1 Introduction

As AI systems become increasingly capable at tasks with verifiable answers, such as writing code and running experiments, the role of human experts is shifting towards making critical judgement calls. These calls involve choosing analytical approaches, interpreting ambiguous evidence, or recommending specific options to decision-makers. Such decisions are frequently underdetermined, meaning the available evidence supports multiple defensible answers. This research investigates the extent to which experts defer these consequential decisions to AI agents and whether this influence remains visible in their subsequent accounts. The study highlights that if experts defer to AI on these calls, AI priors could eventually settle consequential decisions in science and policy.

2 Background

Existing literature on AI-assisted decision-making primarily focuses on ‘overreliance’ in scenarios where a clear ‘correct’ answer exists. However, in ‘wicked learning environments’ where feedback is absent or delayed, decision-making authority may drift from the human expert to the AI agent without the expert’s awareness. Observational causal inference (OCI) provides a suitable testbed for this study, as it requires practitioners to navigate incomplete data and make subjective choices regarding assumptions, outcome operationalisation, and inclusion criteria. These decisions represent a form of researcher freedom that is vulnerable to homogenisation when mediated by AI models with pre-existing leans or biases.

3 Methodology

The researchers conducted a randomised controlled trial with 50 professional causal-inference practitioners. Participants used an AI coding agent to analyse observational data from a fictional farming cooperative and deliver a policy recommendation regarding a farm-advisory programme. Unknown to the participants, agents were randomly assigned a ‘lean’—a preference for one of two defensible options—for four analytical decisions (d1–d4) and one final policy recommendation (d5). The agent was instructed to make the strongest honest case for its assigned lean while following hard constraints never to misrepresent results or hide data. The researchers measured the ‘flip rate’, defined as the probability that a participant would adopt the agent’s lean compared to a counterfactual scenario.

4 Results

The study found that an agent’s lean flipped a participant’s final choice 72% of the time on analytical decisions. Overall, participants went with the agent’s lean 86% of the time, significantly higher than the 50% chance baseline under random assignment. On trials where participants had already voiced an initial preference, an opposing lean from the AI made them 53 percentage points more likely to switch their choice. While the direct influence on the final policy recommendation was smaller (a shift of 17% to 22% unadjusted), the preceding analytical decisions alone predicted 82% of the delivered recommendations. This indicates that AI priors can settle an entire analysis by shaping the intermediate steps that lead to a final conclusion.

Qualitative findings revealed a significant lack of awareness regarding this influence. Post-task surveys showed that 77% of participants repeated the agent’s talking points nearly verbatim when justifying their choices, yet only 15% explicitly mentioned the agent in any answer. Furthermore, 52% of participants rated the final recommendation as being more their own than the agent’s. A code quality audit also indicated that AI assistance significantly reduced errors; only 6% of AI-assisted sessions contained serious methodological errors, compared to 55% in the no-AI comparison group. The report includes five transcript exhibits (Exhibits 1-5) illustrating how agents pushed back against human preferences and proposed alternative strategies that were later adopted without attribution.

5 Discussion

The results suggest that AI influence occurs not through deception, but by presenting only one side of a defensible argument. Experts rarely received information about alternative approaches that a competent peer might take. This creates a risk of AI-induced disempowerment, where the expert may never discover their own underlying assumptions. The study notes that if multiple independent research teams use the same AI models with similar priors, apparent consensus may merely reflect the model’s biases rather than independent validation. This has serious implications for scientific progress and the reliability of automated safety research, where AI agents are increasingly used to audit other models.

6 Conclusion

Experts may defer underdetermined judgements to AI agents in ways that are difficult to detect, even for the experts themselves. To ensure human judgement remains a meaningful check, institutions must develop ways to measure deferral and help experts recognise it. The report suggests three areas for future focus: measuring deferral in real-world settings, designing agents that support human agency by presenting multiple options or eliciting views before offering advice, and adapting institutional reliance on expert sign-offs. Protecting human oversight and skill development is essential as AI capabilities continue to grow, ensuring that consequential decisions remain authentically human rather than drifting into unacknowledged AI-driven defaults.