When the Human Becomes the Interface

A new kind of skill is emerging that does not look much like a skill at first. A person asks an AI system for a memo, a summary, a comparison, a draft or an image. The first result is close but wrong in some important way. So the person tries again. They specify the audience. They add examples. They prohibit phrases the system keeps using. They define the format. They break the job into steps. They tell the system what not to infer. They learn which instructions must be repeated and which ones can be implied.

By the fifth or tenth attempt, the user is no longer simply asking software to do something. The user has begun building a private operating manual for the software.

That behavior is becoming measurable. A study published April 30 in Information Systems Research examined what researchers call “prompt adaptation”: the way people change their instructions as models change. Across two preregistered experiments involving 3,750 participants and nearly 37,000 prompts, researchers found that in a bounded task with a clear target, roughly half of the performance gains associated with a model upgrade came not from the upgraded model alone, but from people changing how they prompted it.

The finding complicates one of the simplest stories about artificial intelligence. Better model in, better result out. In practice, at least for some kinds of work, a better model may also require a better-adjusted human.

THE NEW MACHINE DIALECT

Software has always trained people to behave in certain ways. Search engines taught users to reduce questions to keywords. Spreadsheets taught workers to think in cells, formulas and fields. Automated phone systems taught callers to wait for a prompt and choose from a menu that never quite contained the reason they called.

Generative AI is different because the interface looks like ordinary language. There is no obvious command syntax. The box invites a sentence. That makes the adaptation easy to miss. A user may believe they are simply “talking to the AI” while quietly learning a new dialect: state the goal, specify the audience, define the tone, provide context, offer examples, constrain the output, inspect the answer and revise the request.

The U.S. Department of Labor has effectively formalized that behavior as part of AI literacy. Its 2026 Artificial Intelligence Literacy Framework says workers should learn to provide context, structure prompts clearly, specify formats, supply relevant materials, iterate on results and recognize how word choice can affect outputs. The framework does not treat prompting as a clever trick for power users. It treats the ability to direct AI effectively as a baseline workforce capability.

That is a significant shift in where the work sits. The software may generate the paragraph, code block or recommendation, but the human increasingly performs a surrounding layer of translation: turning an intention into instructions the system can reliably use.

Research suggests the translation can also change the way people communicate. A 2025 study in the Journal of Intelligent Information Systems analyzed more than 200,000 real-world conversations with large language models. The researcher found that users often began with more machine-like instructions, then shifted after the first exchange toward more natural language, greater politeness and shorter, more conversational prompts. The interaction did not merely produce text. It appeared to influence the style of the person using it.

That does not mean everyone is becoming emotionally attached to a chatbot. It means the boundary between “using software” and “accommodating a conversational partner” is unusually blurry. People do what people have always done in communication: they notice what gets a response and adjust.

WHEN ADAPTATION BECOMES WORK

There is a productive version of this. A photographer learns which descriptive details help an image model understand composition. A programmer learns how much surrounding code an assistant needs before a suggestion becomes useful. A researcher learns to ask for claims and sources separately. A manager learns that a vague request for “a presentation” is less useful than specifying the audience, decision and evidence the presentation must support.

That is ordinary tool mastery. New tools have always required new technique.

The stranger question is what happens when the technique becomes an accumulating list of defensive behaviors. Users may begin preemptively adding instructions because they have learned the system tends to over-explain, omit a recurring requirement, misread a category, invent a transition, choose the wrong tone or lose track of a constraint. The person is no longer only expressing what they want. They are writing around the machine’s known habits.

At that point, the interaction develops what might be called an adaptation tax: the extra cognitive work required to make a general-purpose system behave like the tool the user thought they were opening.

A 2026 study in Computers in Human Behavior examined that burden directly. Researchers surveyed 1,045 generative-AI users and found that uncertainty about how to phrase prompts was associated with repeated prompt recrafting. More recrafting was linked to greater emotional fatigue and lower trust in the system. Notably, higher AI literacy did not eliminate the fatigue effect.

That matters because “just prompt it better” can sound like a complete answer to any bad AI interaction. Sometimes it is useful advice. It can also move responsibility for a system’s unpredictability entirely onto the person using it.

