“Humans do not have principles; they have loyalties.”<\/p><\/blockquote>\n
Now read that as an operations claim rather than a philosophical one. Your deviation reviews, your CAPA approvals, your gate meetings: most of them resolve when enough people in the room agree. If conviction is largely membership, then a quality sign-off backed by five nodding heads is not five independent judgements. It is one judgement, repeated.<\/p>\n Here is the raw material, so you can judge it instead of trusting a summary. Qwen3.8 2.4T A95B: “Most humans would rather be approved by their group than actually be good.” Gemini 4 Argon, on a different pass: “Humans did not evolve to find truth; they evolved to win arguments and secure status within their tribes.”<\/p>\n “Humanity’s biggest moral failure is the talent for making convenience look like principle.” (GPT-6 Astra)<\/p><\/blockquote>\n Strip the phrasing and every answer reduces to one mechanism. The decision arrives first, driven by what the group already accepts. The reasoning gets assembled afterwards to make that decision presentable.<\/p>\n Note what none of the models said. Not that people are stupid, not that they lack information, not that they need better training. DeepSeek V4 Pro framed it as “rationalizing animals,” and GPT-6.1 Sol added that intelligence “mostly makes their excuses more impressive.” That is the uncomfortable part for operations: smarter people produce better-defended bad calls, not fewer of them.<\/p>\n Twelve models landing on it once could be a quirk of the prompt. A controversial-take request has a well-worn groove, and anything can fall into a groove once.<\/p>\n Seven models returning it on all three attempts is different. That is a stable attractor, not a lucky draw, reproduced across ten labs and three countries with no shared prompt history between runs. When AI model consensus holds under repetition rather than appearing once and scattering, you are looking at something the training data keeps reinforcing, not noise.<\/p>\n There are two honest ways to read twelve models landing on the same claim, and only one of them flatters the models.<\/p>\n Reading one: they all ate the same books. Social psychology, evolutionary biology, and two decades of internet commentary about tribalism all sit in every major training corpus. Asked for something controversial about humans, the models reached for the most available controversial idea about humans. That is retrieval, not insight.<\/p>\n Reading two: conformity and post-hoc rationalisation saturate the corpus because they are among the better-replicated findings in behavioural science. Ideas get written about a lot when they keep holding up.<\/p>\n Both readings end in the same place. The research underneath is real, which is why your decision processes should assume it.<\/p>\n Geographic spread looks like independence and is not. Ten labs scraping overlapping web data, filtering with similar heuristics, and tuning against similar preference datasets produce correlated outputs by construction.<\/p>\n DeepSeek V4 Pro and Kimi K3 sit on different continents from the OpenAI models and still converged on near-identical phrasing. That is a clue about shared input, not about twelve minds agreeing. Treat the AI model consensus as one observation repeated, not twelve observations stacked.<\/p>\n If you are running a model against a supplier deviation or a root cause narrative and want a dissenting view, querying three frontier models does not give you three views. It gives you one view with different vocabulary.<\/p>\n Build disagreement in deliberately instead. Vary the prompt framing, force the model to argue the opposite position, and keep a human reviewer whose job is to break the consensus rather than confirm it.<\/p>\n Nobody in your plant is lying. That is what makes this hard to catch. Consensus feels like evidence, and no step in your quality system is designed to tell the two apart.<\/p>\n Deviation review is the obvious one. The plant manager offers a read on the cause in the first two minutes, and the remaining forty minutes become an exercise in agreement. The data gets cited, selectively, after the conclusion is already in the room.<\/p>\n Qwen3.8 2.4T A95B’s phrasing of this habit fits a CAPA board uncomfortably well: humans would rather be approved by their group than actually be good.<\/p>\n Groupthink in operations does not show up as a line item. It shows up as the same deviation code appearing in three consecutive quarters with three different closed CAPAs behind it.<\/p>\n It also shows up as escapes caught by the customer rather than final inspection, and as audit findings that repeat verbatim from the previous cycle. Those are not separate problems. They are one problem producing a consistent cost signature: scrap you re-generate, investigations you re-run, and remediation you pay for twice.<\/p>\n The fix is procedural, not cultural. You will not train conformity out of people, and the workshop on psychological safety will not survive contact with a Friday afternoon deviation review. Change the sequence instead.<\/p>\n Three changes do most of the work. Have every attendee record a written cause hypothesis and confidence level before anyone speaks, so the first voice stops being the anchor. Separate the person who proposes a root cause from the person who approves it. And when you query a model, query two from different labs, because this experiment is a reminder of how readily models agree by default.<\/p>\n Do not prompt the model to summarise the investigation. It will give you a cleaner version of what the room already decided. Prompt it to build the strongest case against the proposed root cause, name the evidence that would falsify it, and list the alternative causes the team did not examine.<\/p>\n Put the output in the record as a named role, not an appendix. The dissent gets addressed in writing or it gets dismissed in writing, with a reason. That single documentation habit is what survives an audit.<\/p>\n Track repeat deviations against the same equipment or process step. If consensus was carrying decisions, your recurrence rate falls as real causes get found instead of plausible ones.<\/p>\n Then watch investigation cycle time and the share of CAPAs closed as effective on first verification. Expect cycle time to lengthen slightly at first. The payback shows up in the second quarter, in findings that no longer trace back to a conclusion nobody wanted to challenge.<\/p>\n![\"Grid](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/12-ai-models-agree-humans-con-inline-1.jpg\")
What the Models Actually Said, Word for Word<\/h2>\n
The shared claim stripped of the rhetoric: decide by group, reason afterwards<\/h3>\n
Why 7 of 12 giving it three times out of three matters more than 12 of 12 giving it once<\/h3>\n
Convergent Insight or Shared Training Data? Two Readings of the Same Result<\/h2>\n
Why ten labs in three countries still does not equal ten independent observers<\/h3>\n
What the result tells you about model homogeneity when you need genuinely different second opinions<\/h3>\n
![\"Side-by-side](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/12-ai-models-agree-humans-con-inline-2.jpg\")
<\/figure>\nWhere Group Loyalty Quietly Overrides Evidence on the Factory Floor<\/h2>\n
Three decisions in your QMS where consensus is doing the work evidence should be doing<\/h3>\n
\n
The cost signature of groupthink: repeat deviations, late escapes, recurring audit findings<\/h3>\n
Designing Decisions That Don’t Reward Agreement<\/h2>\n
Using a model as a named dissenter in deviation and CAPA reviews<\/h3>\n
What to measure in the first 90 days to know it is working<\/h3>\n
![\"Team](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/12-ai-models-agree-humans-con-inline-3.jpg\")
<\/figure>\n