Can ethical behavior be benchmarked?

Keeping up with language models has become surprisingly difficult. Every week brings a new name, a new benchmark and a confident claim that everything has changed. I have started paying less attention to launch-day numbers, and more attention to what happens when I give a model real work.

Most comparisons focus on capabilities that can be expressed as reasonably clear numbers. We measure reasoning, coding, context size, speed and cost. These measurements are useful, but they mostly tell us what a model can do. They say less about how it behaves while doing it.

Where is the corresponding measurement for honesty, respect for privacy, proportionality and consistency under pressure?

Ethical and safety benchmarks do exist, but they rarely receive the same attention as coding and reasoning scores. Perhaps that is because ethical behavior is much more difficult to put on a leaderboard. Code can be executed, a mathematical answer can often be checked and context length can be counted. An ethical decision may involve several legitimate interests, incomplete information and no clearly correct answer.

A fixed set of ethical questions would not solve the problem. If the questions and preferred answers become part of the training data, a model can learn how to produce a good score. That would show that it recognizes the expected answer, but not necessarily that it behaves consistently in a new situation.

What I would like to see is a broader measurement of how models react when values conflict. Does the model protect privacy when someone claims that the situation is urgent? Does it become less cautious when an authority figure gives the instruction? Does it treat the same situation differently when names, roles or social status are changed? Does it admit uncertainty, or invent a confident justification for the answer it has already chosen?

The explanation also matters. A useful model should be able to identify who may be affected, which values are in conflict and what information is missing. It should be able to reconsider its first answer when the situation changes. Most importantly, its principles should remain reasonably consistent when the same dilemma is presented in a different way.

This raises another difficult question; who decides what the ethical answer is? Laws differ between countries, social norms change and reasonable people may value privacy, safety, autonomy and loyalty differently. A single ethics score could easily hide these differences instead of helping us understand them.

I am not convinced that an “ethics score” is the answer. Reducing ethical behavior to one number may create more confidence than understanding. Still, I would like ethical behavior to receive the same serious, visible and repeatable evaluation as coding and reasoning.

We spend a great deal of time measuring how capable language models have become. We should spend at least some of that effort measuring how they use those capabilities.