Is AI Making Work Feel Harder? Your Prompting Habits Could Be Part of the Explanation

The findings connect more structured prompting habits with lower reported mental strain and greater confidence in handling AI responses.

When using AI tools to generate answers, verifying the accuracy and appropriateness of the terminology is crucial. A translation may read smoothly but still miss the precise meaning that a specialist needs, leaving someone to untangle the mistake and rewrite the instructions. For people working with technical language, that cycle of asking, checking, and correcting can become mentally tiring.

A study published in Frontiers in Psychology examines this pressure in "Generative artificial intelligence and terminology work: structural associations of prompt strategy preference, cognitive load, and self-efficacy." Researchers Kangxi Li, Liye Li, and Jingwen Lu surveyed 1,760 people who used large language models for terminology tasks. The findings connect more structured prompting habits with lower reported mental strain and greater confidence in handling AI responses.

The research highlights the person doing the checking, whose workload can disappear behind the speed of an AI-generated answer. Participants who reported using prompting strategies more extensively tended to feel more capable of directing the technology and less burdened by the work. The survey captures relationships between these experiences, without establishing that better prompts caused the differences.

The work behind getting the right words

Terminology work involves identifying specialised concepts, defining them accurately, and keeping their meaning consistent across languages and professional settings. Participants included terminologists, translators, localisation specialists, and researchers in language-related fields, all of whom had used an AI language model for relevant tasks during the previous three months. Recruitment took place through online surveys and professional networks, with 1,760 responses retained from 1,920 submissions after quality checks.

The questionnaire covered seven prompting approaches: assigning a role, supplying examples, requesting reasoning steps, providing reference texts, giving feedback, setting boundaries, and organising instructions clearly. These approaches describe how someone directs a model, such as specifying an approved glossary or showing examples of acceptable translations. Researchers measured participants' reported preference for these strategies rather than observing their actual prompts during a work session.

Participants rated their experiences on seven-point scales, including confidence in correcting AI responses by changing instructions. An adapted NASA Task Load Index assessed mental demands, time pressure, effort, frustration, physical demands, and perceived performance. The team used statistical modelling to examine relationships between these measures and identify groups with similar prompting habits.

Three prompting patterns and a difference in strain

The analysis identified three groups: 470 participants with low strategy use, 875 with medium strategy use, and 415 with high strategy use. These represented roughly 27%, 50%, and 24% of the sample, making moderate use the most common pattern. People in the high-use group reported greater use across all seven approaches, rather than favouring just one particular technique.

The high-use group reported the lowest mental workload and the greatest confidence in managing AI interactions. People in the low-use group reported the highest workload and the lowest confidence, with the middle group falling between them. Across the full sample, the relationship between prompting preference and lower workload was modest, so the findings do not suggest that structured instructions remove most of the strain.

More demanding tasks were associated with greater mental workload, and better perceived output quality was associated with lower workload. The central relationships remained after the researchers accounted for age, education, task complexity, and concerns about trusting the system. A separate check that removed physical demands and perceived performance from the workload measure produced a similar overall pattern.

Confidence matters across career stages

Confidence in handling AI, described in the paper as "AI interaction self-efficacy," means believing that you can guide, revise, and use a model's responses effectively. Greater use of prompting strategies was associated with stronger confidence, which was itself associated with lower mental strain. Confidence accounted for part of the relationship between prompting and workload in the researchers' model, leaving a separate direct association between the two.

The authors interpret this pattern through cognitive offloading, the practice of moving some mental work into external tools or instructions. Writing down terminology rules and boundaries could reduce the need to keep every requirement in mind during repeated checks, and feeling capable of correcting a response could make uncertainty less frustrating. These are plausible explanations for the survey findings, rather than measured proof that participants saved time or completed fewer correction rounds.

Professional seniority did not clearly change the overall pattern linking prompting, confidence, and workload. Comparisons among 837 novices, 490 intermediate participants, and 433 experts did not support the expectation that experienced workers would show a stronger relationship between prompting and reduced strain. The direct prompting–workload link was less conclusive within the expert group, but formal comparisons did not establish a difference between career groups.

What workplaces can learn from the findings

The authors suggest that AI training should address workers' confidence and mental workload alongside the quality of generated text. Their proposals include structured prompt templates, terminology constraints built into interfaces, and tools that help people divide complicated tasks into manageable parts. These ideas could give employees clearer ways to direct a model and recover from errors, but the study did not test a training programme or interface redesign.

The findings come from a survey taken at one point in time, leaving the direction of the relationships unresolved. People who already feel less overloaded may have more energy to write detailed prompts, and confidence could influence both their prompting habits and their workload ratings. Self-reported answers can be affected by memory, and an overall workload score cannot capture every spike in frustration during a difficult exchange.

The sample was highly educated, with about 95% holding at least a master's degree, which limits how confidently the findings can be extended to other groups. The researchers propose follow-up studies using training interventions, observations over time, interaction records, eye tracking, or brain-activity measurements to examine what changes during actual work. Their study offers a useful starting point for understanding AI-assisted work: the quality of the experience includes how much effort people spend managing the tool and how capable they feel when it gets something wrong.