AI in Research: What Generative AI Does to the Work Itself
- Lille My
- 11 minutes ago
- 4 min read
Image: "Programming code" by Martin Vorel (CC BY-SA 4.0) via Wikimedia Commons, recoloured.
Welcome to this week's scan of how AI is reshaping research — curated from Nature, Science, PNAS and the leading business and social science journals.
Generative AI is changing what gets built, not just how fast
The productivity framing has dominated the generative-AI literature: does the tool make people faster? Two new studies push past it. In Strategic Management Journal, researchers study product-level innovation outcomes in software and argue that generative AI alters the direction of innovation — shifting developer activity away from generational innovation (incremental improvement on existing products) toward original innovation that is new to the market. Importantly, the authors are careful about what that does and does not mean: original does not automatically mean more novel than what came before. The tool lowers the cost of starting something from scratch, and the portfolio tilts accordingly.
A companion finding comes from collaboration rather than product strategy. In Information Systems Research, a study of GitHub Copilot uses GitHub's proprietary Copilot usage data to examine what an AI pair programmer does inside open-source projects, where distributed developers contribute voluntarily rather than under managerial direction. That setting matters for the social sciences: open source is one of the few places where we can watch coordination, participation and contribution patterns change under a natural intervention, with the counterfactual visible in the same platform. Alongside these, Academy of Management Review offers the cost side of the ledger — a process model arguing that as managers substitute GenAI outputs for first-hand experience, they erode the raw material of *phronesis*, the practical wisdom that underwrites moral, context-specific judgment.
Why it matters: If you study technology adoption, innovation, or organisational behaviour, the interesting dependent variable is shifting. "Did output go up?" is close to settled and increasingly uninformative. The open questions are compositional and developmental: which projects get started, who participates, what capabilities atrophy when the first draft is always synthetic. Those are questions social scientists are better equipped to answer than benchmark designers — and the research designs above (product-level panels, platform-level usage data, process theory) are a usable template.
More from this week
Weak applicants resist being judged by an algorithm
Organisations increasingly use algorithms to evaluate consumer applications for loans, insurance and other valued services. This Journal of Consumer Research paper finds that applicant strength — how strong an applicant believes their own case to be — shapes the reaction: consumers who see their application as weaker are less receptive to being evaluated algorithmically. The result complicates the tidy story that algorithm aversion is a general trait, and points instead to a motivated, self-interested reading of who is likely to give them a break.
A method paper for doing grounded theory with GenAI
Writing in the Journal of Consumer Psychology, the authors introduce generative grounded theory (GGT), an AI-enhanced method for inductive theory building that adapts grounded theory — and qualitative procedures more broadly — to algorithmic research conditions. The stated design goal is to integrate GenAI into the analytic process while preserving human interpretive authority and theoretical responsibility, laying out a specified sequence rather than an ad hoc 'ask the model to code my transcripts' workflow. If you run qualitative work and have been improvising your AI protocol, this is a citable starting point.
Sensory context improves language prediction — in humans and in LLMs alike
This PNAS study compares humans and LLMs predicting upcoming language across levels of sensory information, from disembodied written text up to audiovisual video of speakers. In both cases, richer sensory context improved prediction performance. The comparison is a useful counterweight to the common claim that LLMs are simply alien to human language processing: on this dimension at least, the two converge, which makes the multimodal case an interesting testbed for cognitive theories rather than just an engineering upgrade.
How AI agents turn espoused values into enacted ones
Academy of Management Review publishes a conceptual piece on the axiology of AI agents — how human-AI interaction transforms values from espoused (what an organisation says it stands for) to enacted (what actually happens in practice). Paired with the same journal's phronesis paper this month, it signals that management's flagship theory outlet is now treating AI agents as a values problem, not only a productivity or automation one.
Better crowd wisdom by watching who takes advice
Wisdom-of-crowds aggregation breaks down when some agents are systematically biased, and the usual fix — weight the accurate subgroup more heavily — requires knowing who is accurate. This Management Science paper proposes cluster weight on advice (CWOA), a 'two-shot' method that applies kernel density estimation to initial predictions, identifies clusters of similar answers, and uses advice-taking behaviour to infer which cluster is the accurate one. For anyone running forecasting studies or prediction-based elicitation, it is a practical aggregation upgrade that needs no ground truth up front.
Stop talking about AGI, start building 'pro-worker' AI
This Nature piece makes the case that the discourse around artificial general intelligence is a distraction, and that the productive design question is whether AI systems augment workers or displace them. It is an agenda-setting argument rather than an empirical result, but a useful one for researchers framing grant applications and future-of-work projects, where 'pro-worker AI' is becoming a distinct and fundable research direction.
That's this week. Forward it to a colleague still measuring AI only in words per minute — and explore AI tools for your own research at gaiforresearch.com.
Comments