The Strange New Value of Human Friction in an AI-Accelerated World

---

AI is making many forms of knowledge work easier to start, faster to execute and cheaper to scale. Drafts appear in seconds. Code can be generated before a developer has finished describing the problem. Research assistants can summarize material that once required hours of reading.

That creates an unexpected problem: when the cost of producing an answer falls, the value of stopping to question that answer may rise.

This is what can be called human friction the deliberate moments when people slow down, verify, disagree, rethink, edit, test, or decide that a machine-generated answer is not good enough. It is not an argument for rejecting AI or returning to manual work. It is an argument for recognizing that some of the effort AI removes was doing useful cognitive and organizational work.

Recent research illustrates the tension. Generative AI has produced substantial productivity gains in some real-world settings, while other experiments have found that AI can perform unevenly across tasks or even slow experienced professionals down. A 2025 randomized study of experienced open-source developers, for example, found that participants took 19% longer on average when AI tools were allowed, despite expecting the tools to make them faster.

The emerging lesson is not that AI is inherently productive or unproductive. It is that productivity depends increasingly on where humans place the friction.

Key Takeaways

• AI can reduce the effort required to produce work without necessarily reducing the effort required to judge its quality.
• Research shows that AI performance varies sharply by task, creating a “jagged” boundary between useful and unreliable automation.
• Human verification becomes more important when AI output is easy to accept but difficult to independently evaluate.
• Some friction protects learning, expertise and judgment by forcing people to formulate, question and reconstruct ideas themselves.
• The best AI workflows may not eliminate human effort; they may deliberately redirect it toward decisions machines handle poorly.
• Organizations should measure the quality and consequences of AI-assisted work, not simply how quickly work is produced.

When Faster Stops Being the Same as Better

The strongest argument for AI is straightforward: it can remove unnecessary work.

That benefit is real. In a field experiment involving 5,179 customer-support agents, researchers found that access to a generative AI assistant increased issues resolved per hour by 14% on average, with a 34% improvement among novice and lower-skilled workers. The effects were not uniform: experienced and highly skilled workers saw much smaller gains.

That distinction matters.

If AI helps a less experienced employee access the knowledge of more experienced colleagues, automation can reduce an old form of organizational friction: waiting, searching, asking and repeatedly solving problems that someone else has already solved.

But another type of friction remains.

Someone still has to determine whether the answer is appropriate for the particular customer, codebase, legal situation, scientific question or business decision.

That is where the simple equation of less effort = more productivity begins to break down.

A Harvard Business School and Boston Consulting Group experiment involving 758 consultants found that GPT-4 improved performance on tasks within what researchers called a “jagged technological frontier.” But on tasks outside that frontier, AI assistance could make performance worse.

The implication is important for managers: AI capability is not a single number that can be applied uniformly to a job.

A task can look similar to another task while requiring a completely different kind of judgment.

The Hidden Cost of Removing Every Pause

Human beings have always used tools to reduce cognitive effort. Calculators removed arithmetic from many jobs. Search engines reduced the need to memorize information. GPS reduced the need to remember routes.

Generative AI goes further because it can participate directly in the production of language, code, analysis and ideas.

That creates a different kind of dependency.

When a person has to formulate a problem, search for evidence, compare alternatives and construct an answer, those steps create opportunities to notice contradictions. When an AI system produces a polished answer immediately, those intermediate opportunities can disappear.

Research into automation bias predates generative AI. A systematic review of 74 studies found that people can become over-reliant on decision-support systems, with workload, task complexity, time pressure, trust and confidence among factors affecting the phenomenon. The review also identified training and explicit user accountability as potential mitigations.

The relevance to generative AI is not that every chatbot interaction produces automation bias. It does not.

The more useful lesson is narrower: when a system supplies an answer, humans still need an active mechanism for deciding whether the answer deserves trust.

That mechanism takes time.

In other words, verification is friction.

And friction has a cost but so does skipping it.

The Evidence Is Becoming More Interesting Than the Hype

The most revealing AI research increasingly produces mixed results rather than universal productivity claims.

Consider software development.

In 2025, the nonprofit research organization METR conducted a randomized controlled trial involving 16 experienced open-source developers and 246 real tasks in repositories they knew well. Developers were randomly assigned tasks where AI tools were either allowed or prohibited. The study found that allowing AI increased completion time by 19%. Participants themselves had expected AI to make them substantially faster.

That result should not be interpreted as evidence that AI coding tools generally make developers slower. The study examined a specific group, specific tasks and AI tools available in early 2025. METR itself subsequently changed its experimental design after finding that later data were affected by selection effects as AI adoption increased.

But the result exposes something valuable.

AI-generated code is not the same thing as completed software work.

A developer working inside a mature codebase has to understand architecture, conventions, dependencies and unintended consequences. Generated code must be inspected, integrated and tested. A tool can therefore add output while also adding review work.

The machine reduces one kind of friction and creates another.

That is increasingly the central question for AI adoption.

Human Friction Can Be Designed

The answer is not to make people manually perform tasks that machines can perform well.

Instead, organizations can deliberately decide where human effort is most valuable.

The U.S. National Institute of Standards and Technology makes a similar distinction in its AI Risk Management Framework. NIST emphasizes defining human roles and responsibilities around AI systems and recognizes that human-AI arrangements can range from fully manual to highly autonomous. It also warns that representing complex human phenomena through data and models can remove context that matters for understanding impacts.

That suggests a practical model for AI-assisted work:

Automate production. Preserve judgment.

For a low-risk task, the human checkpoint might be minimal.

For a high-consequence task, it might include independent verification, a second reviewer, source checking, testing or explicit approval.

The friction should therefore be proportional to the consequences of being wrong.

