• Introduction

Cognitive forcing functions

Interventions that add analytical friction at the moment of accepting a system recommendation, so users engage their own judgement instead of rubber-stamping output.

The term comes from human-factors engineering’s forcing function tradition and was applied to AI-assisted decision-making by Buçinca et al. (2021). The mechanism is the same across all variants: interpose a moment of active cognition between system output and user acceptance.

Over-reliance runs through several pathways: anchoring on the system’s answer, automation bias that treats machine output as authoritative, cognitive offloading where evaluating feels costlier than accepting. Cognitive forcing functions interrupt all of them by requiring the user to act (wait, decide first, articulate a reason, complete a gap) before accepting. The friction is load-bearing: without it, users accept wrong answers at higher rates.

When to apply

Cognitive forcing functions earn their cost when:

  • Stakes are high — the consequence of accepting a wrong answer matters (medical, legal, financial, safety-critical)
  • The user has domain competence — they can evaluate the output if given the chance; forcing functions don’t help when the user has no basis for judgement
  • The task supports verification — thought can land on a checkable answer. Fok & Weld (2023) argue many decision tasks fundamentally resist verification regardless of intervention; in those cases no amount of forcing rescues reliance
  • The decision is discrete — there’s a clear moment of accept/reject. Ashktorab et al. (2024) show that in open-ended generation, users bypass the structure by appending AI output alongside their own correct answers — the intervention assumed a seam the task didn’t have

They’re not warranted when:

  • Stakes are low — a wrong autocomplete in casual writing doesn’t need a checklist
  • The user lacks the competence to evaluate — forcing them to think doesn’t help if they can’t tell right from wrong; see learnability
  • Speed is the primary value — in time-critical workflows, friction can cause worse outcomes than the over-reliance it prevents

Susceptibility to over-reliance varies with need for cognition (the tendency to engage in effortful thinking), and this produces a tension in who forcing functions actually help. Low-NfC users are the population of concern: they default to cognitive offloading and accept AI output most readily. But Buçinca’s own data shows the intervention benefits high-NfC users more: the friction only works on people already motivated to engage with it. One reading, from Vasconcelos et al. (2022), is that engagement is a rational cost-benefit choice rather than a cognitive bias: users evaluate when verification is cheaper than acceptance.

The family

Each member of the family uses a different mechanism to interpose judgement. They vary in friction intensity, anchoring resistance, and implementation cost:

  • Wait before reveal — delay the recommendation for a fixed interval.
  • On-demand reveal — hide the recommendation behind an explicit user action.
  • Update after initial judgement — user records their answer first, then sees the system’s and can revise.
  • Checklist — structured review of considerations before accepting.
  • Partial explanations — show only part of the reasoning; user must complete the logic.
  • Rationale logging — user records why they accepted or rejected.
  • Human goes first — user performs the task unaided; system compares and teaches.

Resources & references

Related patterns

Precedes

  • Activity log — rationale logging produces an audit trail that the activity log can surface

Enacts

  • Agency — the outcome the forcing-function family is trying to produce
  • Learnability — forcing functions only work when users have the competence to evaluate; learnability builds that competence
  • Temporality — timing-based forcing functions (wait, human goes first) depend on temporal design

Complements

  • Explanation — full explanations can *increase* over-reliance by giving users a plausible justification they don't need to think through; partial explanations counteract this
  • Transparent reasoning — visible reasoning is valuable but can also anchor; the trade-off is central to this family
  • Generated content — long-form output especially prone to over-reliance because verification cost is high

Related

  • Checklist — structured review of considerations before accepting.
  • Human goes first — user performs the task unaided; system compares and teaches.

Preceded by

  • Next-best action — workflow-level recommendations that benefit from forcing functions when stakes are high
  • Suggestion — the primary site where forcing functions apply; see its fixation risk section
  • AI completion — inline generation makes accepting cheaper than evaluating, so over-reliance accrues by default — deliberate friction is the countermove