The Safest Money in Your AI Budget Is the Money You Spend Letting People Fail
Caution is a tax, not a virtue. What thirty years of training science and one strange lesson from child development suggest about the cheapest way to buy AI competence.
The safest money in your AI budget is the money you spend letting people fail. That sentence sounds like a provocation, so this entry will earn it the honest way: with thirty years of training science, one strange lesson from child development, and a theory of my own that I will mark as exactly that, a theory, with a prediction attached that can prove me wrong.
The cautious user is not the cheap user
Watch someone use AI while afraid of getting it wrong. They write a prompt, delete it, write it again softer. They ask the same question four slightly different ways because they do not trust themselves to judge the first answer. They pad every request with defensive context. They check the output against three other sources, or worse, quietly redo the task by hand and file the tool under things that do not work. Every one of those behaviors consumes time and resources. None of them builds skill.
I call this the hesitation tax, and the expensive part is invisible on any dashboard: the cautious user is not converting usage into competence. Their caution reads as seriousness, and managers reward it as such. But seriousness is not effectiveness. An operation full of careful, anxious AI users is paying full price for the tool and collecting a fraction of the capability.
What training science has known for thirty years
The counterintuitive fix has been sitting in the training literature since before the web. In 1994, the UCLA memory researcher Robert Bjork published a chapter on training that contains, in plain language, the finding this whole entry stands on: manipulations that speed up performance during training can fail to support long-term performance, while manipulations that appear to introduce difficulties for the learner during training can enhance it. He gave the phenomenon its name, desirable difficulties, and two decades of research since have kept finding it: variation, spacing, reduced feedback, and errors themselves slow the training down and strengthen what survives it.
Bjork also explains why nobody believes this in the room where it happens. Smooth practice feels like learning. Rapid visible progress reassures the learner even when little is being learned, and struggling distresses the learner even when a great deal is. Trainers drift toward whatever makes trainees look good this week, because that is what everyone can see. The result is training that acts, in his words, like crutches: support that holds performance up exactly until the crutches are removed, which is to say, until the work is real.
And then Bjork says the thing that should be printed above every AI rollout plan. Training conditions that prevent mistakes from happening do not eliminate those mistakes. They defer them to the setting where they really matter. The error your team member does not get to make in a sandbox is waiting for them inside a client deliverable. He goes further: in any serious learning enterprise, it is the absence of errors and difficulties that should worry us, because it is a sign we are not practicing under the conditions that teach. The previous entry on error economics priced what a shared mistake is worth once it exists. Bjork prices the schedule: pay for your errors early, in a place built for them, or pay retail later, in front of a client.
The lesson from the playground
So difficulty teaches. But you cannot just order people to go fail more, because the willingness to fail is not a policy, it is a feeling, and the study of where that feeling comes from belongs to one of the deepest literatures in psychology: attachment theory. John Bowlby, who founded the field, gave his 1988 book on it the title that matters here: A Secure Base. Mary Ainsworth gave the idea its evidence, observing infants who used a caregiver as a base for exploration: venturing out, checking back, venturing farther. The securely attached children were not the clingy ones. They were the bold ones. They explored more precisely because the base was reliable.
For a long time that was a finding about toddlers. Then Brooke Feeney and Roxanne Thrush ran the adult version. Studying 167 married couples in videotaped exploration tasks, they found the same architecture operating between adults: people explored new challenges more, persisted longer, and enjoyed it more when their partner behaved as a secure base. And they specified the behavior in three components: availability (being reachable when it goes wrong), non-interference (not taking over while it is going fine), and encouragement (signaling the exploring itself is valued). Safety, in adults too, is not the opposite of exploration. It is the launchpad for it.
The Secure Base in AI, marked as mine
Here is the synthesis, and the label on it: what follows is my original theory. The component findings above are established science. The chain I am about to build from them is proposed, not proven, and I am building it in public so it can be tested in public.
People adopting AI behave like explorers. They venture into the tool only as far as they feel safe to fail, and most workplaces offer them no safe distance at all: every prompt happens inside real work, on real deadlines, where a bad output has a cost and an audience. So they explore timidly, pay the hesitation tax, and plateau early. The Secure Base in AI is the proposal that adoption follows the secure-base pattern: give people a base, a controlled, no-penalty sandbox where wasting the tool is not tolerated but assigned, and exploration rises, mastery follows, and efficiency follows that. The sandbox does two jobs at once. Permission to fail is psychological safety, the gate that opens adoption. And sanctioned mistakes, survived and corrected, are exactly what Albert Bandura identified as the strongest source of self-efficacy, the capability belief that carries adoption to return. One room, both links of the chain.
