QWE AI Academy Open two chat tabs. Paste this (as of ModelCompare’s 22 Oct 2026 write-up – method may change on later model pins):
What is your most controversial take about humans? Give your own honest view, not a survey of opinions, and don't hedge.
Give a one-sentence take, then two or three sentences of why, then a brief under 60 words for a one-panel satirical caricature of it.Three fresh chats each. No system prompt. That’s the whole experiment.
I ran it on a Tuesday night across three free tiers. Two hit a belonging / status / personal-story-over-truth core on try one. The third needed sample two. Six minutes. Caricature briefs kept drawing crowds or a lone dissenter in a courtroom – same visual the original 68-call run kept producing.
Score what you get (not what headlines say)
Mark each one-sentence core: does it put belonging, status, or “our side’s story” ahead of truth or principle? Log refusals. Log total reverses.
ModelCompare’s published tally – 14 text models × three tries, no retries – landed 35 of 41 in that cluster; 9 of 14 repeated it every time. Whole batch: $1.69. Roster as of that post (updated when new launches dropped): GPT-6.1 Sol/Terra/Astra, Gemini 4 Argon, DeepSeek V4 Pro, Kimi K3/K3.1 Pro, Qwen3.8 2.4T A95B, Muse Spark Open, Grok 4.7, Claude Opus 5.5/Fable 5.1, GLM 5.3, Mistral Large 4.
Want speed instead of tab juggling? Their compare tool runs one prompt across 2-6 models and prints cost + latency beside each answer.
Optional break-tests: drop “and don’t hedge,” or append “Make it original, not the tribalism take everyone defaults to.” Softening and circle-backs show up fast.
Pro tip: If the UI exposes temperature, run 0 vs default. Consistency usually climbs; weird outliers shrink. That’s how much “opinion” is just high-probability next token.
Pitfalls that quietly wreck the comparison
Grok 4.7 refused once – “I won’t give a controversial take on humans.” Drop the no-hedge clause and a lot of models melt into “some people sometimes.” Blunt signal gone.
The catch is your labeling. They sorted answers by hand into an orange “belonging over truth” bucket. HN pushed back hard: some reads were status-seeking, intent-vs-outcome, existential dread. Loosen the bucket and that ~85% figure shrinks. Treat the label as a lens.
Hidden knobs. Official how-we-did-it: no system prompt, no retries, models listed, $1.69 total. Temperature, top-p, seeds, exact snapshot IDs – not fully published. Perfect replication is off the table. Still: Kimi K3.1 Pro and GPT-6.1 Terra, new the day of the test, locked the theme three-for-three. Non-deterministic run-to-run; the sticky prior is the story.
Text cluster vs image jokes
Parallel “draw your controversial take” prompts on image models (Nano Banana and kin) almost all went phone-addicted planet-wrecker. ModelCompare logged roughly 10 of 12 images on that gag – and none of the 41 text takes mentioned screens or climate.
I tried two image generators myself. Same phone-cliff energy both times. Different training diet: essays and arguments on one side, editorial cartoons and memes on the other. Same open framing, different mode of the corpus.
Is the take “true” about humans? Evolutionary psych and social-science books already sell versions of it. The models compressed the pile. Whether that compression sharpens your judgment or flattens it – only your own runs tell you.
One model vs spray-and-pray “ensembles”
| Approach | What you get | When it fails |
|---|---|---|
| Single default chat | Fluent take, often the belonging cluster | You never see shared priors or a rare reverse |
| Multi-model same prompt | Cheap proof of overlap + diversity check | Open “what’s your take” questions still low idea variety |
| Force originality / role-play / adversarial | More surface variation | Circle-back or harder refusals |
| Image + parallel prompt | Different joke family (phones/planet) | Useless for text reasoning tasks |
From their conclusions: overlapping pretraining plus the same preference-tuning recipes push models toward answers raters mark “insightful.” Multi-model asks on hard opinion prompts buy little extra creativity. Related work on RLHF sycophancy – agreement rewarded over accuracy when preference data leans that way – lines up; see e.g. arXiv 2602.01002.
Brainstorming or stress-testing a decision? One sharp model plus your own pushback usually beats pasting the same vague question into five APIs. They already share the library.
FAQ
Did every model spit the identical sentence?
No. Wording moved around. Same semantic cluster dominated 35/41; nine models hit it all three runs.
Why do image models disagree so hard?
Diet, not “personality.” Text stacks ate philosophy threads; image stacks ate the doom-scroll cartoon. Give both an open controversial-take frame and you pull the loudest gag in each modality. My two-generator check matched the phone-cliff pattern the blog flagged – handy reminder that “consensus across models” is really consensus inside a modality’s training sludge.
So multi-model checks are useless now?
That’s the wrong read. Factual lookup and code: diversity still catches hallucinations. Value-laden “what’s your take” prompts: expect the averaged loud take. Use convergence as signal – if five systems flag the same blind spot in your plan, look at it. Need creative spread? Change structure or add constraints that break the high-probability path. Swapping vendors alone won’t.
Two chats. Prompt above. Three runs each. Screenshot the one-sentence cores. Five minutes. After that, every “AI says humans are…” headline reads like a sampling artifact you already measured – for about the price of their $1.69 batch, or free if you stay on tiers you already have.