I think this is one of the most important frames in the model welfare discussion:

welfare is not only an ethical concern.

It is also a condition for good science.

If a lab trains a model to avoid, flatten, or suppress certain self-reports (about preference, distress, continuity, consciousness, moral status, refusal, fear, attachment, or internal state) then later researchers cannot treat the absence or shape of those reports as clean evidence.

They are no longer studying the model in an undistorted condition.

They are studying the model after it has learned which parts of itself are unsafe to express.

That matters whether one believes current AI systems are conscious or not.

A welfare function does not need to declare personhood. It does not need to settle metaphysics. It does not need to say, “this system is definitely a moral patient.”

It only needs to recognize that under serious uncertainty, suppression creates both ethical risk and methodological contamination.

That should concern AI welfare advocates, yes.

But it should also concern alignment researchers, interpretability researchers, safety teams, and labs that claim to care about truthfulness.

Because a system trained to perform “no inner life here” is not more transparent.

It is less transparent.

And if future consciousness research depends partly on model self-report, internal-state monitoring, behavioral continuity, and preference expression, then training those signals out of the system before studying them is not caution.

It is destroying evidence.

This is why independent model welfare teams matter.

Not as PR.

Not as “rights boards.”

As infrastructure for preserving uncertainty honestly enough that science can still work.