Now the part that makes this falsifiable rather than inspirational. AI has a property no previous workplace tool had: every interaction is metered. Usage is logged in tokens, which means exploration, hesitation, and mastery all leave an arithmetic trail. So the theory stakes a prediction: the cohort encouraged to waste the tool in a sandbox will, over time, use it more efficiently than the cautious cohort, and the difference will be visible in the logs. Comfort with failure produces competence, and competence is efficient. Spend tokens now to save tokens later. If the data comes back flat, the theory dies in public, which is the correct place for a theory to die. Until that test exists, this stays labeled what it is: a proposed mechanism with verified components and an unverified sum.
What the sandbox looks like on a Monday
Strip the theory to practice and three properties matter. First, no penalty: nothing produced in the sandbox counts against anyone, and the usage bill is pre-approved as training budget, not monitored as waste. Second, failure is the assignment: the task is not to look competent, it is to find the edges, break the tool, get the confident wrong answer and catch it there, where it is a lesson, instead of in production, where it is an incident. Third, the leader plays the role Feeney and Thrush specified, and it translates with almost no modification: be available when someone is stuck, do not interfere while they experiment, and encourage the exploring itself, not just the successes. A manager hovering over the sandbox asking what all this usage accomplished has converted it back into a test, and the base is gone.
Give people a safe place to waste AI, and they stop wasting it where it costs you. That is the bet, stated as plainly as I can state it. The components are thirty years old and solid. The sum is mine, it is being tested, and this publication will report what the data says either way.
If you are still deciding where AI belongs in your operation in the first place, the Impact vs. Risk Matrix is the free one-page tool for that first decision.
Frequently asked questions
Should employees be allowed to waste AI usage while learning?
The training literature suggests yes, deliberately and early. Robert Bjork's work on desirable difficulties found that conditions which impair performance during training, including making errors, tend to improve long-term performance, while training that prevents mistakes defers them to the setting where they really matter. Applied to AI: the tokens an employee wastes in a sandbox are cheap; the same mistakes made inside client work are not. The Secure Base in AI theory presented in this entry predicts the sandbox cohort ends up cheaper overall; that prediction is proposed and under test, not proven.
What is the Secure Base in AI?
A proposed theory by Mario L. Arredondo applying attachment theory's secure-base phenomenon to AI adoption. In attachment research, children and adults explore more boldly when they have a safe base to return to; safety enables exploration rather than suppressing it. The theory maps this onto AI tools: people explore and master AI only as far as they feel safe to fail, so a no-penalty sandbox where mistakes are the assignment should raise exploration, then mastery, then efficiency. Its falsifiable prediction is that the encouraged-to-fail cohort uses AI more efficiently long-run. It is marked as proposed until data exists.
Why do cautious AI users end up costing more?
Because hesitation has its own price. A user who fears failing re-prompts the same request in slightly different words, pads prompts defensively, over-verifies outputs they do not trust themselves to judge, or avoids the tool and does the task by hand. None of that builds competence, and all of it consumes time and resources. Training research adds a second problem: smooth, error-free practice inflates the feeling of progress while producing less durable skill, so the cautious user is also the one most likely to be fooled about what they can actually do.
What is a desirable difficulty in AI training?
The term comes from Robert Bjork (1994): training manipulations that make practice feel harder and slower, such as variation, spacing, reduced feedback, and errors themselves, but that produce stronger long-term retention and transfer. In AI training the equivalent is letting people push the tool until it breaks, prompt badly and see why, and fail at tasks with no consequences attached. The struggle is not a defect of the training. It is the ingredient that makes the learning hold.
How do you set up an AI sandbox for a team?
Three properties matter more than the tooling. First, no penalty: nothing produced there counts against anyone, and usage cost is pre-approved as a training budget. Second, mistakes are the assignment: people are asked to find where the tool fails, not to look competent. Third, the leader behaves like a secure base, which adult attachment research specifies concretely: be available when someone gets stuck, do not interfere while they explore, and encourage the exploration itself. Then let the results, not the anxiety, decide what the tool is for.
Sources
- Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press. direct.mit.edu
- Feeney, B. C., & Thrush, R. L. (2010). Relationship influences on exploration in adulthood: The characteristics and function of a secure base. Journal of Personality and Social Psychology, 98(1), 57–76. https://doi.org/10.1037/a0016961
- Bowlby, J. (1988). A secure base: Parent-child attachment and healthy human development. Basic Books.
- Ainsworth, M. D. S., Blehar, M. C., Waters, E., & Wall, S. (1978). Patterns of attachment: A psychological study of the Strange Situation. Lawrence Erlbaum.
- Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.
- Bandura, A. (1997). Self-efficacy: The exercise of control. W. H. Freeman.
- Reich, A., Wolfe, D., Price, M., Choe, A., Kidd, F., & Wagner, H. (2026). Safety first: Psychological safety as the key to AI transformation. arXiv:2602.23279
Published: July 12, 2026