Trained skills fade on a curve that has been measured, and the measurement is less comforting than most AI enablement sales decks suggest. The largest review of skill retention to date, published in Psychological Bulletin in 2025, puts the half-life of accuracy-based performance gains at roughly six and a half months with no practice in between. AI fluency is an accuracy-based skill. Buy a session in January, do nothing after, and by late July you own about half of what you paid for.
That is the short answer. The longer one is that two problems share the name, the research answers only one well, and neither shows up on your dashboard.
Two problems, one name
Search “AI skills decay” and almost everything you get back is about the first problem: human judgment weakening because people use AI. Call it deskilling. It has a serious literature, an aviation precedent, a BCG survey and a clinical study of endoscopists losing unaided detection skill behind it.
The second problem is the one a COO actually has. You paid for an AI enablement session. People came out able to do things they could not do before. Six months later the team is back to pasting a one-line prompt into a chat window and shipping whatever comes out. Nothing eroded anyone’s judgment. The skill went unpracticed and left, the way any trained skill does.
Keep them apart, because the cures differ. Deskilling is fixed by changing how AI gets used: review rules, work design, keeping humans on the hook. Decay is fixed by practice on a schedule. Confuse them and you buy a governance policy for a retention problem.
They belong in one article because the second causes the first. A team whose fluency has faded does not stop using AI. It keeps using AI badly, accepting output without checking it, which is the exact behavior the deskilling studies measure.
What the decay research measures
Skill decay was measured long before anyone needed the word for AI, which means there are numbers instead of opinions.
The 1998 review by Winfred Arthur and colleagues in Human Performance pulled 189 data points from 53 studies and found what you would expect in shape, not in size. Right after training, loss is nil, at an effect size of d = −0.01. After more than 365 days of not using the skill, it reaches d = −1.4. A year of no practice leaves the average trained person well below where the training left them.
The 2025 update is the one worth keeping on hand. A review of procedural skill retention and decay in Psychological Bulletin gathered 1,344 effect sizes from 457 reports and was the first to treat the gap since training as a sliding scale instead of buckets. Prompting a model and checking its answer are procedures, which puts AI fluency inside its scope. Its rates:
| Measure of performance | Decay rate | Half of the gain is gone at |
|---|---|---|
| Accuracy-based | 0.08 per month | about 6.5 months |
| Mixed | 0.06 per month | about 11 months |
| Speed-based | 0.06 per month | about 13 months |
The authors are blunt about what the figures are for: they tell you how much loss to expect when a skill goes unused, and therefore when refresher training belongs on the calendar. That is a planning instrument, and almost nobody in AI enablement is using it.
Two numbers in wide circulation are worth retiring before they reach your board pack. “Skill shelf life has dropped from years to months” traces to HR-platform blog posts with no study attached. The Ebbinghaus figures, 70% lost in a day, come from one man memorizing nonsense syllables in the 1880s and were never about workplace skills. The meta-analyses point to a slower, more expensive truth: the loss is not immediate, which is why it is so easy to miss.
Why AI fluency sits in the fastest-fading category
Both reviews name the same moderators, and AI fluency scores badly on every one.
Arthur’s review found that physical, natural and speed-based tasks hold up, while cognitive, artificial and accuracy-based tasks fall off fastest. Riding a bike is physical and natural. Writing a usable brief for a model, spotting where it has invented a citation, deciding which half of a draft to keep: that is cognitive, learned rather than instinctive, and judged on whether the answer is right rather than how fast it arrives.
| Attribute | Holds up | Fades fast | Where AI fluency lands |
|---|---|---|---|
| Nature of the task | Physical | Cognitive | Cognitive |
| Origin of the skill | Natural | Artificial | Artificial |
| What success means | Speed | Accuracy | Accuracy |
| Practice between uses | Frequent | Intermittent or none | Usually none on the hard parts |
The last row is the one you control. The 2025 review lists intermittent performance opportunities among its moderators, a careful way of saying that survival depends on whether the person gets to use the skill. Here is the catch specific to this subject. A team does keep using AI daily, so the easy part, the part that needs no skill, gets practiced constantly. The hard parts, structuring a difficult task, checking an answer you cannot verify at a glance, rejecting a confident wrong one, come up rarely and are what was taught. Daily tool use gives them no reinforcement at all, while producing a log file that says everything is fine.
