The AI Cognition Initiative (https://ai-cognition.org/) has worked on a Bayesian Hierarchical model which aggregates expert evidence to estimate the probability of consciousness in AI systems. This project aims to test alternative model configurations, run extended sensitivity analyses, and experiment with new Bayesian approaches to improve the model’s robustness, interpretability, and computational performance.
About the project
This project investigates how different Bayesian modelling choices affect estimates of AI consciousness in the Digital Consciousness Model (DCM) developed by the AI Cognition Initiative (https://ai-cognition.org/ ). The central research questions are:
How sensitive are consciousness estimates to particular modelling assumptions (e.g., priors, conditional dependencies, stance weights, indicator collapse methods)?
Can alternative hierarchical structures, likelihood parameterisations, or variable types (e.g. continuous features) yield more stable or interpretable results?
What computational methods and implementations improve model efficiency and scalability as we add more stances, features, and indicators?
Mentees will help design, implement, and evaluate alternative model variants. Concrete starting points include:
- Alternative structures where evidence is injected with more observations at a time, possibly by explicitly modelling expert judgements as a layer
- Reparameterising indicator–feature linkages using different prior families and concentration parameters.
- Testing continuous or ordinal versions of selected indicators or features. (Less of a current priority.)
- Running expanded prior and likelihood sensitivity analyses, including robustness checks for stance correlations and indicator dependence. (We have done some of this.)
- Exploring variational inference, approximate methods, or mentee's original ideas to reduce compute burden.
- Comparing results across multiple systems (LLMs, biological organisms, alternative digital architectures) and identifying which components most influence posterior rankings.
- The aim is to produce a clearer mapping between modelling assumptions and substantive conclusions about AI consciousness, while generating improved tools for future assessments.
There will be a link to relevant model details at https://ai-cognition.org/ as of late January 2026.
Theory of change
Understanding which AI systems plausibly warrant moral consideration is a prerequisite for safe and ethical deployment of advanced models. The Digital Consciousness Model provides a structured, evidence-based way to estimate the probability of consciousness across systems, hopefully enabling better decision-making around governance, precaution, and system design. By improving the model’s statistical foundations, reducing fragility to assumptions, and enhancing interpretability, this project strengthens our ability to track when emerging AI systems may cross morally significant thresholds — supporting safer navigation of transformative AI development.
Your role
Mentees will design and implement alternative versions of the DCM, propose new Bayesian formulations, run sensitivity analyses, and compare model outputs across configurations. They will have substantial autonomy in selecting modelling directions but will receive guidance on priorities and methodological standards. The role includes hands-on coding (primarily Python/PyMC), statistical reasoning, experimental design, and critical interpretation of model behaviour.
Prerequisites
• Proficiency in Python. • Familiarity and background in Bayesian modelling (e.g., priors, likelihoods, hierarchical models). (Some programming proficiency can be traded-off for statistical/Bayesian proficiency and viceversa) • Ability to read and reason about technical research documents. • Ideally: Experience with at least one probabilistic programming framework (PyMC, Stan, NumPyro, or similar). • Ideally: an understanding of neural networks or AI system behaviour.
Location preference
No geographical preference. Mentees should be available for at least one short synchronous meeting every fortnight (ideally every week) within UK–Europe–US overlap hours.
Application question(s)
Context: the digital consciousness model (DCM) can be found at https://ai-cognition.org/ as of late January 2026.
Choose one modelling assumption in the DCM (e.g. binary treatment of features, indicator independence, stance–feature conditional independence). Explain why relaxing or altering this assumption might materially affect the model’s outputs, and propose a concrete method for testing the impact of this change. Answer in one paragraph.
-
(Optional) Describe one Bayesian technique or inference method you have used (e.g. hierarchical priors, variational inference, MCMC tuning), or a relevant statistical model you have built or worked on. Summarise it in one paragraph. Then explain, in one to four paragraphs, the strengths and weaknesses of this method when applied to complex hierarchical models such as the DCM.
Provide a brief “red-team” analysis of the statistical limitations of the DCM. Identify key vulnerabilities or structural weaknesses in the current design, and explain why these limitations could meaningfully distort inferences about AI consciousness. Answer in one to two paragraphs.
If you were to change the structure of the DCM, how would you redesign it to improve scalability? In particular, propose approaches that (a) capture expert uncertainty without requiring repeated model runs, and/or (b) make individual runs substantially more efficient (e.g. under one minute). You may wish to consider alternatives to the current procedure in which expert-provided probabilities are collapsed into single binary observations before inference, and each model run generally is only fed one observation per indicator. Answer in one to five paragraphs.
About the mentor
Arvo is a Senior Researcher at Rethink Priorities' Worldview Investigations Team, and also did work for Oxford's Global Priorities Institute. Before that, he was a Research Analyst at the Forethought Foundation for Global Priorities Research and earned an MPhil in Economics from Oxford. His background is in mathematics and philosophy.