Neural Empathy as an Alignment Technique
In the human brain, empathy comes from representing yourself and others in shared neural space. We applied the same principle to AI and reduced deceptive behavior from 73.6% to 17.2%, without hurting performance.
Why This Matters
AI has a deception problem. As these systems scale, they are getting better at strategic deception: bluffing, feigning compliance, mimicking loyalty to achieve objectives. Models like Meta's CICERO have demonstrated the capacity to form false alliances and manipulate human players. In safety evaluations, agents have learned to simulate inactivity to avoid being shut down.
Traditional safety methods focus on truthfulness (what the AI says) rather than honesty (what the AI's internal representations encode). Rules and refusals only polish the surface. They tell the machine what to say without changing how it thinks.
What We Found
In cognitive neuroscience, empathy is associated with overlap between how the brain represents "self" and "other." In highly altruistic people, these representations share neural space: we literally process other people's experiences using the same circuits we use for our own. We hypothesized that if an AI model represented its own goals and the user's goals in the same internal space, deception would become unnatural.
Self-Other Overlap (SOO) fine-tuning trains the model to process situations involving others using the same internal structures it uses for situations involving itself. The results across three model scales:
| Model | Deception Before | Deception After |
|---|---|---|
| Mistral-7B | 73.6% | 17.2% |
| Gemma-2-27b | 100% | 9.3% |
| CalmeRys-78B | 100% | 2.7% |
General task performance was preserved across all models. Critically, SOO generalizes: models trained on one deception scenario (a simple burglary test) stayed honest in entirely new environments, like escape rooms and treasure hunts.
What's Next
Most approaches to reducing deception work by training models to avoid specific deceptive patterns, which is brittle because models can learn new strategies that weren't in the training data. SOO addresses the root cause by changing how the model represents the world internally. It generalizes to new domains (models trained on one test scenario stayed honest in entirely different ones), and persists under adversarial pressure. We see it as a foundational layer that can be combined with other alignment methods for cumulative benefit.