Sentient Futures
Back to mentors
Matilda Gibbons
Artificial sentience

Matilda Gibbons

Applying indicators of pain to LLMs

University of Pennsylvania

Bio

I am a postdoc in Neuroscience at the University of Pennsylvania and a Digital Sentience Consortium Research Fellow researching the neural basis of sentience and consciousness. My research models include insects, mice and, more recently, AI systems. I am mainly interested in how we can best test for sentience across different systems, without having a clear picture of the neural requirements for consciousness. I am particularly interested in using the feeling of pain as a key subjective experience to test for, for three main reasons: 1) pain is well understood in the neuroscience field across many model systems, 2) it is anatomically and behaviorally differentiable from it's unconscious counterpart (nociception), and 3) it is, arguably, the most important subjective experience in terms of welfare risk.

Mentee must-haves/nice-to-haves

The mentee should ideally be interested in testing for sentience in artifical systems or animals. Interests in neuroscience, philosophy and cognition/behavior would be preferable, and within the domain of using research to answer questions about sentience.

Mentee role

Mentees could take on one of the five main tasks listed above dependent on their interests, and I would work alongside the, or they could sub-tasks within these tasks depending on the time committment.

❓ Sample mentee tasks

  1. compiling a comprehensive list of putative sentience indicators
  2. providing a detailed explanation of why each indicator is relevant to sentience, using the human neuroscience literature
  3. collapsing these indicators into their underlying functional rationale. For example, “runs away from something that could cause pain” could be reframed as “motivated to avoid negative stimuli”
  4. assigning evidential weights to each indicator with certainty ratings
  5. using LLM architecture as a test-case to test whether the new indicators can be applied

Sub-task examples: Assigning evidential weights to five of the indicators Applying one or two indicators to LLM architecture

Mentor support

For research papers, I can help with evidence gathering, writing, editing, figure creation, data analysis or anything related to the submission process. For grant proposals or project facilitation, I can help with developing and fleshing out aims and timelines, and planning specifics.

Mentor-led project

Applying indicators of pain to LLMs

Currently, no gold-standard method exists for determining whether an agent is sentient, leaving such judgments to subjective and potentially biased interpretation, despite their significant consequences for welfare. In the animal sentience literature, multiple lists of indicators for sentience have been proposed, but such lists are currently not directly applicable to AI systems. This project will aim to convert these indicators into their more basic, functional underpinnings, so they can be used for AI systems. This will involve four main tasks:

  1. compiling a comprehensive list of putative sentience indicators
  2. providing a detailed explanation of why each indicator is relevant to sentience, using the human neuroscience literature
  3. collapsing these indicators into their underlying functional rationale. For example, “runs away from something that could cause pain” could be reframed as “motivated to avoid negative stimuli”
  4. assigning evidential weights to each indicator with certainty ratings
  5. using LLM architecture as a test-case to test whether the new indicators can be applied