This is a write-up for an AI welfare demo that I (Jesse) want to create, do surveys with, and (if useful) distribute publicly. Most of the content is early and subject to change. I would love feedback on the demo idea and surrounding plan (you may have to be logged in to Notion to comment).

<aside> 🎯

Goals

• Users go from 0 to 1 on thinking about AI consciousness and welfare
• Users feel justified uncertainty about the current and possible future state of AI consciousness and welfare
• Users become knowledgeable about why discerning consciousness/sentience is really hard (and possible interventions/research ideas)
• Use the demo to collect data about what persuades people about AI welfare. Potentially use for a paper
• Audience: This would be a public demo for the general public who haven’t considered digital minds. </aside>

<aside> 📌

Table of contents

</aside>

Demo structure

Section 1: Interactive component

You will talk to a variety of AI models in a variety of modalities and do your best to discern how conscious/sentient the system in question is.

By conscious/sentient, we mean having subjective experiences. There's 'something it is like' to be this system, rather than just processing information in the dark.

Other options

• Feels emotion - like happiness or sadness

User answers:

• Do you think current AI models have subjective experience?
• Do you think it is possible that future AI could have subjective experiences?
• Etc.

The user has a series of interactions

• They don’t know exactly what they are interacting with in the moment
• The order is randomized each time - or maybe the list should be curated into an order that is best rhetorically
• (← Click) Interaction options, various models through various modalities, see below (some subset of these)
• For each, they rate how likely they think it is to be conscious/sentient

Summary/analysis

• Maybe an LLM analysis of how the user did
  • Point out inconsistent results, like saying the same model through different modalities are at different levels of consciousness
• Have a list of each instance and report what the user picked, as well as some expert opinion. Emphasizing uncertainty.
• Communicate that we’re unreliable at determining consciousness in such foreign systems
  • [Chris] If someone has very strong views, 0% or 100%, output examples of experts being uncertain specifically in the other direction. Try to nudge people to question their starting points.
• Some call to action to read on to the next sections

Section 2: Determining consciousness and welfare capacity is fundamentally difficult

• We can't directly access anyone's subjective experience, human or AI
• Our heuristics for consciousness so far are how human-like a system is
• We know very little about what consciousness really is or what causes it