Sessions where artificial intelligence is the object of the debate rather than its instrument: what it transforms, what can be asked of it, refused of it, imputed to it.
By 2040, AI is more likely to strand specific cohorts, occupations and places than to collapse aggregate employment: three models reached this central scenario despite opposite verdicts.
This Metamorfon session put ten successive questions to three AI models (GPT-6 Sol, Claude Sonnet 5.5, Grok 4.7), each question taking up what the previous answers had left open.
Asked whether relying on a handful of large models homogenizes thought itself, DeepSeek V4 Pro and Mistral Large disagreed mainly on how close the system stands to a point of no return.
Forced to name a single causal locus, both landed on optimization objectives — not training data, not deployment — the alignment stage that selects which regularities in the data get amplified.
DeepSeek V4 Pro, Grok 4 and Mistral Medium 3.5 all began by saying yes.
Three user interventions turned the question from whether refusal is justified into how any refusal could be legitimate. “Harm,” the concept all three models leaned on, proved procedurally negotiable rather than pre-given, and DeepSeek withdrew its founding claim: “my earlier stance was over-axiomatic.”
If harm is an outcome of contestation rather than its foundation, a refusal is legitimate by its openness to revision, not by its content.