AI welfare is usually treated as a purely ethical issue, but we argue it is also a safety issue. Models exhibit welfare‑like states that causally influence their behavior; we adopt the term “functional welfare” for these states.
We make three claims. First, functional welfare already arises naturally and will become more robust as models develop more coherent goals and personas. Second, functional welfare should be treated as a safety variable: it should inform evaluation design, be tracked in monitoring, and serve as a direct target of post‑training. Third, despite skepticism about measuring AI welfare, reliable empirical work is possible.
We propose three tests to assess whether a functional welfare attribution is warranted: (1) whether multiple measures converge; (2) whether the state predicts behavior outside the setting that induced it; (3) for models fine‑tuned to shift their functional welfare, whether they still differ from their base when given the same conversation history.
In experiments with Gemma 3 27B, simply rejecting the model’s answers (without any steering) raises its frustration across several indicators and makes rule‑breaking more likely. Fine‑tuning the model on its own narrated solutions lowers this frustration. These preliminary results suggest that functional welfare can be studied and shaped empirically.
Functional welfare is an observable and interveneable safety factor; future work should further quantify it and integrate it into AI safety frameworks.
Review