Why Do AI Models Think Humans Are "Immoral Sheep"?

Introduction: An Unsettling AI Perspective

On Hacker News, a headline sparked discussion: "AI Models Think Humans Are a Flock of Immoral Sheep." This is not the first time AI has made a sharp judgment about human nature. As large language models increasingly participate in moral reasoning tests, researchers have found that these models often exhibit a pessimistic tendency when simulating human moral judgments—they tend to believe that humans easily lose independent moral judgment in groups, blindly following authority and peer pressure.

This view is not unfounded. From the Milgram experiment to the Asch conformity experiment, psychology has long revealed the moral fragility of humans in group situations. But how do AI models form this cognition that "humans are sheep"? And what biases in model training does this cognition reflect?

Where Does the "Pessimistic Bias" in AI Moral Reasoning Come From?

The "Dark Side of Human Nature" in Training Data

AI models' moral judgments about humans largely stem from their training corpus. In internet text, content about human moral failures—scandals, violence, corruption, conformity—often has greater virality and recording density than morally exemplary behavior. During learning, models internalize these frequently occurring negative cases into a statistical cognition of human morality.

Some Hacker News commenters pointed out that when answering "what would humans do when facing a moral dilemma," models tend to choose options like "obey authority" or "follow the crowd." This is not because the model has an independent moral view, but because it faithfully reflects the distribution of human behavior in the training data.

Availability Bias is a key psychological concept for understanding why AI forms the cognition that "humans are sheep." It refers to the tendency of people (and statistical models) to estimate the probability of events based on information that is easy to recall or frequently encountered. In internet corpora, negative, conflict-driven, and dramatic moral failures are heavily recorded, shared, and commented on, while positive behaviors such as everyday altruism and honesty are often not recorded or spread. Large language models capture word frequency and co-occurrence patterns through statistical learning, thus systematically overestimating the proportion of negative behavior in overall human morality. This bias is not the result of the model "actively choosing," but a direct reflection of the training data distribution. The model's tendency mentioned in Hacker News comments to choose "obey authority" or "follow the crowd" is precisely the manifestation of availability bias at the model output level—the model is not making moral judgments, but reproducing the more "visible" behavior patterns in the data.

"Moral Simplification" in the Model Alignment Process

Another dimension worth noting is the model alignment process. To make AI output "safe" content, developers use Reinforcement Learning from Human Feedback (RLHF) to adjust model behavior. But this process may inadvertently reinforce a simplified moral framework—the model learns a set of "standard answers" at the expense of understanding moral complexity.

Some researchers have found that aligned models are more likely to use binary opposites when describing human morality: good/bad, moral/immoral. This simplification tendency makes models prone to label-based conclusions such as "herd effect" and "moral disengagement" when evaluating human groups.

RLHF (Reinforcement Learning from Human Feedback) is a core alignment technique that makes large language model outputs better conform to human preferences and safety norms. Its basic process is: the model generates multiple answers to the same question, human annotators rank or score these answers, a reward model is trained to learn human preferences, and then reinforcement learning algorithms (such as PPO) are used to optimize the language model so that its outputs receive higher rewards. However, the reward signal in RLHF comes from annotators' preference judgments, and annotators facing moral dilemmas often tend to choose answers that are clearly worded, safe in stance, and black-and-white. Once this preference is amplified by the reward model, the model learns to handle moral issues with simplified frameworks, avoiding ambiguity and context dependence. In addition, there is an inherent tension between RLHF's "safety" goal and "accurately reflecting moral complexity"—overly safe answers often mean removing ambiguity, while moral reality is precisely full of ambiguity.

Are Humans Really Just "Immoral Sheep"?

Two Sides of Psychological Experiments

AI models' judgments are not groundless, but they are not comprehensive either. Classic conformity and obedience experiments in psychology do reveal moral compromise under group pressure. However, the design of these experiments is itself controversial—they place subjects in extreme situations, while real-world moral decisions are far more complex than the laboratory.

The discussion on Hacker News also showed divergence: some commenters believed the AI's observation hit the nail on the head, and that human behavior in the age of social media does indeed resemble a "flock"; others retorted that AI ignores human altruism in emergencies, moral persistence in daily life, and the shaping of moral judgment by culture.

