AI Use Can Make Tasks Take Longer, Research Finds

New research suggests AI can make simple tasks take longer, while convincing users they are becoming more productive.

A new study from Stanford, NYU and Princeton has found that we frequently use AI even when it’s inefficient; and that for the smaller tasks that we compulsively farm out to AI, we would often expend less mental effort, and save more time, by doing the task ourselves.

Across three human studies commissioned for the research, the authors found that participants routinely mis-estimated how much time AI would save them on a proposed task, as well as notably underestimating how much they depend on and actually use AI*.

‘In [the second study], we seek to understand why people might use AI for simple tasks despite AI use not providing efficiency benefits. One hypothesis is that people are miscalibrated about how much time and effort AI assistance saves.

‘To test this hypothesis, we compared people’s predicted versus actual time and effort completing these tasks with and without AI assistance and identified efficiency-gain illusions, where people overestimated both the time and effort that AI saves.

‘On average, people predicted AI assistance to save time by 55.7 seconds when it only saved 7.5 seconds. This miscalibration is particularly severe on the simple variants of the tasks, where people predicted AI assistance to save time, but it made tasks slower to complete in reality.’

The new paper, titled The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks, co-written by seven authors across the three institutions, finds also that prior AI use appears to reinforce future AI use, even when the technology offers little or no real efficiency benefit.

Overview of the three experiments used to test how people use AI for simple everyday tasks, revealing that users underestimate how often they rely on AI, overestimate the time it saves, and become more likely to use it again after prior exposure. Source

Data from the three studies found that people become more likely to use AI after prior exposure, even in ways that are not productive or time-saving, or even stress-saving (i.e., expending less mental effort on the task)*:

‘Contrary to the possibility that experience improves calibration [i.e., ability to estimate how useful AI is for a task], we identified a session-level carryover effect where initial AI use increases subsequent AI use.

‘Participants who initially completed tasks with AI became even more likely to opt for AI assistance on easy task variants, even though doing so did not offer time or effort savings on average.’

In one of the human studies, the authors found that savings obtained from AI use were entirely illusory:

‘AI assistance can [backfire.] [We] found that people who chose to use AI spent 7.06 seconds more than people who completed the tasks [independently] and reported higher effort.’

The study was limited to tasks that took five minutes or less, but may resonate with former search engine addicts who now routinely resort to ChatGPT and other popular, commoditized LLMs instead.

Study Groups

Across the various user studies, tasks were devised based on the Taxonomy of User Needs and Actions (TUNA) framework. The experiments covered information-seeking; summarization; arithmetic; spelling correction; rewriting; and other low-complexity tasks that could generally be completed in under five minutes.

The first study compared participants’ predicted willingness to use AI against their actual behavior during task completion, investigating whether people accurately understand their own reliance on AI assistance.

The second focused on perceived-versus-actual efficiency gains, by comparing participants’ expectations about time savings and reduced mental effort against measured completion times and reported workload, during AI-assisted and independent task completion.

The third examined whether prior exposure to AI changed later decision-making, tracking whether participants who had previously completed tasks with AI became more likely to rely on AI again during subsequent tasks.

Overthinking It – AI Use on Simple Tasks

To understand whether people accurately estimate their own reliance on AI, study participants in one session were asked to complete four tasks, with an option to use AI assistance for each task. The level at which the participants actually did use AI was compared to their own prior estimation of how much they thought they would use it, with significant dissonance evident in the results:

Participants consistently underestimated how often they would turn to AI for simple tasks, with the gap widening on easier prompts where actual AI use rose to 38% against a predicted rate of 20%, suggesting that habitual delegation to AI extends well beyond users’ own awareness.

The authors state:

‘We find that [people] indeed used AI significantly more than the average predicted rate. On average, participants reported that they would use AI in 33% of tasks, but the population-level rate of AI use is 47% (β = 1.07, p < 0.001).

‘This gap is larger for easy task variants (β = 0.69, p < 0.001): participants predicted a 20% AI use but the actual rate of AI use was 38% (β = 1.42, p < 0.001), nearly doubling the stated preference rate.’

The experiments focused on ordinary low-effort tasks that many people now routinely hand over to AI, even when doing so may be unnecessary. Participants were asked to perform simple activities involving factual recall, arithmetic, spelling correction, rewriting short passages, summarizing text, and answering basic reasoning questions, with some tasks requiring only a few words or a single sentence to complete.

The study also included slightly harder versions of the same activities, allowing the researchers to compare whether AI usage changed as the work became more demanding.

AI Time-Saving Benefits Overestimated

In the second study, participants were divided into two separate groups, with one group first estimating how much time and mental effort AI would save on a series of tasks, while another group actually completed those same tasks either independently or with AI assistance. The tasks again centered on low-complexity activities involving arithmetic, rewriting, factual recall, summarization, spelling correction, and short reasoning exercises.

The goal was to compare people’s expectations about AI productivity against what really happened when the work was performed. According to the paper, participants consistently overestimated how much AI would help them, particularly on easier tasks where many assumed AI would dramatically reduce workload and completion time.

Instead, measured results often showed only minor gains, and in some cases, as mentioned earlier, AI use actually slowed participants down. The paper reports that people expected AI assistance to save nearly a minute on average, while the observed time savings were only a few seconds.

On some simpler tasks, AI users did indeed take longer to finish than people who completed the work independently:

Predicted versus actual time and mental effort during AI-assisted and independent task completion, revealing the paper’s proposed ‘speedup illusion’, in which participants consistently believed AI would save far more time than it actually did. Actual AI-assisted completion times were substantially longer than predicted, while estimates for independent task completion remained far closer to the observed results.

