Performative AI use is when employees overstate or stage their use of AI to look compliant or capable, without the use improving the work. A Visier survey of 1,000 US employees, published September 23 2026, found that 48% admit exaggerating their AI usage or expertise to colleagues or leadership, and 45% feel pressure to use AI without confidence in how. Any adoption metric built on what people report, or on a count they know is being watched, is measuring the incentive as well as the behavior.
Two things are true at once. Adoption is a real signal worth tracking, and it is now one that people have a reason to shape.
What is performative AI use, and how common is it?
The best public evidence is two recent surveys, and both are self-report. They agree on direction, measure different populations, and are not added together here.
What the Visier survey found
Visier's report, "AI Usage at Work: 'Performative AI' Stems from Lack of Confidence in Leaders," draws on 1,000 US employees. Beyond the 45% and 48% above, 59% doubt their employer has a clear AI plan. CIO Dive's coverage adds that more than half of respondents said their roles have significantly changed because of AI in the last two years, and that 7 in 10 worry about career or job-loss effects. Andrea Derler, Visier's principal of research and value, told CIO Dive that if leaders do not create room for employees to be transparent about what does and does not work, they get a blanket answer that AI is helping everywhere.
What the GoTo survey found
GoTo and Workplace Intelligence surveyed 2,500 respondents from November 2025 to January 2026, half full-time knowledge workers and half IT decision-makers, across the US, Canada, the UK, Ireland, Germany, Austria, Switzerland, India, Mexico and Brazil. In the published release, 60% say they feel pressured to use AI to increase productivity, and 47% say AI use is considered in their performance reviews. 42% say they would rather risk errors with AI than look "slow" or "replaceable," and 55% admit they do not always check AI outputs, even for high-stakes tasks. The release attributes these figures to employees, from a sample that is half IT leaders, so read them as directional.
What these surveys cannot tell you
Both are vendor-authored surveys and neither is published as peer-reviewed research. An admission of exaggeration is itself a survey answer, so these figures show the behavior exists, not its rate inside your organization. Visier's page does not state a fielding window. Neither survey can tell you which of your own teams has the problem.
Why does performative AI use break adoption metrics?
Because every common adoption metric has a way to be satisfied without doing better work. This table is our own analysis of the mechanism, not a finding from either survey.
| Adoption signal | What it records | How performative use contaminates it | What it can still support |
|---|---|---|---|
| Self-reported usage (surveys, manager check-ins) | What people say they do | Overstated to look capable or compliant | Sentiment and confidence, not usage |
| Seat or login activity | Whether a licensed person opened the tool | Logging in satisfies it | License right-sizing |
| Prompt or token counts | Volume of consumption | Volume can be generated to be counted | Cost forecasting |
| Manager-observed adoption | A manager's impression | Shaped by what people show managers | Coaching conversations |
| Outcome against a comparable group | Whether the work changed | Hard to stage, because it needs the comparison | Impact claims, with limits stated |
We covered the input-metric problem when a large employer removed AI usage from its review guidance, and the seat-count version in why a seat count is not an adoption metric. The new evidence adds the human side: 47% of GoTo's respondents say AI use is already considered in their performance reviews.
Is self-report ever reliable evidence of AI's effect?
Not on its own, and the cause is not always exaggeration. In a randomized trial by METR, 16 experienced open-source developers completed 246 tasks, each randomly assigned to allow or disallow AI tools. Before starting, developers forecast that AI would cut completion time by 24%. Afterward they estimated it had cut it by 20%. Measured completion time rose by 19%.
Nobody in that study was performing. The developers were sincere and the perception still missed by a wide margin. The study is small, covers experienced developers on familiar codebases using early-2025 tools, and its authors note that experimental artifacts cannot be entirely ruled out. It is a result about one setting, not a general slowdown. What it supports is narrower: perception and measurement can diverge even when people are honest. A related gap between reported and measured gains appears in the Fed research we examined in the AI productivity paradox post.
So there are two separate problems. Some people overstate on purpose, and some misjudge in good faith. A survey sees neither.
How do you measure AI use without encouraging performance?
- Stop scoring AI use in individual ratings. Pressure is the input to performative behavior. Keep usage as context for a conversation, and keep it out of a number that decides a rating.
- Read behavior from the systems where work happens. Cycle time, rework, handoffs and queue time from the tools people already use are harder to stage than a survey answer or a prompt count.
- Report impact as a comparison. Set adopting teams against comparable teams that have not adopted, controlling for tenure, role and tool stack, and attach a confidence level to every result.
- Aggregate before interpreting. Individual-level surveillance is the fastest way to generate more performance. Team-level views should require 5 or more people.
- Make it safe to say what does not work. Visier's principal of research and value, Andrea Derler, told CIO Dive that organizations will not succeed without an environment where employees can be open about what does and does not work. Where reporting a failed use case has a cost, people will report success.
Where Levos fits
Levos is a Human Capital Operating System: an intelligence layer above the existing stack that reads behavioral signals from where work happens. AI Impact is one of its six signal families, and the measurement methodology compares adopting teams with non-adopting teams, with a confidence score on every result and the limits disclosed. By default, individual-level data is visible to the person and their direct manager, and anyone above that sees team aggregates of at least 5 people, apart from a logged escalation path for urgent issues; the privacy model is on our operating page. We do not claim full attribution.
The Visier survey comes from a workforce analytics vendor and the GoTo survey from a communications vendor working with a research agency, and we cite both as published. A buyer still needs the observed layer beside what people say.
Frequently asked questions
What is performative AI use?
Performative AI use is when employees overstate, stage or inflate their use of AI tools so they look compliant, capable or indispensable, without the use being tied to better work. A Visier survey of 1,000 US employees, published September 23 2026, found 48% admitted exaggerating their AI usage or expertise to colleagues or leadership, and 45% felt pressure to use AI without confidence in how. The figures are self-reported and come from a vendor survey, so they show the behavior exists, not how common it is inside any one company.
Why do AI adoption metrics mislead when employees feel pressure to use AI?
Because any adoption number that people know is watched, rewarded or scored measures the incentive as well as the behavior. Self-reported usage can be overstated, and system-observed usage such as logins, prompts or tokens can be inflated by activity that exists to be counted. A GoTo and Workplace Intelligence survey of 2,500 respondents across 10 countries, half knowledge workers and half IT decision-makers, fielded November 2025 to January 2026, found 47% said AI use is considered in their performance reviews and 60% felt pressured to use AI. It is a different population from the Visier sample and the two are not combined here. Adoption remains useful as an input measure, but it cannot carry an impact claim.
Can employee surveys reliably measure AI productivity?
Not on their own. Surveys record what people believe and are willing to report. In a randomized trial by METR with 16 experienced open-source developers and 246 tasks, developers forecast that AI would cut completion time by 24% and later estimated it had cut it by 20%, while measured completion time rose by 19%. That result reflects honest misjudgment on a small sample and early-2025 tools, and it should not be generalized. It shows why perception needs a measured comparison beside it.
How should organizations measure AI use without encouraging performative behavior?
Keep AI use out of individual ratings, read behavior from the systems where work happens rather than from what people say, and report impact as a comparison between adopting and comparable non-adopting teams on outcomes the business already tracks. Aggregate before interpreting, with team-level views only for groups of 5 or more people, and attach a confidence level to every result. Make it safe for people to report what does not work, because where leaders do not allow that, they get a blanket answer that AI is helping everywhere.
Measure what changed, not what was shown
If your AI readout rests on survey answers or usage counts, and finance or the board is asking what it returned, that gap is solvable. Request a Demo and we will walk through what a comparison-based readout contains.
Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.
