AI Safety

Rethink Priorities' Moral Weights project faces critique over animal-friendly bias

New analysis reveals flaws in the model that equates 14 bees to 1 human.

Deep Dive

Rethink Priorities' Moral Weights project attempts to quantify the intensity of suffering across species, assigning a 'moral weight' relative to humans (1.0). Their model combines empirical proxies (60%), neurophysiological data (30%), and an equality assumption (10%). The results are famously animal-friendly—e.g., 14 bees equal 1 human. Critic Bill Jackson presents two critiques of the empirical proxies, which drive these high weights.

First, 'functional analogues' lead to double counting: a pig scoring 'likely yes' on five related negative-valence behaviors (anxiety, fear, depression, panic, flexible self-protection) inflates its score because these are not independent but rather multiple ways of asking if the animal exhibits distress. More broadly, all proxies depend on a single contentious claim—that behavioral/cognitive functions predict subjective experience even with vastly fewer neurons. Second, a Bayesian critique: black soldier flies have only ~100,000 neurons yet score positively on 12 of 46 proxies (including depression-like behavior and hyperalgesia). Such high scores from tiny-brained animals suggest the proxies are poor discriminators of sentience, not that flies are highly sentient. These critiques undermine the project's foundational methodology.

Key Points
  • Double counting of correlated behavioral proxies (e.g., five negative-valence indicators in pigs) inflates moral weight scores.
  • All proxies rely on the untested assumption that behavioral functions predict subjective experience regardless of neuron count.
  • Black soldier flies (100K neurons) scoring on 12/46 proxies suggest the model's proxies are weak and overly inclusive.

Why It Matters

Challenges the foundation of animal welfare estimates used in effective altruism and AI alignment research.

📬 Get the top 10 AI stories daily