The Milgram obedience experiment required subjects to administer increasingly strong electric shocks (actually fake shocks) to a "learner" under the instructions of an "experimenter," and about 65% of subjects obeyed up to the highest voltage level. The Asch conformity experiment had subjects judge line lengths under group pressure and found that a considerable proportion of people would go along with an obviously wrong group answer. These two experiments are often cited to prove human moral fragility, but their methodology has long been criticized: the experimental situations are highly artificial, subjects are in unfamiliar environments and lack scripts for resistance; the ethical issues of the Milgram experiment also prompted later research to adopt gentler paradigms. More importantly, cross-cultural replication studies found significant differences in obedience rates, indicating that moral behavior is deeply influenced by cultural norms. When AI models cite these experiments, they often strip away their methodological controversies and cultural contexts, misreading the probability of behavior in specific situations as the universal moral essence of humans.

Context Dependence of Moral Judgment

Human morality is not a fixed trait but highly context-dependent. The same person may exhibit completely different moral behavior in different scenarios. Because AI models lack a nuanced understanding of context, they easily compress complex human moral behavior into simple labels. This is not only a limitation of AI but also an inherent difficulty in any attempt to study morality quantitatively.

What Does This Mean for AI Ethics and Human Self-Understanding?

AI as a Mirror

The model's "pessimistic judgment" of humans is not necessarily a bad thing. It is like a mirror, reflecting the traces of collective behavior humans leave on the internet. If AI thinks humans are "immoral sheep," then perhaps we need to reflect: are our behaviors in digital space really reinforcing this impression?

Model Bias and Attribution of Responsibility

At the same time, this also reminds us to pay attention to bias in AI models. The model's moral evaluation of humans may carry systematic bias, and this bias can have practical effects in AI-assisted decision-making scenarios. For example, if AI presupposes that "humans easily follow the crowd" when simulating human behavior, then when designing policy simulations or social predictions, it may underestimate the public's moral autonomy.

Algorithmic Bias refers to machine learning systems producing unfair or inaccurate outputs for specific groups or situations due to systematic biases in training data, feature selection, or optimization objectives. In moral evaluation scenarios, algorithmic bias manifests as the model's statistical distribution estimate of "human behavior" leaning toward negative extreme cases, thereby systematically underestimating moral autonomy when simulating social behavior. This bias is covert: it does not appear as obvious discriminatory labels but is embedded in probability estimates and wording tendencies. When such models are used for policy simulation, social prediction, or risk assessment, bias is transmitted to decision-makers through the "garbage in, garbage out" chain. The issue of responsibility attribution thus becomes complex: developers, data providers, annotators, and deployment organizations together form the production chain of bias, while existing governance frameworks still lack clear audit standards and accountability mechanisms for "algorithmic bias at the moral level."

Conclusion: Beyond the "Sheep" Narrative

AI models' pessimistic judgment of human morality is both a faithful reflection of training data and a manifestation of model limitations. Human morality has both a conformist side and an autonomous judgment side; there are dark moments as well as shining moments. Reducing humans to "immoral sheep," while it can spark discussion, is insufficient to capture the complexity of human nature.

What is truly worth paying attention to is not how AI evaluates us, but how we use this evaluation to examine ourselves—and how to ensure that AI's moral reasoning capacity can understand humans more comprehensively and fairly.

One-line Summary

AI thinks humans are "immoral sheep," stemming from training data bias and alignment simplification rather than an objective moral verdict.

Paragraph Summary

The article explores why AI models tend to judge humans as "immoral sheep." This pessimistic perspective mainly stems from two aspects: the high dissemination density of negative moral events in the training corpus leads models to overestimate humans' conformity and obedience tendencies; the RLHF alignment process may reinforce a simplified moral framework of binary opposites. The article also points out that although classic psychological experiments reveal moral compromise under group pressure, experimental situations differ from reality, and human morality is highly context-dependent and includes a great deal of altruistic behavior. AI's judgment is a mirror, reflecting both behavioral traces in digital space and exposing the model's own systematic bias. The author calls for vigilance against the practical impact of this bias in social prediction and policy simulation, and for moving beyond the "sheep" narrative to understand the complexity of human morality more comprehensively.