This article is written by a simulator provider, and it runs against a simulator provider's short-term commercial interest. We are publishing it anyway, and we explain why at the end.
The industry's founding statistic is false
"Virtual reality achieves 80% retention versus 20% for traditional methods." It also circulates as "we remember 10% of what we read, 20% of what we hear, 90% of what we do", attributed to the learning pyramid or Dale's Cone.
Letrud and Hernes traced the chain of citations and found that it is circular and has no empirical origin. The percentages were invented and added decades after Edgar Dale's original work, and Dale never put numbers on his cone. The NTL Institute, to which the supporting study is attributed, has acknowledged that it cannot produce it. The figures are all multiples of five or ten, something impossible with real empirical data.
What to use instead: the Howard et al. (2021) meta-analysis of 184 studies: d = 0.541. It is a real, moderate, and defensible effect.
"One hour of virtual reality equals ten hours of classroom training"
It also circulates as 1:6, 1:8, and 1:12.
There is no primary study that reports 10x. It is an abusive extrapolation of the 4x that PwC reported in 2020 —30 minutes of virtual reality versus 2 hours of classroom— in a soft-skills course for office supervisors, not in industrial operations. Once equipment setup time is factored in, PwC itself reduces it to 3x.
What to use instead: the Al-Ahmari et al. (2021) figure: 25% less training time in a real industrial case with a control group.
"82% reduction in accidents with virtual reality training"
It appears attributed to an unnamed "leading manufacturer". It is a vendor's commercial case study, published without a company name, without methodology, without a control group, and without a measurement period. It cannot be verified in any way.
What to use instead: the Tyson Foods case, a reduction of more than 20% in injuries over one fiscal year, reported by the company itself in the trade press, which stated that virtual reality was one of several initiatives.
"Digital twins improve OEE by 25% to 28%"
Also seen as "35% less downtime with digital twins". There is no primary source. Every variant comes from vendor blogs, LinkedIn articles, or non-indexed magazines. None publishes the plant, measurement period, baseline, or method for calculating OEE. It is a circular citation chain among companies that sell the product.
What to use instead: the official U.S. Department of Energy guide: a 35% to 45% reduction in downtime and 70% to 75% elimination of breakdowns in a mature predictive maintenance program.
"Augmented reality increases first-time fix rate by 20% to 30%"
Widely used by remote-assistance vendors. No primary source was found. Every appearance comes from vendors in the industry, with no sample, no baseline, and no operational definition of the metric.
What to use instead: the Boeing / Iowa State University study: roughly a 90% improvement in first-time quality, with a university experimental design.
"US$2 million per hour of downtime"
Also as "US$129 million a year per plant". It is not false, but its basis is 56 self-reported online interviews —or 181 accumulated over five years in the most recent version— plus an extrapolation to the Fortune Global 500. It is an order of magnitude, not a measurement. Citing it without disclosing the sample is a risk in front of any reliability manager who knows the report.
What to use instead: the NIST study: 45.7% reactive maintenance and US$119.1 billion in preventable losses, with 71 establishments, stratification, and a stated margin of error.
And four more
"219% ROI over three years with virtual reality" — this is a Forrester study commissioned and paid for by Meta, the company that sells the headsets. Studies in that format are not independent research.
"Only 10% of training transfers to the job" — an estimate of disputed origin, known in the literature as the 10% delusion. It is repeated so often that it looks like a measurement, but no one has been able to reproduce the original figure.
"Walmart cut training time by 96%, from 8 hours to 15 minutes" — it appears only on the vendor's website and contradicts what Walmart itself published in its newsroom, which refers to a 10% to 15% improvement in assessment scores.
"94% of employees would stay longer at a company that invests in their learning" — a survey by a vendor that sells courses, about stated intention rather than actual behavior. Stated intention systematically overestimates behavior.
The study that makes us most uncomfortable, and that we publish anyway
A University of Colorado study published in the ASCE Journal of Construction Engineering and Management evaluated virtual reality with haptic feedback among construction workers and found no improvement at all in hazard recognition, risk perception, or risk tolerance.
It contrasts with most of the literature, and for that very reason it is the most instructive study in the set. It confirms what the 184-study meta-analysis already showed as a moderator: when the fit between the task and the technology is low, the effect collapses to d = 0.207, which is practically zero.
The lesson is uncomfortable for anyone selling this: poorly designed virtual reality does not work. It doesn't work a little less: it doesn't work.
Why we are publishing this
For years, several of these figures were in our own materials, as they were in almost everyone else's in the industry. We took them out.
Not out of political correctness, but by calculation: the client who buys these systems is an engineer, checks what they are shown, and the day they discover an invented figure, they stop believing the true ones as well.
We would rather show up with a smaller, sustainable number. If, when comparing proposals, you see an 80% retention rate or an 82% reduction in accidents with no source, you know exactly what to ask.
Sources: Letrud & Hernes — Cogent Education; Thalheimer — Work-Learning Research, 2018 · Howard, Gutworth & Jacobs — Computers in Human Behavior, 2021 · Hans, Mortimer & Horan — Virtual Reality (Springer), 2023 · University of Colorado — ASCE Journal of Construction Engineering and Management, 2025