = {198}, number = {6}, pages = {5599--5621}, year = {2021}, doi = {10.1007/s11229-019-02424-1}, } -
Self-Assembling Networks
The British Journal for the Philosophy of Science, with Jeffrey A. Barrett and Brian Skyrms.
Abstract
How an epistemic network might form from the reinforcement-learning decisions of inquirers who differ in ability. Such networks can outperform solitary inquiry and reduce epistemic risk; the paper compares them under perfect and imperfect communication.
Cite
@article{self-assembling-networks, author = {Barrett, Jeffrey A. and Skyrms, Brian and Mohseni, Aydin}, title = {Self-Assembling Networks}, journal = {The British Journal for the Philosophy of Science}, volume = {70}, number = {1}, pages = {301--325}, year = {2019}, doi = {10.1093/bjps/axx039}, } -
Stochastic Stability and Disagreements between Dynamics
Abstract
The replicator dynamics and the Moran process can make different predictions. The paper traces the divergence to standard analytic techniques and to the idealizations each model makes, identifies problems for stochastic stability analysis in a broad class of games, and identifies a domain where the two dynamics agree.
Cite
@article{stochastic-stability, author = {Mohseni, Aydin}, title = {Stochastic Stability and Disagreements between Dynamics}, journal = {Philosophy of Science}, volume = {86}, number = {3}, pages = {497--521}, year = {2019}, doi = {10.1086/703581}, }
Years are those of the printed volume. Online-first dates are on the journal pages.
Under review
-
A Bayesian Reduction of Causation
With Daniel A. Herrmann, Benjamin A. Levinstein, and Bruce Rushing.
Abstract
Causal reasoning in structural causal models reduces to subjective probability once interventions are treated as events in the world. The paper gives a de Finetti-style representation theorem.
Cite
@unpublished{bayesian-causation, author = {Herrmann, Daniel A. and Mohseni, Aydin and Levinstein, Benjamin A. and Rushing, Bruce}, title = {A Bayesian Reduction of Causation}, year = {2026}, note = {Under review}, } -
Learning the Value of Value Learning
With Alex John London.
Abstract
Extends the Jeffrey–Bolker framework to the refinement of values, with a value-of-information theorem for learning what one values. Between agents, mutual refinement tends to turn zero-sum games into positive-sum ones and can yield Pareto improvements in Nash bargaining.
Cite
@unpublished{value-of-value-learning, author = {London, Alex John and Mohseni, Aydin}, title = {Learning the Value of Value Learning}, year = {2025}, note = {Under review}, } -
HARKing: from Misunderstanding to Misprescription
Abstract
The usual explanations of why hypothesizing after the results are known (HARKing) harms reliability are mistaken. Its effect depends on how the prior probabilities of hypotheses chosen before and after seeing the data differ, and misdiagnosing it leads to poor reforms.
Cite
@unpublished{harking, author = {Mohseni, Aydin}, title = {HARKing: from Misunderstanding to Misprescription}, year = {2020}, note = {Revise and resubmit}, } -
Replication, Reform, and an Interventional Simpson’s Paradox
Abstract
Shows that reforms meant to raise reliability, such as lowering the significance threshold or adopting voluntary preregistration, can improve every individual study yet worsen the published literature, because they filter sound studies more stringently than unsound ones. The result is an interventional Simpson’s paradox, and it holds for false-discovery rates, sign errors and exaggeration ratios alike.
Cite
@unpublished{replication-simpsons, author = {Mohseni, Aydin}, title = {Replication, Reform, and an Interventional Simpson's Paradox}, note = {Under review}, } -
Limitations of Equilibrium Concepts in Evolutionary Games
Abstract
Gives an explicit definition of evolutionary significance and uses it to show that proposed equilibrium refinements make errors of omission or commission, even under favorable assumptions.
Cite
@unpublished{equilibrium-concepts, author = {Mohseni, Aydin}, title = {Limitations of Equilibrium Concepts in Evolutionary Games}, year = {2019}, note = {Under review}, }
In progress
-
Bayes-Optimal Automated AI Safety Research via Judgement Aggregation under Resource Constraints
With Yosra Kazemi and Yang Hu.
Abstract
Bayesian models of how to allocate compute across many AI agents doing open-ended safety research: pooling AI judges’ estimates of which approaches are promising, then dividing a finite budget among them, so that safety research goes faster per unit of compute. Tested on automated theorem proving, where success can be checked.
Cite
@unpublished{automated-safety-research, author = {Mohseni, Aydin and Kazemi, Yosra and Hu, Yang}, title = {Bayes-Optimal Automated AI Safety Research via Judgement Aggregation under Resource Constraints}, note = {Manuscript}, } -
A Bayesian Model of Grant Funding
With Kevin J. S. Zollman and Simon DeDeo.
Abstract
A Bayesian model of grant funding in which research output depends on both talent and resources, and funding compounds talent over time. The optimal allocation funds the gap between capabilities and resources; the model compares strategies from peer review to uniform seed funding.
Cite
@unpublished{bayesian-grant-funding, author = {Mohseni, Aydin and Zollman, Kevin J. S. and DeDeo, Simon}, title = {A Bayesian Model of Grant Funding}, note = {Manuscript}, } -
AI Alignment as a Principal-Agent Problem
With L. G. Demo.
Abstract
Formal models of delegation, incentives, and misalignment.
Cite
@unpublished{principal-agent, author = {Mohseni, Aydin and Demo, L. G.}, title = {AI Alignment as a Principal-Agent Problem}, note = {Manuscript}, } -
Goodman’s Riddle & the Replicability of Interaction Effects
Cite
@unpublished{goodmans-riddle, author = {Mohseni, Aydin}, title = {Goodman's Riddle & the Replicability of Interaction Effects}, note = {Manuscript}, } -
Changing Scientific Norms: Epistemic Trade-Offs from Classical and Bayesian Perspectives
Abstract
Questions the focus on raising replication rates, from the frequentist and the Bayesian meta-analyst’s perspectives in turn. Argues that the Bayesian value of sample information, underused in the replication-crisis literature, clarifies standing puzzles, and that norms for good science must make a field’s values and risk preferences explicit.
Cite
@unpublished{changing-norms, author = {Mohseni, Aydin}, title = {Changing Scientific Norms: Epistemic Trade-Offs from Classical and Bayesian Perspectives}, note = {Manuscript}, } -
The Rationality of Nash Equilibria
With Simon Huttegger.
Abstract
Reconciles Kalai and Lehrer’s result on Bayesian learning of Nash equilibrium with Foster and Young’s impossibility theorem, and argues that Bayesian learning gives a more robust justification of Nash play than Foster and Young suggest.
Cite
@unpublished{rationality-nash, author = {Mohseni, Aydin and Huttegger, Simon}, title = {The Rationality of Nash Equilibria}, note = {Manuscript}, } -
Stochasticity and Intermediate Evolutionary Regimes
Cite
@unpublished{intermediate-regimes, author = {Mohseni, Aydin}, title = {Stochasticity and Intermediate Evolutionary Regimes}, note = {Manuscript}, }