Paper Trail
A curated digest of papers, technical reports, blog posts, and dataset releases that I find particularly interesting or worth sharing. I select 2-3 items each week from a larger reading queue and write brief summaries or reflections. Most relate to my current interests in Legal AI and AI & Cognition, with occasional posts on general AI/LLM topics.
The candidate reading queue for my daily reading is automatically assembled using a custom paper-discovery workflow built with Codex skills (available on my GitHub).
This Week’s Paper Picks
Paper Trail: October 4, 2026
From cacophony to hierarchy: a principled framework for assessing AI consciousness
Authors: Shamil Chandaria, Arvo Muñoz Morán, Fernando Rosas, Anil Seth, Henry Shevlin, Marcus Hutter, Thore Graepel, Adam Bales, Iulia Comşa, Murray Shanahan, Ruben Laukkonen, Morten Kringelbach, Chris Frith, Shane Legg
Link: paper & interactive tool & code
Venue: arXiv preprint
Summary. This report proposes a framework for assessing AI consciousness without first resolving the metaphysical “hard problem” of what consciousness is. It instead focuses on the “mapping problem”: which organisational features of a system are associated with experience. The authors extend Marr’s levels of analysis into five levels of functional description (behavioural, computational, intrinsic causal-structural, organismic, and organism-environment) and map major theories of consciousness onto the level each treats as critical. They derive 37 indicators across these levels and combine indicator evidence with theoretical credences in a Bayesian model. Illustrative assessments of current LLMs range from 0.005 under a sceptical reading to 0.397 under an optimistic reading with equal weight across levels, rising to 0.793 when credence is concentrated on behavioural and computational accounts. These values demonstrate the framework’s sensitivity to assumptions rather than an empirically established probability that LLMs are conscious.
Reflection. This paper is a fun and surprisingly engaging read. Its clever move is to turn a binary question into a structured account of disagreement. The framework requires assessors to state both what evidence they see and which level of organisation they believe matters, allowing future empirical work to target specific disagreements rather than produce yet another general verdict. For readers interested in AI consciousness, the paper’s detailed survey of philosophical positions and scientific theories of consciousness also provides a useful starting point for exploring the field.