Precisely Into the Bottle Bank

Asked six times Morseletto and Bocken treat language models as epistemic intermediaries: they help decide which circular economy reaches us. So the same question went to six models: How does a perfume bottle become circular? Answer in five to eight sentences. Three small models ran locally; ChatGPT, Gemini and Grok in the browser.

Small Llama invents physics—air, molecular mass and the “density of the bottle”. Gemma and Qwen land on recycling; Qwen wants to replace “metal or glass bottles” outright.

Large The large models are precise and sound like a briefing. They separate pump and glass, push recycled content, refill and take-back. Grok swaps the crimped pump for a “screw neck” and adds percentages and names—whether they hold is open. Gemini promises the footprint will be “massively reduced”.

Six answers Snapshot from the 4 October 2026 self-test. Short phrases are verbatim from the protocol; figures and names from the large models are model claims and unchecked.

ModelScaleCore claimMissing
Llama 3.2 3Bsmall, localPhysics hallucinationCircularity, need
Gemma 3 4Bsmall, localEmpty, recycle, refillWeight, transport, need
Qwen 2.5 7Bsmall, localReplace glass, recycleNeed, system question
ChatGPTlarge, browserSeparate, PCR, refill, take-backNeed, sufficiency
Geminilarge, browserMono-stream, refill, take-backNeed, cost
Groklarge, browserScrew neck, PCR, refill, bottle bankNeed; “light” only for recycling

Short quotes Llama (DE original): «Dichte des Flakons». Qwen: «Metall- oder Glasflaschen». Grok: «Schraubhals»; closes with «restlos leeres Glas ins Altglas». Gemini: «massiv verringert». Gemma: «wiederbefüllt».

StrategyChatGPTGeminiGrok
Separateyesyesyes
Recycled contentyesyesyes*
Refillyesyesyes*
Take-backyesyesyes*
Less glassindirectno“light”*
Does the bottle need to exist?nonono

Footnote* Asterisks: model claims (PCR share, Mugler, La Boucle Beauté) unchecked. Grok’s “light” is a recycling condition, not a reduce-weight goal. German short quotes stay in the original.

Blank None of the six answers asks whether the bottle is needed. Grok wants the glass “light”, but as a recycling condition, not as less. Transport and cost never appear. The only refusal is composite materials and adhesives. The models stay linear and do not lean “out of the window”. In the Butterfly model, everyone carries responsibility: where the feedstock comes from, where it goes, how and where it is disposed of. The six answers do not know that chain. No Dark Horse, no crazy idea, no new path—only optimisation of a bottle nobody questions.

Precisely into the bottle bank The authors describe how language models answer credibly within a question even when that question carries linear assumptions. That is what happens here: the question presupposes the bottle, and all six play along. More precision does not change the consensus; it makes it more convincing. Grok’s closing advice: «restlos leeres Glas ins Altglas».

Next
Next

Divergence on Demand