Your successes are lying to you
You finished the project on time. The client was satisfied. The morning routine ran smoothly for three weeks straight. The presentation landed well. Everything confirms that your system works.
Except it does not tell you that. Success confirms that your system worked under those specific conditions. It tells you nothing about where the system is fragile, which assumptions are load-bearing, or what happens when conditions shift. Success is a flat signal. It says "keep going" without saying where you are headed or what you are standing on.
The previous lesson established that every correction carries a cost. But there is a corollary that reverses the framing entirely: every error carries information. Not as a metaphor. Not as a motivational platitude about "learning from failure." As a literal, measurable fact about the information content of different feedback signals. Errors reveal the boundaries, assumptions, and structural weaknesses of your systems in ways that success categorically cannot. The question is not whether errors are costly — they are. The question is whether you are extracting the information they contain, or wasting it.
Prediction error: how your brain actually learns
The neuroscience of learning is not a mystery. Your brain learns primarily through prediction errors — the discrepancy between what you expected to happen and what actually happened.
In the 1970s, Robert Rescorla and Allan Wagner formalized this in what became the Rescorla-Wagner model: learning occurs not when a stimulus is present, but when the outcome differs from what was predicted (Rescorla & Wagner, 1972). When everything goes as expected, the teaching signal is zero. Your brain has nothing to update. When the outcome deviates from the prediction — when you are wrong — the discrepancy generates a learning signal proportional to the size of the error.
Wolfram Schultz and colleagues later discovered the neural substrate of this mechanism. Dopamine neurons in the midbrain fire not in response to rewards, but in response to unexpected rewards — reward prediction errors (Schultz, Dayan, & Montague, 1997). When you receive more reward than predicted, dopamine neurons fire above baseline: a positive prediction error. When you receive exactly what you predicted, they remain at baseline: no teaching signal. When you receive less than predicted — when reality disappoints your expectation — dopamine drops below baseline: a negative prediction error.
This is the mechanism. Not a metaphor. Your brain literally generates a stronger learning signal when things go wrong than when things go right. A fully predicted success produces zero dopamine-based teaching signal. An unexpected failure produces a large one. The neurochemistry is telling you the same thing the information theory says: errors carry more bits of information than confirmations.
This is why you can repeat a comfortable routine for years without improving. If every outcome matches your prediction, there is no prediction error, no teaching signal, and no learning. The routine feels productive because it is familiar. But familiar and productive are different things.
Productive failure: Kapur's experimental proof
Manu Kapur, an educational researcher now at ETH Zurich, spent over a decade testing a counterintuitive hypothesis: what if letting students fail before they receive instruction produces deeper learning than teaching them the correct method first?
The answer, across more than 166 experimental comparisons involving over 12,000 participants, is unambiguous. Students who struggled with complex problems before receiving instruction — who failed first — significantly outperformed students who received instruction first on measures of conceptual understanding and transfer (Kapur, 2014; Sinha & Kapur, 2021). Both groups achieved equivalent procedural knowledge. But the group that failed first understood why the procedures worked and could apply them to novel problems that the instruction-first group could not solve.
Kapur calls this "productive failure." The term is precise. The failure is productive because it activates prior knowledge, exposes gaps in understanding, and generates what Kapur calls "knowledge activation and differentiation" — the learner discovers what they do not know in a way that passive instruction cannot replicate. The subsequent instruction lands on prepared ground. The student has already mapped the shape of the problem, encountered the boundaries of their current understanding, and generated candidate solutions that did not work. The correct solution, when it arrives, fits into a structure the errors built.
This is not an argument for careless mistakes. It is an argument for the informational asymmetry between success and failure. When your system works, you learn that it works. When it fails, you learn where it fails, why it fails, and what assumption was wrong. The second kind of learning is structurally richer.
Aviation and Google: industries that harvest error feedback
Some of the highest-performing systems in the world are organized around a single principle: errors are the primary source of system improvement.