Most retention advice treats AI fluency as one skill with one curve. It is at least two, fading at different rates, and your usage report cannot tell them apart. The companion piece on what survives six months after a session covers which habits hold.
Checking fades before doing does
If one finding deserves to change how you measure, it is this one. The skill that goes first is verification.
A three-wave longitudinal pilot tracking problem-solving over six months, posted as a preprint in January 2026, found daily AI use climbing from 52.4% to 95.7% while something else moved the other way. Reliance on AI was heaviest for the hardest tasks, at 73.9%, exactly where confidence in one’s own verification sat at 68.1% and accuracy was worst at 47.8%. The gap between believed and actual performance widened to 34.6 points. It is a small academic pilot and a preprint, so treat the figures as directional. The shape matches the larger survey evidence: Microsoft Research and Carnegie Mellon’s CHI 2025 study of 319 knowledge workers across 936 real use cases found that the more confidence people placed in the tool, the less critical thinking they reported doing.
Picture that on a Tuesday. Output volume is holding. Turnaround times are good. Nobody has complained. And the share of work going out with nobody having read it against the source climbs month by month until one piece of it reaches a client.
So a decay problem shows up first as a quality problem with no obvious cause. BCG’s June 2026 survey of 70 C-suite and senior executives found almost 90% naming overreliance on AI output without stress-testing it, and half saying they already see capability loss in their own organization. Seventy executives is opinion rather than measurement. The useful figure is the last: only one in ten of those companies had an organization-wide response, and about a third had never discussed it.
Why the admin console cannot see any of this
Decay is latent. It stays hidden for as long as nothing changes, then surfaces all at once when something does.
Researchers at Aalto University studying skill erosion describe exactly this in an accounting firm that leaned on an automation system. The erosion stayed invisible while the system ran, then appeared the moment management took it away and found staff could no longer do the core work by hand. Their list of triggers is short: a tool withdrawn, a model updated, the environment shifting under the workflow.
For anyone measuring AI adoption from a seat-license report, that is the whole problem. Every instrument most buyers have, license counts, weekly active users, prompt volume, measures whether the tool is open. None measures whether the person at the keyboard can still do the hard thing the session taught. Usage can rise while capability falls, and on the evidence above that is the normal case. The measurement work that proves capability moved sets out the layers that do catch it, and the case against leaving a program unmeasured prices what the missing baseline costs later.
The decay calendar
Read this as a schedule of bills. The dates come from the rates above; the decisions land whether or not anybody measured.
| Month | What has likely gone | The decision it forces |
|---|---|---|
| 1 | Nothing yet. Enthusiasm is doing the work | Whether to fund reinforcement while it is still cheap |
| 2–3 | Verification habits on hard tasks | Whether a quality wobble is a training gap or a scope gap |
| 4–5 | Task structuring. People fall back to one-line prompts | Who needs a refresher and who needs a different role |
| 6–7 | About half of accuracy-based gains, per the published rate | Renewal, with or without evidence |
| 9–12 | The workflows taught, if nobody wrote them down | Whether to buy a second program because the first “did not work” |
| 12+ | On the 1998 figures, most of the measurable gain | Whether AI enablement gets a budget line next year |
Month six is the expensive one. A renewal talk with no capability measurement is a negotiation where the vendor knows more than you do about what their session achieved. Month nine quietly kills programs: a gap appears, nobody can say whether the session covered it, and the fastest story in the room is that the training failed.
Reinforcement sized to the curve
Reinforcement is not a culture program. It is scheduled practice on the parts of the skill daily work does not exercise, and the published rates tell you both how often and on what.
Three to four touches in the first six months, not twelve. The decay is gradual, so the cadence should be too. Weekly nudges burn through attention long before the curve bends: the fatigue curve of internal AI messaging bends well before month six, and a reminder nobody reads is not reinforcement.
Practice the hard half. A touch that reminds people the tool exists reinforces the skill that was never at risk. One that counts puts people in front of a task they cannot do with a one-line prompt and has someone look at the result. Thirty minutes of that beats an hour of recap.
