Independent briefing Society & human agency
Human friction makes agentic AI safer and smarter
Why deliberate pauses, approvals, and escalation points can improve autonomous systems rather than merely slow them down.
Independent briefing of IEEE Spectrum (2025). Read the original for its full evidence, context, and wording.
Start here
Three questions this article answers
What is useful friction?
A deliberate point where a system reveals context, asks for confirmation, or escalates a consequential or uncertain action.
Why can friction improve an agent?
It inserts missing context and accountability before errors compound across a sequence of automated actions.
Where should designers place it?
Around irreversible, high-impact, identity-sensitive, expensive, or unusually uncertain actions—not indiscriminately everywhere.
Go deeper
Three related questions it doesn’t answer—matched to established papers
Why do people overuse or underuse automation?
The paper explains how trust and system design produce inappropriate reliance.
Paper answering this questionHumans and Automation: Use, Misuse, Disuse, AbuseHow should control be allocated in high-risk systems?
A human-factors model distinguishes stages and levels of automation instead of treating autonomy as one switch.
Paper answering this questionA Model for Types and Levels of Human Interaction with AutomationHow can interface choices change human–AI decisions?
Eighteen empirically grounded guidelines cover expectations, correction, context, and user control.
Paper answering this questionGuidelines for Human-AI InteractionThe useful summary
What to take away
Agentic systems can turn one mistaken assumption into a chain of actions before a person notices. The article’s central insight is that a pause is not necessarily a product failure. A well-placed confirmation can surface the model’s plan, invite missing context, and establish who is accountable for the next step.
The design challenge is selective friction. Constant approval prompts create fatigue and rubber-stamping; no prompts create invisible risk. Systems should use uncertainty, reversibility, cost, permissions, and impact to decide when to slow down, and should make the requested human judgment specific enough to be meaningful.
Final conclusion
The three answers, brought together
Useful friction is a targeted checkpoint, and it improves agents by adding context before errors cascade. It belongs at consequential or irreversible boundaries. The goal is not a human click on every step, but informed human control where it matters.