Commercial aviation has the lowest fatality rate of any transportation mode, and the architecture behind that safety is not better pilots or better planes — it is a systematic infrastructure for extracting maximum information from every error. The Aviation Safety Reporting System (ASRS), established in 1976, provides confidential, non-punitive reporting of safety incidents. Crew Resource Management (CRM) protocols, developed after analyzing fatal accidents in the 1970s, restructured cockpit communication to ensure errors are surfaced rather than suppressed. Every accident and serious incident triggers a formal investigation whose goal is not blame but structural understanding: what failed, why, and what systemic change would prevent recurrence.
Google's Site Reliability Engineering (SRE) team adopted the same architecture for software systems. Their blameless postmortem process, documented in the Site Reliability Engineering book (Beyer et al., 2016), operates on an explicit principle: "You can't fix people, but you can fix systems and processes to better support people making the right choices." When a system outage occurs, the postmortem investigates the systemic conditions that made the error possible — not who made the mistake, but what about the system's design meant that a reasonable person, acting on the best available information, would make that mistake.
The insight in both cases is identical: errors are not aberrations to be punished. They are the richest source of feedback about system weaknesses. An organization that suppresses error reporting — through blame, punishment, or cultural shame — is an organization that has voluntarily blinded itself to its most valuable information stream. Amy Edmondson's research on psychological safety at Harvard confirmed this empirically: teams that feel safe reporting errors learn faster and perform better than teams where errors are hidden (Edmondson, 1999).
The personal parallel is exact. If your response to an error is shame, self-criticism, or a vague resolution to "do better," you are suppressing your own error reporting system. You are the organization that punishes the messenger.
The AI parallel: error is the only teacher
In machine learning, error is not one input among many. It is the only mechanism through which a neural network learns.
During training, a network processes an input and produces an output. A loss function computes the discrepancy between that output and the desired output — the error. Backpropagation then calculates how each weight in the network contributed to that error, and gradient descent adjusts the weights to reduce it. The next iteration produces a slightly smaller error. Over millions of iterations, the network converges on weights that minimize the loss function.
Remove the error signal and the network learns nothing. It does not matter how many inputs you feed it, how large the architecture is, or how much compute you allocate. Without the gradient — the precise, per-weight feedback about how each parameter contributed to the error — the network is a static function that never changes. The error signal is not a component of learning. It is learning.
The architecture of backpropagation also reveals something about the structure of error feedback that applies directly to human systems. The gradient is not a single number that says "you were wrong." It is a vector — a detailed, high-dimensional signal that tells each weight in the network how much it contributed to the error and in which direction it should change. The richness of this error signal is what makes deep learning possible. A simple "right/wrong" signal would be insufficient to train a network with millions of parameters. The network needs error feedback that is specific, directional, and granular.
Your own error feedback works the same way. "That meeting went badly" is a scalar — a single, low-information signal. "The meeting went badly because I presented the options before establishing the decision criteria, which meant the group anchored on their initial preferences before understanding the tradeoffs" is a gradient — a specific, directional signal that tells you exactly which parameter to adjust.
Why success feedback is structurally impoverished
Information theory offers a formal way to understand why errors carry more information than successes.
Claude Shannon's foundational insight was that the information content of a signal is inversely proportional to its probability (Shannon, 1948). A highly probable event — one you expected — carries little information. An improbable event — one that surprised you — carries a lot. When your system works as predicted, the signal has low information content: it confirms what you already believed. When your system fails in a way you did not predict, the signal has high information content: it reveals something about reality that your model did not capture.
This is not an argument against paying attention to success. It is a structural observation about the distribution of learning opportunities. If your system succeeds 80% of the time, then 80% of your experience is confirming your existing model. The 20% where it fails is where your model is being updated — where the actual learning lives. If you allocate your attention proportionally to the frequency of each outcome, you spend 80% of your analysis budget on the 20% of the information and 20% of your budget on the 80% of the information.
The highest-performing individuals and organizations invert this ratio. They treat success as expected and errors as data. Not because they are pessimists, but because they understand where the information density is highest.
