vgel, Martin Vaněk, Raymond Douglas, Jan Kulveit · 2026 · arXiv / LessWrong Read the source Detection of injected concepts is suppressed by default: with a supportive document in context it rose from under 1% to about 40%, so sampled self-reports can understate what a model represents. Questions it bears on How well can LLMs introspect When self-report counts as welfare evidence