Make verification the exercise, not the lecture. Since checking fades first, the cheapest good drill is handing a team an answer with a plausible error in it and asking them to find it. Twenty minutes, and failure in the room is free.
Give it an owner with hours. Reinforcement that belongs to everyone belongs to nobody. The mechanics of a champion network are the usual way to fund a distributed version, provided the hours are real and not goodwill.
Measure before each touch. A short timed task on a named deliverable, run before touch two and before touch four, turns the curve from a belief into a number you can show.
What changes at the point of purchase
Decay is predictable, which makes it a contract term rather than a surprise.
Workshops in this market run roughly $2,000 to $15,000 depending on length and group size. The question to put to any provider is not whether the session is good. It is what happens in months two, four and six, who runs it, and whether it is in the quote. A provider selling one day with nothing after is selling you the top of the curve and leaving you the slope. Often that is what the buyer asked for. It is still an incomplete purchase.
Three things to specify before signing:
- A dated record of what was taught. Without it you cannot tell a training gap from a scope gap at month nine, and you will guess wrong.
- A before-reading on one named deliverable, taken the week before the session. It is the only input you cannot buy back later.
- Named reinforcement touches with owners and hours. Three in six months, with the second booked before the first has happened.
None of this flattens the curve. It makes the curve visible, which is enough to stop you paying twice for the same capability. The broader case against treating license count as capability makes the same argument at the level of strategy: what you bought and what your team can do are different things, and only one shows up on an invoice.
Frequently asked questions
How fast do AI skills actually fade?
About half of accuracy-based trained gains are gone at roughly six and a half months with no practice, at a rate the 2025 Psychological Bulletin review puts at 0.08 per month. Plan against that default; your own rate depends on how often people use the hard parts.
Is my team losing AI skills or losing thinking skills?
Both are real and they differ. Losing the trained AI skill means people have reverted to simpler, worse use of the tool. Losing thinking skills means judgment itself weakened from handing decisions over. Decay is cured by practice, deskilling by redesigning the work. Left alone, the first produces the second.
Can I see decay in our usage dashboard?
No. Usage reports tell you the tool is open. Decay shows up as worse judgment inside work that still looks on time and on volume, which is entirely consistent with rising usage. Catching it needs a short capability check.
Which AI skill goes first?
Verification. Telling whether an answer is right fades before the ability to get an answer at all, and daily use gives it almost no practice. It is also the costliest one to be missing.
How often should we run refreshers?
Three or four real practice touches across the first six months, spaced further apart as the horizon lengthens. More frequent reminders fatigue people without exercising the skill.
Does a longer workshop slow the decay?
Not much. Length raises where the skill starts; it barely moves the slope. What moves the slope is whether people use the skill between sessions, which is why a shorter session with three follow-ups beats a longer one with none.
We trained people a year ago and nobody measured. Is anything recoverable?
The capability picture is, the before-and-after picture is not. Take a reading now, treat it as the new baseline, and schedule from here. Last year’s spend stays unprovable.
Do junior staff decay faster?
They face a different problem. Juniors often never built the unaided skill, so there is nothing to decay, and AI removes the practice that would have built it. BCG’s respondents put slower junior development among the effects they already saw, at 53%. The fix differs too: protected practice without the tool, not a refresher on it.
Is reinforcement just more training?
No, and buying it as more training is the common mistake. Reinforcement is short scheduled practice on the parts of the skill daily work does not touch, usually under an hour and observed. A second full program is what companies buy at month nine when nobody can prove the first worked.
Key takeaways
- AI skills decay on a measured curve: 0.08 lost per month for accuracy-based gains, half gone at about 6.5 months without practice.
- AI fluency sits in the fastest-fading category the research names, and daily tool use practices only the easy half of it.
- Verification fades before production. A team can keep shipping on time while quietly checking less, which no usage report will show you.
- Decay stays latent until a tool, a model or the work changes, so seat-license reporting cannot catch it by design.
- Reinforcement means three or four short observed practice sessions across six months, owned by a named person with funded hours.
- Specify the before-reading, the record of what was taught, and the reinforcement touches at purchase.
Working out whether your team’s AI capability is where you left it, or quietly below, takes a conversation, not a dashboard. Book a free AI-readiness call and we will walk through what to measure first and what the reinforcement should cost.
Arthur Wandzel