A useful workplace framework might look like this:

AI can usually move faster	Humans should often slow down
Formatting and transformation	Defining the actual problem
First-draft generation	Evaluating factual accuracy
Routine classification	Interpreting ambiguous cases
Repetitive coding tasks	Reviewing architecture and consequences
Summarizing known material	Checking whether important context was omitted
Brainstorming alternatives	Selecting which alternatives deserve pursuit
Routine data processing	Deciding what the data actually means

The distinction is not absolute. AI can assist with many items on both sides.

The important point is that the human role changes from doing every step to deciding which steps cannot safely be delegated.

The Skill Problem: What Happens When the Machine Does the Repetition?

There is another reason friction may matter: skills are partly maintained through use.

A person who repeatedly writes, calculates, debugs, researches or analyzes develops mental models of the work. Those models make it easier to recognize errors and handle unfamiliar situations.

If AI performs the difficult parts continuously, people may become more dependent on the system precisely because they have fewer opportunities to practice the underlying skill.

A small 2025 MIT Media Lab study provides an early, highly qualified signal in this direction. Researchers compared participants writing essays with an LLM, a search engine or no external tool and used EEG measurements alongside behavioral analysis. The study reported different patterns of neural connectivity across the groups and lower reported ownership of essays among the LLM group. The researchers explicitly cautioned that the experiment was limited to an educational essay-writing context and that its findings should not be generalized broadly to other tasks or AI systems.

That limitation is crucial.

It would be premature to say that ChatGPT, or AI generally, causes lasting cognitive decline based on this study.

But the research raises a legitimate question: if people routinely outsource the most demanding portions of a task, which parts of the underlying skill do they continue to practice?

That question applies far beyond education.

The New Productivity Metric May Be Judgment Per Unit of Automation

The first generation of workplace AI measurement focused heavily on speed: minutes saved, tasks completed and output generated.

Those measurements remain useful.

They are simply incomplete.

Microsoft’s 2026 Work Trend Index points toward a different emphasis. Its research reports that AI users increasingly identify quality control of AI output and critical thinking as especially important human skills. It also reports that many users treat AI output as a starting point rather than a final answer.

That suggests a subtle change in what expertise means.

A strong professional may no longer be the person who personally produces every sentence, spreadsheet formula or line of code.

Instead, expertise may increasingly be expressed through knowing what to delegate, what to inspect, what to reject and what to solve personally.

That is a much harder capability to measure.

It also explains why the most valuable human skill may not be speed.

It may be judgment under conditions of abundance.

When producing ten possible answers is almost free, selecting the right one becomes more consequential.

The Business Case for Deliberate Friction

Companies should therefore be cautious about treating every reduction in human effort as an efficiency gain.

A workflow that removes ten minutes of drafting but adds fifteen minutes of checking is not necessarily more productive.

A customer-service system that answers more queries but produces more escalations may shift rather than eliminate work.

A coding agent that generates thousands of lines of code quickly can create a maintenance burden if humans cannot adequately understand or test what it produces.

And an organization that automates decision-making without preserving accountability may discover that the hardest part of the process was never generating the recommendation.

It was deciding whether the recommendation should be trusted.

This is why NIST’s AI framework emphasizes ongoing risk management across the AI lifecycle rather than treating deployment as the end of the process. Its core functions govern, map, measure and manage are designed around continuous assessment rather than one-time approval.

For businesses, the practical question becomes:

Where does human involvement create enough value to justify its cost?

That is a better question than simply asking how much work AI can eliminate.

What Good Friction Looks Like

Not all friction is useful.

Some administrative processes, repetitive approvals and duplicated data entry should disappear.

The goal is not to make work slower for the sake of making it human.

Useful friction has a specific purpose. It might require someone to:

• verify a consequential claim against an original source;
• test AI-generated code before deployment;
• explain why a recommendation should be accepted;
• produce an independent judgment before seeing the machine’s answer;
• challenge an AI output when evidence is weak;
• review edge cases that automated systems handle poorly;
• periodically perform a task without AI to maintain underlying competence.

These practices turn human involvement from a ceremonial “human in the loop” into an actual control mechanism.

That distinction matters.

A human who simply clicks “approve” after reading a confident AI-generated recommendation is not necessarily exercising meaningful oversight.

A human who understands the task, knows the system’s failure modes and has enough time and authority to reject the output is.

The Emerging Divide May Be Between Automation and Agency

The next stage of AI adoption may therefore create an unusual reversal.

For decades, technology was often valuable because it reduced the amount of human effort required to accomplish something.

With generative AI, that principle still holds but only up to a point.

Once production becomes extremely cheap, human attention becomes the scarce resource.

The scarce capability is no longer necessarily making something.

It is deciding what should be made, whether it is correct, whether it matters, what was overlooked and who should be accountable for the result.

Microsoft’s 2026 research describes advanced AI users as people who deliberately decide what should be handled by AI and what should remain human work, including intentionally doing some work without AI to keep skills sharp.

That is a more useful definition of AI fluency than knowing how to write increasingly elaborate prompts.

AI fluency may increasingly mean knowing when not to use AI.

Conclusion

The strange thing about an AI-accelerated economy is that human effort does not necessarily become less valuable as machines become more capable.

Some forms of effort become less valuable. Routine production, repetitive transformation and mechanical execution are obvious candidates.

Other forms become more valuable precisely because AI makes them scarce.

Attention. Verification. Taste. Context. Skepticism. Responsibility. Independent thought.

Human friction is valuable when it prevents those capabilities from being silently outsourced.

The objective should not be to preserve every old inconvenience. It should be to preserve the forms of effort that help people understand what they are doing and recognize when something has gone wrong.

The most effective AI workflow may therefore not be the one with the fewest human steps.

It may be the one that removes the right steps and protects the right pauses.