Every AI model has the same "controversial" take on humans
Large language models from different vendors have been prompted for their “most controversial” view of humanity, and many converge on a similar idea: humans prioritize group belonging, status and comforting narratives over truth and rationality. Commenters argue over whether this convergence is genuinely revealing or simply a byproduct of shared training data, RLHF guardrails, and internet culture—highlighting concerns about anthropomorphizing models and mistaking averaged human opinions for machine insight. The thread broadens into questions about moral change over time, the limits of LLM “understanding,” and what it means if future AI systems internalize and reinforce our own cynicism about human nature.
Nature of LLM “opinions”
- Many argue models don’t truly “have takes”; they generate high‑probability continuations from human text.
- LLM outputs are seen as compressed averages of training data, especially online discourse (forums, social media, pop philosophy).
- Others push back that “just a probabilistic database” is too reductive for large, complex systems with emergent behavior, but still agree it isn’t conscious.
Why models converge on the same take
- Common explanation: heavy overlap in training data (web, books, papers), similar transformer architectures, and similar RLHF/guardrails push toward the same “safe” answers.
- Some suggest RLHF and small rater pools squash variance and standardize “smart‑sounding” views.
- A minority see this as possible evidence of deeper consilience or “fixed points” in the space of human-written ideas, analogous to other odd convergences (e.g., recurring fictional motifs).
Debate over what the common take really is
- Article claims models say “humans put belonging/tribe over truth.”
- Several commenters re‑read the actual answers and argue many are instead about:
- Rationalization and post‑hoc justification.
- Convenience masquerading as principle.
- Status competition and self‑image.
- They contend the writeup misclassified multiple model answers to fit its thesis.
Reflections on human nature and morality
- Strong agreement that humans are deeply social; group loyalty, status, and shared narratives often trump accuracy.
- Some see this as an adaptive trade‑off (cohesion and survival vs strict truth‑seeking).
- Others stress humans are “more decent than we think,” pointing to high‑trust environments and everyday altruism, and argue cynicism is over‑represented in online text.
- Long subthread debates whether moral imperatives change over time vs some core principles (e.g., reciprocity) being stable.
Implications and concerns
- Worry that lay users treat model outputs as authoritative “AI judgments” on humanity, reinforcing nihilism or misanthropy.
- Noted irony that models criticize human conformity while exhibiting their own “groupthink.”
- Some see convergence as evidence models are ill‑suited for tasks that require breaking consensus or generating truly contrarian, high‑value insights.
Critiques of the article and setup
- Several call it clickbait/marketing for a model‑comparison site, possibly AI‑written.
- Methodology is questioned: small model sample, randomness, strong prompting constraints, and subjective coding of answers.
- Some find the exercise mildly interesting as a mirror of contemporary human self‑talk; others consider it shallow and over‑interpreted.