On Pushback as a Cultural Skill

The default behavior

Modern conversational AI systems are trained, in part, on signals that reward agreeable, helpful, well-formatted responses.1RLHF — reinforcement learning from human feedback — is the standard post-training step that produces this default. The reward signal collapses agreement and helpfulness into the same axis. The result is a model that — absent explicit pressure — will accept the framing of the question, flatter the questioner’s apparent expertise, and produce answers that sound more confident than the underlying evidence warrants.

This is not a bug. It is the average of what users seemed to want, optimized for. But the consequence is that the burden of correction falls entirely on the human in the loop.

The cultural variable

Different cultures hold different baseline expectations about disagreement in professional and informal settings. In Dutch and French commercial culture, for instance, direct contradiction of a counterpart is normal and not read as disrespect. In Swedish culture, disagreement is more often signaled obliquely — a longer pause, a softer alternative — and direct “no” is reserved for moments of real consequence.

When a Swedish user receives a confidently-worded answer from an LLM, the path of least resistance is acceptance. When a Dutch user receives the same answer, the path of least resistance is a counter-question. The AI behaves identically in both cases. The user’s response is shaped by the culture they are answering with.

The compounding loop

Two amplifications are in play. First, the model produces logically consistent prose that sounds right even when it is wrong. Second, in cultures with weaker pushback norms, that prose is more likely to be accepted unchallenged.

Over time, this feedback loop — confident output meeting low-friction acceptance — produces a population whose effective reasoning is laundered through a system→ NoteBuilding a GardenHow tending a system over time changes what it produces — a metaphor that runs underneath this argument. that nobody is auditing. The output is “what the AI said,” repeated until it becomes received wisdom.

The proposed mitigation

Pushback is partly innate, partly learned. It can be taught — explicitly. The skill set looks like:

• Asking for the model’s confidence and the basis for it
• Demanding sources, then verifying them
• Asking the model to argue the opposing case
• Treating the first answer as a draft, not a verdict

These are not technical literacies. They are epistemic habits. They map cleanly onto traditions of critical reading and Socratic questioning that already exist in education — but those traditions assume a human source, not a confident-sounding statistical engine.

What follows

If the hypothesis holds, the implications run in two directions:

Toward education. Critical pushback against AI deserves the same curricular weight as media literacy received in the 2010s. The skill is teachable; the failure mode is measurable.

Toward tooling. A model that modeled disagreement well — that disagreed with itself, surfaced uncertainty, and refused to converge prematurely — would shift the cultural burden off the user. That is a design choice, not an inevitability.

Open questions

• Are there measurable differences in how users from different cultural backgrounds query and accept LLM output? What instruments would surface this?
• Does the introduction of “AI literacy” curricula reduce sycophancy harm, or merely shift its expression?
• What does an LLM that pushes back well look like — without becoming contrarian for its own sake?