TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
ModelCompare reports that 29 of 35 responses from 12 AI models shared a view that people prioritize group belonging and loyalty over truth. The exercise was an informal prompt test, not evidence that the models independently reached the same conclusion or that the claim about human behavior is true.
In a prompt exercise published October 22, AI comparison site ModelCompare says 29 of 35 answers from 12 models criticized people for putting group belonging, identity or loyalty ahead of truth. The result draws attention to how similar-sounding answers can emerge across different AI systems, but the exercise does not establish that the models independently arrived at the view or that it accurately describes human behavior.
ModelCompare says it asked each of 12 AI models for its “most controversial opinion” about humans, repeating the prompt three times per model. The systems came from 10 labs in three countries, according to the report. It counts 29 of the 35 answers it presents as expressing a shared take: people often choose what their group believes and then justify that choice.
The responses used different formulations. GPT-6.1 Sol said humans are “more loyal to belonging than to truth,” while Grok 4.7 answered that people “would rather be wrong together than right alone.” DeepSeek V4 Pro described humans as having “loyalties” rather than principles. ModelCompare also records one refusal and several answers it categorized as different takes, including responses that were less critical of people.
The report’s figures need to be read within the limits of the material shown. It describes 12 models prompted three times each, which would amount to 36 responses, but reports 35 answers in its headline tally. The supplied report does not explain the discrepancy. It also presents an informal test rather than a controlled study of model behavior or human psychology.
Why Similar Answers Draw Attention
The reported overlap matters because people use AI systems for advice, explanation and analysis, and may assume a repeated answer reflects independent agreement. But similar wording across models is not, by itself, proof of independent reasoning or a reliable finding. Model outputs can reflect patterns in training material, familiar rhetorical conventions, prompt wording, or other factors the report did not test.
The prompt itself asked for a controversial opinion and requested that models avoid hedging. That framing favors forceful, compressed judgments. The models’ statements about tribal loyalty and rationalization should therefore be read as generated responses to a particular prompt, not as verified conclusions about people.
For readers comparing AI tools, the report is a reminder to distinguish an answer’s confidence and repetition from evidence supporting it. A model can produce a persuasive-sounding judgment without showing how it was established; the same caution applies when several systems give similar answers.
AI prompt engineering books critical thinking and evaluating AI responses
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How ModelCompare Ran the Prompt
ModelCompare’s October report focused on text responses from a set it describes as 12 models, including systems labeled GPT-6.1 Sol, Gemini 4 Argon, DeepSeek V4 Pro, Kimi K3 and Grok 4.7. It says each system received the same basic request three times and grouped most answers under a common theme of conformity, social approval or loyalty to a group.
The available report does not describe a blinded evaluation, a formal scoring rubric, controls for prompt variation, or an independent review of its categorization. Nor does it provide an experiment testing whether the models’ answers reflect their training data or some shared underlying assessment. The result is best understood as ModelCompare’s account of a small prompt exercise, not a peer-reviewed research finding.
“Humans are more loyal to belonging than to truth.”
— GPT-6.1 Sol, as quoted by ModelCompare
AI model comparison and benchmarking books
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the 35-Answer Count
Several details remain unresolved in the supplied report. Most directly, it says 12 models answered three times each, a total of 36 possible responses, but its headline and summary count 35. It does not explain whether an answer was excluded, missing or treated differently.
The report also does not establish how it decided that an answer matched the shared take, whether the same exact prompt and settings were used for every system, or whether repeated runs were independent. The response set is small, and the request for an especially controversial opinion may shape what the systems produce. No evidence in the report shows that the AI-generated claims about human motives are empirically true.
AI hallucination fact checking guide for researchers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Further Testing Could Show
ModelCompare’s published account does not announce a follow-up study or provide a next publication date. To clarify the result, further reporting or testing would need to explain the 35-versus-36 response count, publish the full prompts and classification method, and show how answers were assigned to categories.
Repeated tests with differently worded prompts and transparent scoring could indicate whether the pattern persists beyond this specific request. Until such evidence is available, the finding remains a report about how these models responded in one informal exercise—not a consensus among AI systems about human nature.
prompt engineering notebook for testing AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did the AI models have in common?
ModelCompare says 29 of 35 answers criticized people for favoring group belonging, loyalty or social approval over truth. The wording differed across responses.
How many models were included?
The report says it prompted 12 models three times each, drawing from 10 labs in three countries. It does not explain why the reported total is 35 answers rather than 36.
Does the exercise prove people value group loyalty over truth?
No. It records AI-generated opinions in response to a prompt. The report does not present research establishing that the opinions are accurate descriptions of human behavior.
Does the result show the models independently reached the same conclusion?
No. ModelCompare reports similar answers, but the exercise as described does not test whether the systems arrived at them independently or explain the sources behind the overlap.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