A protocol for extracting structural learning from errors
Understanding that errors carry valuable feedback is necessary but not sufficient. You need a systematic process to extract that value. Most people skip this step entirely — they feel the error emotionally, resolve to avoid it, and move on with zero structural learning captured.
Here is a protocol that converts error feedback into system improvements:
Step 1: Separate the signal from the noise. Within 24 hours of the error, write one sentence describing what happened mechanistically. Not "I failed" or "it went wrong," but "The proposal was submitted two days late because I underestimated the review cycle by three days." Strip the emotion. Isolate the mechanism.
Step 2: Identify the incorrect assumption. Every error reveals a gap between your model of reality and reality itself. What did you assume that turned out to be wrong? "I assumed the review would take two days because it took two days last time" — this surfaces the assumption that past cycle times predict future ones without accounting for reviewer availability.
Step 3: Determine if the error is systematic or stochastic. Some errors are random — a one-time confluence of factors unlikely to recur. Most errors that feel random are actually systematic: they are the predictable output of a structural weakness that will produce the same class of failure again. If you have seen a similar pattern before, it is systematic.
Step 4: Design a single, testable adjustment. Not a vague resolution. Not a wholesale system redesign. One specific change to your process that addresses the identified assumption. "Add two buffer days to any estimate that depends on external review timelines." This adjustment is testable: you can observe whether it reduces the error rate.
Step 5: Expect the adjustment to produce its own errors. This is critical. The adjustment is a hypothesis, not a solution. It will interact with your existing system in ways you cannot fully predict. The errors it produces are the next round of feedback.
From punishment to protocol
The shift this lesson asks you to make is not motivational. It is architectural.
Most people have an error response that looks like this: error occurs, emotional reaction fires, vague resolution forms, attention moves to the next thing. Zero structural information captured. Zero system parameters adjusted. The error is processed as punishment — something to feel bad about — rather than as signal — something to learn from.
The previous lesson (L-0497) showed you that error correction has a cost. This lesson shows you that error feedback has a value — and that value exceeds the value of success feedback by a wide structural margin. Your brain is wired to learn more from prediction errors than from confirmations. The research on productive failure demonstrates that struggling and failing before mastering a concept produces deeper understanding than getting it right the first time. The highest-performing industries in the world — aviation, elite engineering — are organized around systematic extraction of information from errors.
The next lesson (L-0499) takes this further. If errors are your most valuable source of feedback, then a system that never produces errors is not a perfect system. It is a system that is not being tested at its boundaries. Building error tolerance into your expectations means designing systems where errors are expected, budgeted for, and welcomed as the information stream they are — not catastrophes to be avoided at all costs.
You are not trying to eliminate errors. You are trying to eliminate wasted errors — errors that occur, inflict their cost, and produce no learning because you lacked the infrastructure to capture what they were telling you.
Sources:
- Rescorla, R. A., & Wagner, A. R. (1972). "A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement." In A. H. Black & W. F. Prokasy (Eds.), Classical Conditioning II: Current Research and Theory. Appleton-Century-Crofts.
- Schultz, W., Dayan, P., & Montague, P. R. (1997). "A neural substrate of prediction and reward." Science, 275(5306), 1593-1599.
- Kapur, M. (2014). "Productive failure in learning math." Cognitive Science, 38(5), 1008-1022.
- Sinha, T., & Kapur, M. (2021). "When problem solving followed by instruction works: Evidence for productive failure." Review of Educational Research, 91(5), 823-861.
- Beyer, B., Jones, C., Petoff, J., & Murphy, N. R. (2016). Site Reliability Engineering: How Google Runs Production Systems. O'Reilly Media.
- Edmondson, A. (1999). "Psychological safety and learning behavior in work teams." Administrative Science Quarterly, 44(2), 350-383.
- Shannon, C. E. (1948). "A mathematical theory of communication." Bell System Technical Journal, 27(3), 379-423.