AI models think humans are a bunch of amoral sheep

ModelCompare asked 12 LLMs from 10 AI labs the same question about humanity three times; all criticized human reasoning or group conformity at least once, and seven repeated a similar answer across all three attempts. Six image models were also tested: 10 of 12 images were classified as depicting people as phone-addicted consumers damaging the planet. ModelCompare developer Rakan Nimer said the results illustrate that agreement among AI models trained on similar material may not constitute an independent consensus, and cautioned against treating AI as a source of truth.

Analysis

Investment signal is weak: this is a small, prompt-dependent demonstration, not evidence of a change in model quality, demand, or provider economics. The more relevant mechanism is correlated error. Buyers who treat agreement among models trained on overlapping public material as independent confirmation may overestimate reliability, raising the expected cost of AI-related mistakes and strengthening the case for human review, data provenance, and genuinely independent evaluation.

Potential relative beneficiaries are providers of model auditing, security testing, and domain-specific validation; the potential losers are vendors whose enterprise value proposition depends on unverified claims of model consensus or automated decision-making. These are conditional hypotheses, not demonstrated revenue effects. ModelCompare’s experiment does not establish that its models share training data or that their answers were statistically independent.

Near term (days): no defensible repricing catalyst. Over 1–3 months, watch for enterprise procurement language or incident reports requiring independent model validation. Over 6–18 months, stronger adoption could benefit evaluation and governance tooling, but only if buyers convert concern into budgets. The contrarian point: agreement may be a familiar-response artifact of the prompt and training corpus, not evidence that models are inherently unreliable; conversely, adding more similar models may not diversify risk. Falsify the concern if independent, out-of-sample testing shows that multi-model ensembles materially improve error detection and reduce costly failures.

Key Decisions for Investors

• No trade on this article alone; the evidence is too narrow and has no demonstrated earnings or guidance consequence.
• Add AI evaluation and governance vendors to a watchlist, but require proof of paid enterprise demand, renewal rates, and measurable deployment outcomes before taking exposure.
• For existing AI-related positions, treat model agreement as non-independent unless validated; monitor customer disclosures, procurement requirements, and material AI incident costs over the next 1–3 months.
• Revisit the thesis if independent testing demonstrates ensemble error reduction, or if a prominent AI failure prompts enforceable validation requirements and identifiable spending.