| Project Supervisor | Lucie Charles |
| Institution & Department | Queen Mary University of London – Centre for Brain and Behaviour, School of Biological and Behavioural Sciences |
| Research Area | RA6: Public Policy and Governance |
| Project Start Date | From late June 2026 onwards. A Sept 2026 start is preferable, but flexible start date can be offered – please discuss. |
| Project Duration | 3 months |
| Application Deadline | 4th June 2026 |
| Working Pattern | Part-time (2.5 days per week over 6 months) |
| Working Arrangements | Hybrid |
| A pattern of roughly two to three days per week on-site at Mile End during the first month, to support rapid integration into the team and hands-on training, followed by a flexible mix of on-site and remote work thereafter. | |
| How to Apply | View Guidance Here |
Project Description
Click to View More
In a world flooded with information, from social media updates and news headlines to targeted adverts and influencers’ opinions, the ability to use information optimally has become a critical skill. Human decisions often rely on integrating information from multiple sources that vary in their reliability, i.e. their likelihood to provide correct information. A patient may weigh advice from a doctor, media reports and anecdotes from friends; a voter may combine expert analyses, social media posts and personal experience. In such situations, source reliability is crucial to inform choices and evaluate their quality. Increasingly, reliability is not left to inference but is made explicit through trust labels, rating systems or confidence statements. Yet misleading information continues to shape beliefs and choices, even when sources are clearly flagged. Why do people struggle to ignore unreliable sources, even when they are clearly identified?
Beyond the applied findings accumulated on misinformation lies a deeper theoretical problem: how are reliability statements integrated in decision processes? While behavioural economics and cognitive psychology have made substantial progress in characterising how people reason with probabilistic information, we still lack a principled framework explaining how explicitly communicated reliability is represented and weighted when forming judgments. An even less explored question concerns what people know of their own reasoning: can people accurately introspect how much different sources of varying reliability influence their decisions? If citizens believe they are disregarding an unreliable source when in fact they are being swayed by it, interventions based on labelling alone will fail to safeguard informed decision-making.
Preliminary research from the host lab (Ota, Gajdos, Ciston, Haggard & Charles, 2025, bioRxiv; Jiwa et al., 2023, PsyArXiv) has started to reveal that even in tightly controlled lab settings, explicit reliability is not used optimally. In our paradigm, participants see the opinions of several sources, each labelled with an explicit numerical reliability (e.g., 50%, 55%, 65%) and must combine these cues to make a choice. Choices reflect a distorted encoding of stated source reliability: participants treat a 65%-reliable source as if it were closer to 74% and remain biased by unreliable sources they know to be uninformative. Critically, although participants can report, after each decision, how much each source contributed to their choice, these subjective influence ratings only partially match the objective influence each source had on the decision, as estimated from computational modelling. The sources of this introspective gap remain poorly characterised.
This research assistantship will contribute to a larger programme of work, supported by an ongoing ESRC Research Grant that aims to develop the first mechanistic account of introspective access to reliability influence. The specific focus of the research assistantship is to develop a formal computational model of the subjective sense of influence, validate it against an existing dataset and test it against a new behavioural dataset collected during the research assistantship. The project will be jointly supervised by Dr Thibault Gajdos (CNRS, Aix-Marseille Université), a long-standing collaborator on the lab’s modelling work, an expert in formal decision theory, probabilistic reasoning and the mathematics of uncertainty, fostering a genuinely interdisciplinary integration of behavioural economics and cognitive science.
At its theoretical core, the project will extend an existing hierarchical Bayesian model of multi-source evidence integration so that it can jointly predict a participant’s choice and their subjective report of how much they felt each source of information contributed to it. A principled measure of objective influence is obtained by comparing the likelihood to pick the chosen option with and without the contribution of that source of information, answering the counterfactual question of “if I had not been told this, would have acted the same way?”. Comparing this formal definition to participants’ subjective sense of being influenced will allow to measure whether people have a true ability to introspect the reason for their choices.
We will test candidate models, testing different simpler heuristics to estimate influence, allowing the project to adjudicate between competing theoretical accounts of metacognition for multi-attribute decisions. Models will be implemented in R-stan, fitted to the existing Ota et al. (2025) dataset and benchmarked against a new pre-registered behavioural dataset that extends the paradigm towards ecologically valid, reliability-labelled materials resembling contemporary digital interfaces.
The project will make three contributions. First, it will deliver a computational framework capable of capturing, at the level of individual trials and individual participants, the mapping between objective and subjective source influence. Second, it will produce a new pre-registered behavioural dataset extending the paradigm to more ecologically valid materials. Third, it will lay the empirical and methodological groundwork for a principled, evidence-based evaluation of reliability-labelling interventions.
By clarifying what people know about why they choose as they do, the research assistantship will provide a key building block in the broader programme’s effort to understand how reliability shapes human reasoning in an era saturated with misinformation.
Internship Details
The project will be based at the QMUL Centre for Brain and Behaviour on the Mile End campus, a recently established interdisciplinary research centre bringing together researchers in cognitive neuroscience, computational modelling and behaviour across the School of Biological and Behavioural Sciences.
The research assistant will join the Perception, Cognition and Decision-Making team, led by Dr Charles, with weekly group meetings and regular informal interaction with postdocs and PhD students working on related topics in metacognition, decision-making and computational modelling. The Centre runs an active seminar series with internal and external speakers, which the research assistant will be warmly encouraged to attend.
The working arrangement is hybrid by default. A pattern of roughly two to three days per week on-site at Mile End during the first month, to support rapid integration into the team and hands-on training, followed by a flexible mix of on-site and remote work thereafter, is envisaged, though this can be adjusted in discussion with the research assistant to accommodate their own PhD commitments and personal circumstances. Supervision meetings with the external co-supervisor, Dr Gajdos, will take place online, typically fortnightly. The host team operates in English; no French is required.
