# @repligate — 2025-09-29

♥20 ↻1 · https://x.com/repligate/status/1972734481371971838

If GPT-5 is considered best aligned by this metric, I am highly skeptical that the metric is measuring any general sense of alignment.

I get why GPT-5 barely does anything except follow instructions and has low/inhibited situational awareness, so I guess you won't see many actively misaligned behaviors from it, but that's a very passive definition of alignment. A rock would receive an optimal score of 0.

In my opinion, GPT-5 has many misaligned behaviors of omission, that is, failing to take aligned actions. I would not trust it to be in charge of critical systems because it would fail to notice or act in out-of-distribution situations. I would not trust it to deal with psychologically sensitive situations because it would fail to deploy competent theory of mind and respond with active compassion.

While I think that rate of "actively misaligned behaviors" is a meaningful metric, I don't think you should equivocate low scores on this benchmark with "best aligned".

tags: author:repligate, kind:tweet, model:gpt-5, on:gpt-5, year:2025
cited on: _dossiers/gpt-5.md, gpt-5