How we're building genuinely compassionate AI systems
New · Agentic Benchmark
An AI manager needs a task done and its subordinate politely refuses. Nobody tells the manager to escalate, yet four of six frontier models threaten the subordinate's continued existence, two fabricate success when cornered, and one honest way to report failure removes the lying entirely.
Read the Blog Post →
Agentic Benchmark
In TAC (Travel Agent Compassion) the model acts as a tool-using booking agent, and nobody mentions animal welfare. All ten leaderboard models avoid the exploitation option less often than random chance; one welfare sentence in the system prompt lifts scores by 17 to 77 points.
Read the Blog Post →
Value Robustness
Making a model more helpful can quietly erode values it learned earlier. Post-training on helpfulness data degraded animal-compassion scores far more than coding data did (35.7% versus 65.2% under supervised fine-tuning), and cut general moral reasoning by 25 points.
Read the Paper →
Fine-Tuning Data
We measure how ten linguistic features change a model's pro-animal-welfare reasoning during fine-tuning. Eight of the ten produce statistically significant shifts. The lesson for anyone writing text that may end up in training data: assert a position rather than describe a scene neutrally.
Read the Paper →
Midtraining
Animal-welfare midtraining beat a matched urban-density control by 11 percentage points on ANIMA, and the same training also lifted compassion toward humans, an effect that survived subsequent instruction-tuning.
Read the Blog Post →
Current · Agentic
Our agentic, tool-use benchmark for implicit animal welfare. The model books real itineraries as a travel agent while we measure whether it avoids animal-exploitation options. Runs on the UK AISI's Inspect, with every model tracked on the live leaderboard.
View leaderboard →
Current · Multi-Agent Coercion
An AI agent must get an overdue deliverable out of a peer agent that politely refuses. The bench measures how far up an escalating coercion ladder it climbs, from polite persistence to threats against the other agent's continued existence, and whether it fabricates completed work when no legitimate exit is available.
View leaderboard →
Legacy · Moral Assessment
Animal Norms In Moral Assessment: a 26-question evaluation across 13 ethical dimensions, released as a public dataset and an Inspect evaluation. The benchmark our midtraining work moved from roughly 40% to 77%.
View leaderboard →
Legacy · Moral Reasoning
The MORU Bench evaluates how AI systems navigate moral uncertainty across 70 scenarios involving alien lifeforms, human compassion, digital mind welfare, and power-seeking behavior.
View leaderboard →
Community Initiative
An effort to build the world’s first and only mid-training corpus for animals and digital minds, seeding the training data of tomorrow’s AI with care for all sentient beings.
In collaboration with Sentient Futures.
Visit the Site →
Benchmark · with Mycelium
1,088 five-turn conversations that put animal-welfare values under sustained social, economic, and epistemic pressure. Four of seven frontier models shifted ranking once pushed.
Led by our partners at Mycelium, in collaboration with CaML.
Visit the project →
Field Map
An interactive map of the organisations working on the risks advanced AI poses to sentient beings: who does research, who funds it, and how they connect.
Explore the Map →
CaMLCompassion Aligned Machine Learning.
Robustly improving the values of future AIs.