When Do We Trust the Machine? Algorithm Aversion, and Governing AI in Scholarship
- Lille My
- 5 days ago
- 3 min read
Welcome to this week's scan of how AI is reshaping research — curated from Nature, Science and the leading business and social-science journals.
Algorithm aversion isn't all-or-nothing — it's asymmetric
For a decade, 'algorithm aversion' has been one of the most-cited reasons that good models go unused: people, the story goes, simply distrust advice once they learn a machine produced it. A new paper in Information Systems Research by Saunak Basu, Alan R. Dennis, Aravinda Garimella and Wencui Han complicates that tidy picture. Across four studies — two randomized controlled experiments, a field study and a qualitative interview study — set in a Fintech investment context, they find that aversion to AI advice is asymmetric: it depends on what the advice actually recommends, not just on the fact that an algorithm gave it.
The mechanism turns on how investments are judged. Some signals are explicit and codifiable; others rest on tacit, experience-based judgement that human experts believe machines cannot fully access. When AI advice lines up with an expert's read, it is absorbed easily. When it cuts against that tacit judgement — telling the expert their instinct is wrong — acceptance drops. In other words, the resistance isn't a blanket suspicion of algorithms; it concentrates precisely in the cases where the AI is most likely to be adding new, contrarian information. That is a far more actionable finding for anyone designing decision-support systems than 'people don't trust algorithms.'
Why it matters: For researchers studying human–AI collaboration — and for the marketing, finance and operations scholars whose fields increasingly deploy algorithmic advice — the lesson is that adoption hinges on the structure of the task, not on a fixed personality trait called 'aversion.' The most valuable AI recommendations, the ones that contradict expert intuition, are also the ones humans are most primed to reject. Designing for that asymmetry — surfacing the tacit cues a model used, calibrating when it should push back — may matter more than making the model marginally more accurate.
More from this week
A governance blueprint for AI in peer review
Responding to a proposal to 'fight fire with fire' by infusing AI into peer review, Zhenhui Jiang argues in a Management Science commentary that AI is already in the review pipeline informally — and that the real task is moving from unmanaged 'shadow AI' to governed, transparent use. The piece frames what accountable AI-assisted review could look like, a question of direct concern to any scholar who submits to or reviews for major journals.
Robust inference when your regressors come from ML
Empirical researchers increasingly train a model to predict some variable, then plug those predictions into a regression as if they were observed — a shortcut that can quietly bias inference. Gordon Burtch, Edward McFowland, Mochen Yang and Gediminas Adomavičius propose EnsembleIV, a method that constructs instrumental variables from ensemble learners to recover valid statistical tests with ML-generated regressors. It's a practical tool for the growing number of studies that mine text, images or other unstructured data for their measures.
Modeling the academic job market in the age of AI
As AI changes how research aptitude is demonstrated and screened, the pathways into academia are shifting too. This Research Policy paper develops a game-theoretic model of the pre-doctoral academic labor market under the influence of AI, examining how hope and costly signaling interact when, as the title puts it, 'silicon' enters the picture. A conceptual lens on a labor market many early-career researchers are navigating right now.
Rethinking how we govern autonomous AI
When machines act without human input, they expose people to new risks. The dominant business-ethics response — the 'moral machine thesis' that autonomous systems should emulate morally right human decision-making, and that regulation should follow suit — is questioned in this Journal of Business Ethics article by Elsa Kugelberg and Henrik D. Kugelberg. They argue for a more practice-dependent account of AI governance, with implications for how firms and regulators frame accountability.
Scroll tracking as a window into choice
Ilkka Leppänen introduces 'scroll tracking' in Management Science, a process-tracing method that uses touch-scroll behavior on responsive mobile web pages to record a response curve in pixel–millisecond space as a decision unfolds. Pitched as a naturalistic alternative to lab eye-tracking, it offers consumer and decision researchers a low-friction way to study the moment-to-moment cognition behind preferential choice on the devices people actually use.
That's this week. Forward it to a colleague weighing whether to trust their model — and explore AI tools for your own research at gaiforresearch.com.
Comments