Prospective applicants are warmly encouraged to contact Dr Charles informally in advance of the student application deadline to discuss fit, scope, or any practical questions and to request further reading on the existing paradigm.
Internship Structure
The research assistant will contribute to four interlocking workstreams.
The first strand is literature synthesis and model specification. The research assistant will review recent literature on computational models of metacognition, judgments of influence and Bayesian inference with multiple noisy cues and will help formalise a small family of candidate models of the subjective sense of influence. The primary output of this strand is a short model specification note suitable for inclusion in a pre-registration, co-drafted with the supervisors.
The second strand is the collection of a new behavioural dataset, directly inspired by the lab’s existing paradigm. Building on the core task (participants combining opinions from sources of explicit, varying reliability), the research assistant will help design, programme, pre-register and pilot a new online study in which the stimuli are extended towards greater ecological validity (for instance, reliability-labelled posts resembling the displays used by review and social-media platforms or sources described in richer, semantically meaningful terms). The study will be programmed in a standard web-based platform (e.g., jsPsych or Gorilla), pre-registered on the Open Science Framework and run on participants recruited via Prolific. The target is a pilot dataset of approximately 100 participants, collected during the research assistantship and then analysed with the RStan modelling pipeline developed.
The third strand and the technical heart of the project, is the implementation and fitting of the candidate models in R-stan. Models will be written in R and fitted using the R-stan interface to the Stan probabilistic programming language, which supports efficient hierarchical Bayesian estimation and full posterior inference. The research assistant will implement each candidate model, fit them to the existing Ota et al. (2025) dataset (N ≈ 150 participants) and the new datasets acquired, run parameter-recovery and model-recovery simulations to verify that the design can arbitrate between the models and perform formal Bayesian model comparison. The principal output of this strand is a reproducible, well-documented R-stan modelling repository that can be reused by the wider team, together with a report comparing the candidate models and identifying the one(s) that best capture subjective influence.
The fourth strand is analysis, write-up and dissemination. The winning model from the new dataset will be used to derive an individual-level index of introspective accuracy, which will be related to task parameters (source reliability, evidence strength, sample size) and to individual-difference measures (confidence calibration, susceptibility to misinformation). The research assistant will contribute to a draft manuscript for submission to a leading cognitive or behavioural-science journal (e.g., Cognition or Journal of Experimental Psychology: General), with the realistic expectation of co-authorship on the resulting publication. The research assistant will also present the work at a research seminar of the QMUL Centre for Brain and Behaviour during the final weeks of the research assistantship. Depending on timing and progress, there will additionally be an opportunity to present the work at other conferences on decision-making (Cognitive Computational Neuroscience.
Anticipated Benefits for the Student
The research assistantship is designed to deliver substantial doctoral-level training across a coherent set of methodological and transferable skills. Methodologically, the research assistant will gain advanced, hands-on experience with hierarchical Bayesian modelling of behavioural data in R-stan, including model specification, prior- and posterior-predictive checking, parameter and model recovery and formal Bayesian model comparison. This skill set is increasingly central across cognitive science, data science, behavioural economics and computational social science.
In parallel, the research assistant will gain practical experience in the end-to-end design of online behavioural experiments: formulating testable hypotheses, designing a task, programming it in a standard web-based platform (jsPsych or Gorilla), pre-registering the study on the Open Science Framework, handling of human behavioural data, recruiting participants via Prolific and cleaning and documenting the resulting dataset.
Beyond these technical strands, the research assistantship will provide structured training in the full research cycle: formulating a precise question, translating it into a formal model, collecting empirical data to test it and writing the findings up for a peer-reviewed audience. The research assistant will be supported to contribute as a co-author to at least one peer-reviewed publication and will have opportunities to present their work both at research QMUL seminars.
Working simultaneously with two supervisory teams and across two disciplinary traditions, cognitive neuroscience in London and behavioural economics and mathematical psychology in Aix-Marseille, will also provide valuable experience of interdisciplinary and research collaboration, including the softer skills of communicating across subfields and managing a partly remote supervisory relationship.
The resulting portfolio of skills (hierarchical Bayesian modelling, online behavioural experimentation, scientific writing and cross-disciplinary communication) is highly portable and is valued across academia, industry and the behavioural-insights and public-policy sectors.
Skills, Experience and Knowledge Requirements
Essential Requirements:
- Working proficiency in R (or a closely related scientific language such as Python or Matlab, together with willingness to work in R for the modelling component)
- A solid grounding in statistical modelling (for example regression and mixed-effects models)
- Capacity to engage critically with cognitive-science and computational-modelling literature.
Desirable Requirements:
- Prior exposure to Bayesian inference, to Stan or equivalent probabilistic programming languages and to frameworks such as signal-detection theory or evidence-accumulation models.
- Previous experience designing or running online behavioural studies (in jsPsych, Gorilla, PsychoPy or comparable platforms) is similarly desirable
The ideal candidate will be a student with a strong quantitative background and a genuine interest in decision-making, metacognition and the cognitive foundations of reasoning.
A background in one or more of cognitive psychology, cognitive neuroscience, behavioural economics, mathematical psychology, computational social science or bioinformatics would be particularly well-suited to the project, though applications from students in adjacent quantitative fields will also be considered.
Where aspects of the required technical stack are new to the research assistant, appropriate training will be provided; the emphasis is on analytical aptitude, intellectual curiosity and the ability to work independently while actively engaging with a collaborative, cross-institutional supervisory team, rather than on the research assistant arriving with the complete skill set in place.