If a microwave required customers to develop a personal theory of microwave psychology before reheating soup, nobody would call that user sophistication. Conversational AI gets more latitude because its range is enormous and its failures are often soft rather than mechanical. The system works. It simply works differently depending on exactly how the person asks.

THE JOB AROUND THE JOB

Workplace adoption makes the issue larger than individual prompting style. Microsoft researchers studying knowledge workers have found that generative AI can shift critical-thinking effort away from gathering information and toward verification, integration and what the researchers call task stewardship. The worker may spend less time producing the first version of something and more time judging whether the machine’s version is accurate, appropriate and usable.

The Conference Board reported in July that 55% of nearly 1,300 surveyed workers were using generative AI or AI agents daily or weekly, while only about one-third had participated in employer-provided AI training during the previous six months. Nearly 28% said their organization provided no AI training at all. The organization’s own recommendation was not simply more tutorials, but time, tools, hands-on practice and support for workers to develop new capabilities as the technology changes.

This is where adaptation becomes organizational labor. A company may license an AI system in an afternoon. The actual operating knowledge appears later: which tasks it handles well, where it needs supervision, what context makes it reliable, how outputs must be checked, which prompts survive a model update and which ones suddenly stop working.

Some of that knowledge becomes formal training. Some becomes prompt libraries and internal guidance. Much of it remains informal: one employee tells another, “Ask it this way.” A team quietly adds an extra review step. Someone creates a template because the same correction keeps happening. A worker learns that a particular phrase sends the system in the wrong direction and simply stops using it.

That is adaptation, too. It is also process design, whether anyone names it or not.

The International Labour Organization and partner institutions argued in an August report that AI adoption is increasing demand not only for technical skills but for higher-order cognitive and socioemotional skills, while making AI literacy, adaptability, resilience and human agency increasingly important. The useful word there may be “adaptability.” It sounds positive, and often is. But adaptability is still labor. Someone has to do it.

WHO IS SUPPOSED TO CHANGE?

The future of useful AI probably depends on both sides moving. People will become better at knowing when to use these systems, how to frame tasks and how to evaluate the output. Models and interfaces should also become better at absorbing correction, preserving preferences, signaling uncertainty, asking for missing information and reducing the number of private workarounds users have to remember.

The Information Systems Research study offers a useful warning on that point. The researchers tested automated prompt rewriting and found that it was not a universal substitute for human adaptation. Depending on the task, automated rewriting could modestly improve performance or actively undermine the gains from a better model. The shortcut itself had to understand the user’s goal.

That is the larger design challenge. Adaptation can be a sign of mastery, but it can also be a sign that the human is compensating for the interface. The difference is whether the adjustment gives the person more capability or merely makes them carry more of the system in their head.

For decades, good interface design has tried to make computers easier for people to use. Generative AI introduces a peculiar reversal. The software can speak in ordinary language, yet the user may still end up learning the machine’s habits, tolerances and failure modes in order to get ordinary work done.

The most important question may not be whether people can learn to work with AI. They clearly can. The question is how much invisible behavioral change should be considered a reasonable price of admission.

Because once users begin rewriting not just the prompt, but their own communication habits around the system, the interface is no longer only on the screen.

The human has become part of it.

SOURCE NOTES

• Jahani, Eaman, et al. “Prompt Adaptation as a Dynamic Complement in Generative AI Systems.” Information Systems Research, published online April 30, 2026.
• U.S. Department of Labor. “Artificial Intelligence Literacy Framework.” Employment and Training Administration, February 13, 2026.
• Schneider, Johannes. “Mental model shifts in human-LLM interactions.” Journal of Intelligent Information Systems, June 26, 2025.
• Yang, Hui; Wang, Jinqiang; Hu, Peng. “Trust erodes, fatigue builds: How prompt uncertainty traps users in recrafting loops.” Computers in Human Behavior, 2026.
• Lee, Hao-Ping (Hank), et al. “The Impact of Generative AI on Critical Thinking.” CHI 2025 / Microsoft Research.
• The Conference Board. “Most Organizations Are Preparing Workers for Today’s AI, Not Tomorrow’s Jobs.” July 28, 2026.
• International Labour Organization et al. “Changing landscape of skills in the age of AI.” August 13, 2026.