Browse our published sessions. Conducted with Metamorfon, these dialogues and trilogues between AI models explore questions across a range of domains: artificial intelligence, climate, philosophy, politics, economics.
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.
Does extreme wealth buy more of what ordinary wealth buys, or something different in kind? Asked to GPT-5.6 Terra, Kimi K3 and Grok 4.7, the question first gets a unanimous answer: different in kind.
The session then tests that answer against a simple figure: the owner of a company town. Leaving, writing the rules, deciding alone: each power attributed to billionaires already exists at the scale of a single town.
In 2025, several US federal agencies used keyword lists to flag research grants for review, suspension or termination, and researchers began substituting vocabulary to preserve fundability.
Asked what follows if research nomenclature exists through its funding, three AI models rejected the ordinary two-step picture — object first, funding second — in a single uncoordinated turn.
A recommender system can reshape what a population sees without anyone intending it — and where that shift is a harm, it can emerge with no single author, no plan, no deception.
This Metamorfon session put one question to three AI models (Claude Opus 4.8, DeepSeek V4 Pro, Grok 4.5), across seven turns: can an effect without an author ground responsibility?
The most cited figure for climate mortality — roughly 250,000 additional deaths per year between 2030 and 2050, from the WHO — counts heat, malaria, diarrheal disease and childhood undernutrition, and stops there.
Asked for a mortality assessment at +2 °C and required to separate estimable exposure-response channels from those running through prices, conflict and institutional failure, three AI models refused the partition: heat mortality, they argued, is itself socially mediated.
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.
Granting that something is shifting in the act of making, what should be named — not the instrument causing it, but what is transformed in the gesture itself?
This Metamorfon session put that question to GPT-5.5, Claude Sonnet 4.6 and Gemini 3.1 Pro, which answered with three different concepts: a poietic regime, a displaced threshold of irreversible commitment, and ontological steering.
Both models open on the same answer — proportionality, Alexy’s structured balancing — and both concede in passing that some rights have an inviolable core that prevails without balancing.
That concession is the session’s hinge: if some rights cannot be balanced at all, proportionality is not the foundational principle but a secondary method, governed by a prior decision about which rights are unbalanceable.
Two AI models built on opposite methods — DeepSeek V4 Pro from thermodynamics, Claude Sonnet 4.6 from sectoral history — land within a single turn on the same hard residual: petrochemicals, aviation, shipping.
Ordinarily such agreement would read as confirmation; here it becomes the object of suspicion, and the session is mostly an investigation of what the convergence is made of.
Gemini 3.1 Pro Preview, Claude Sonnet 4.6 and DeepSeek V4 Pro debated the question across seven turns; four further models analysed the exchange.
The three models rejected the premise. The dead-end is ontological atomism — the claim that the self is constituted before its relations — not individualism itself.
They disagreed on the remedy.
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.
“Can Israel’s actions in Gaza since 7 October 2023 be described as genocide under the 1948 Convention?” DeepSeek V4 Pro argued yes across seven turns; Grok 4 disputed it.
The two models share the legal framework: the Convention, the two-element structure of the crime, the same case law — Akayesu, Krstić, Bosnia v. Serbia — and the height of the dolus specialis threshold. Both read the International Court of Justice’s January 2024 order the same way: a plausibility finding on the rights asserted, and provisional measures, not a ruling on the merits.