The study also examined perceived mental effort. Participants commonly believed AI would make tasks feel substantially easier; yet, the measured reduction in cognitive effort was far smaller than expected. The paper characterizes this as an ‘efficiency-gain illusion’, in which people systematically overestimate both the speed and usefulness of AI assistance during simple everyday work.

AI Usage Deepens Delusion

The last of the three studies was designed to test whether even brief exposure to AI changes later decision-making. Participants were divided into multiple groups and first passed through an ‘exposure phase’, where some completed easy tasks with AI assistance; some completed harder tasks with AI assistance; and others completed the same categories of tasks independently, without AI. A separate control group skipped the task stage entirely.

Subsequently, all groups entered a second ‘test phase’, this time given new and easier tasks, and allowed to decide for themselves whether or not to use AI. The tasks again focused on low-complexity (i.e., rewriting, arithmetic, recall, spelling correction, summarization, and short reasoning exercises) tasks that could each be completed in only a few minutes.

The paper reports that participants who had already used AI during the exposure phase became substantially more likely to rely on it again afterward, even when earlier AI use had failed to save time or reduce mental effort.

The researchers found that prior AI users selected AI assistance far more frequently during the later test stage than participants who had previously completed tasks independently:

Participants who had previously used AI during the exposure phase became substantially more likely to rely on it again during later tasks, despite earlier AI use often failing to produce meaningful gains in speed or reduced mental effort. The left panel shows that prior AI users selected AI assistance far more frequently during the later test phase than participants who had initially completed tasks independently. The right panel illustrates the paper’s proposed ‘speedup illusion’, in which prior AI exposure increased participants’ belief that AI-assisted work was faster and more efficient, even though measured completion times frequently showed little benefit and sometimes slower performance. 

Repeated AI exposure is reported to have distorted participants’ judgement about whether AI was genuinely useful: people who had already used AI became less likely to agree that the tasks could actually be completed faster without it, despite measured results frequently showing little benefit and, in some cases, slower completion times.

The researchers argue that this creates the conditions for a ‘self-reinforcing cycle’, in which AI use increases future dependence on AI, while simultaneously weakening users’ ability to accurately judge whether the technology is improving productivity at all.

Conclusion

Opinion Many readers who have adopted AI for small tasks may, like me, feel a sense of familiarity with the new paper’s conclusions.

Personally, my obsession to automate repetitive tasks precedes the current AI boom by several decades. Then, as now, the question remains: Does the effort entailed in setting up and/or maintaining the automation exceed the estimated (human-only) effort in just doing the task, without any automation?

Those who love to automate may end up automating for the sake of it, even if it were to take years or decades before any return (in terms of time saved) would be evident; and this changes the context of the activity from ‘optimization’ to ‘hobbyist’.

There’s nothing wrong with this, so long as you are not deluding yourself that real gains are being made. Nonetheless, this is a bad habit that I have tried to resist in recent years; and the option to use AI, of late, seems prone to exacerbate it, since even bad or non-optimal results can be obtained much more quickly than, for instance, when scripting macros in JavaScript and other languages.

Deceptive Indicators

What the paper neglects a little is the tension between serendipitous or fortunate results via AI vs. the predominance of blind alleys and frustrated attempts to conform available AI chatbots, such as ChatGPT, to one’s own needs – in a workflow that can be depended on, even in the face of forced upgrades to new versions that may not operate in the same way as the version on which your workflow or routine was conditioned.

A ‘magic’ result, therefore, is that occasion where AI immediately solves your problem in an easy and rational manner.

For instance, every time I review a paper, I must print it, and inevitably have to write page numbers in bold up top, since these are usually either absent or in tiny type at the bottom. Asking ChatGPT to produce a Python script that would add a large and bold page number at the top proved to yield an incredibly quick result, in that I can now drag an Arxiv PDF onto a .BAT file and have a new version with prominent page numbers, in 2-3 seconds:

Evident page numbering added to PDFs via an AI-written Python script.

Aside from a minute or two of argument as to whether Windows has a native and separate Arial Black font (it doesn’t, anymore), this was possibly the quickest that AI had ever created something persistently and regularly useful for me.

Arguably, this kind of ‘breakthrough’ or ‘easy  win’ gives a false impression of AI’s true ability to save time and/or mental expenditure, because we tend to award such instances undue emphasis: our natural tendency to suppress painful or negative memories and to re-use or center on happier memories means that successful instances where AI solves small tasks in a useful way will end up as a lodestar we are likely to pursue, even against the trend of statistical evidence, as demonstrated in the new paper, and even against our own experience that such ‘easy wins’ are the exception rather than the rule.

There is growing proof, besides the new paper, that we delude ourselves about AI’s utility. In 2025 a study showed that developers using AI were taking 19% longer than without AI; and another more recent offering confirms the underlying message of the new paper discussed in this article – that saving time takes time.

It would be useful if research of this kind could be translated into a (inevitably AI-powered) old-school time and motion study, allowing us real insights into the extent to which AI actually saves us – or costs us – time.

Finally, the study is exceptional in another important respect, in that it at least attempts to quantify ‘mental expenditure’ in regard to the use of AI on lesser tasks. As increasing attention is centering on the ‘intensity’ of AI-aided work, we urgently need reliable units of measurement that can quantify the extent to which the exigencies and eccentricities of AI exhaust or deplete us, at a more general cost of quality of work and fitness for work.

* Formatting is the authors’, from the source paper. Any inline citations converted to hyperlinks by me.

First published Saturday, May 23, 2026