SSalarsNet

AI · Article 25 of 64

Synthetic Experience: How AI Could Compress Decades of Learning

Can simulation provide some of the pattern exposure that normally takes decades to accumulate?

By Randy Salars · October 7, 2026

A surgical resident learns to remove a gallbladder by standing at the table while an attending surgeon talks through the anatomy, then taking the instruments when the case is routine enough to allow it. The learning is real but rationed. Hours in the operating room are shared with everything else that must happen there, complications arrive on their own schedule, and the cases most instructive for the trainee are often the least convenient for the patient.

Simulation inverts that arithmetic. A resident can practice the same procedure fifty times in an afternoon, encounter a bile duct variant that appears in one case in a thousand, and make the mistake that would be catastrophic on a living person with no one harmed. The compression is the whole point: it manufactures the pattern exposure that a normal career would deliver slowly, unevenly, or never.

What “experience” is actually providing

Expertise is often described as knowledge, which undersells what long practice builds. A veteran surgeon recognizes an unusual anatomy in a glance, not because a textbook list runs through her head but because the visual configuration fails to match the thousands of previous configurations she has seen. This is pattern recognition built from cases.

The mechanism has a name in the psychology of skill. Ericsson and colleagues, in the 1993 paper that introduced deliberate practice, argued that expert performance grows through sustained, effortful practice on tasks just beyond current ability, with immediate feedback and repeated correction. The raw quantity of time matters less than the structure: hours spent repeating what you can already do produce little. The definition matters here because it sets the bar. A simulation is not automatically deliberate practice just because it is interactive. It becomes deliberate practice when it exposes a specific weakness, provides clear feedback, and forces adjustment.

This also explains why some domains resist compression. Chess and Go have crisp outcomes, immediate feedback, and rules that permit millions of self-play games. Managing a team through a product failure has none of those properties. The gap between the domains that yield to simulation and those that resist it is not about difficulty. It is about how cleanly the environment scores your moves.

The evidence that simulation transfers

The strongest documented case for simulation changing real-world performance comes from medicine, and it is older than the current AI wave. In a 2002 randomized, double-blinded study, Seymour and colleagues assigned surgical residents to learn a gallbladder dissection on a virtual reality simulator or through standard training. The simulator-trained residents then performed the real procedure 29 percent faster, made roughly six times fewer errors, and were far less likely to fail to progress or to injure tissue they should not have touched. This was a small study, and the outcome was measured in an operating room shortly after training. But the design was randomized, the assessors were blinded, and the measured effect landed on real patients.

The broader picture is consistent but thinner than headlines suggest. McGaghie and colleagues, in a 2011 meta-analysis, compared simulation-based medical education with deliberate practice against traditional clinical education. Across fourteen studies, the overall effect size was 0.71, with a confidence interval well clear of zero. Nothing in that range resembles magic, and the authors themselves flagged the small number of studies as a limit on confidence.

Note what the effect size means. It says that, on the measured tasks, simulation with structured practice beat the traditional apprenticeship route. It does not say simulation replaces clinical experience, and it does not establish that the trained skills persist months later or transfer to unfamiliar procedures. Those are separate questions, and the honest summary is that the near-transfer evidence is good and the long-horizon evidence is sparse.

The fidelity trap

The folk theory of simulation is that realism is the goal — that a training environment should look, sound, and feel like the real thing. The evidence does not support that intuition.

Norman, Dore, and Grierson, reviewing twenty-four studies in 2012, found something striking. High-fidelity simulators and low-fidelity simulators both beat no training at all. But when the two were compared directly, almost none of the studies found a significant advantage for higher fidelity, with average differences around one or two percent. Maneuvering a lifelike simulator rather than a plastic model did not reliably teach more.

That result is a useful inoculation against a specific kind of hype. If the mechanism of learning is pattern exposure and corrective feedback, then what matters is whether the practice surfaces the right variation and scores it honestly, not whether the pixels are convincing. A crude model that forces a good decision can outperform an expensive one that merely feels impressive. Realism can even hurt: a simulator that demands elaborate setup or that prioritizes cinematic detail over the reasoning under test spends the learner’s attention on the wrong thing.

For AI-generated scenarios the lesson transfers directly. The tempting failure is to invest in immersion — vivid characters, elaborate settings, seamless conversation — and to neglect the two things that provably matter, which are the quality of the situations presented and the quality of the feedback on what the learner did.

What can be generated, and what cannot

The clearest demonstration that self-generated experience can exceed human knowledge comes from games. AlphaGo Zero, described by Silver and colleagues in Nature in 2017, learned to play Go starting from random play, with no human game records and no hand-coded strategy beyond the rules. It trained by playing itself, using the outcome of each game to update the system that chose the moves. After three days it defeated the version that had beaten a world champion, one hundred games to zero, and it discovered strategies that human players had not found.

This is a genuine existence proof that synthetic experience can build superhuman skill. It is also a caution about generalization, because the game of Go supplies exactly what a messy human domain withholds: a perfect simulator, an unambiguous outcome, unlimited repetitions, and no cost to failure. The system was not navigating ambiguity about whether its decisions were good. It was told, every single game.

Something closer to human social reasoning appears in the Diplomacy work from Meta’s FAIR team, published in Science in 2022. Cicero combined a language model with strategic planning and played forty anonymous online games against human opponents, achieving more than twice the average human score and landing in the top decile. Among the most striking findings was that the system could hold convincing conversations with other players while pursuing its own plan.

The limits are equally instructive. The system sent and received roughly 292 messages per game, and expert review judged about a tenth of its messages to be inconsistent with its plan or with the state of the game. When human players detected the inconsistency, they were less likely to cooperate. So even in a task specifically designed around negotiation, the synthetic social partner could not fully maintain the fiction. The gap between generating plausible language and sustaining coherent social strategy under pressure is exactly the gap that decades of human experience closes.

Where the compression story gets oversold

It is easy to slide from these results to the claim that an AI can hand someone thirty years of judgment in a weekend. Each step of that slide deserves scrutiny.

The first overreach is treating scenario generation as learning. Producing a thousand varied dilemmas is not the same as producing a thousand learning events. Without feedback, a learner cycling through synthetic cases may reinforce their existing habits, including the bad ones, or drift into plausibility that is never corrected. Volume without scoring is entertainment.

The second is the fidelity illusion already described. Elaborate scenarios feel like learning and often are not, and there is no reliable internal signal that distinguishes the two — the same fluency illusion that misleads people evaluating their own comprehension.

The third is the simulator’s blind spots. A synthetic world is built from a model, and the model has assumptions. Where those assumptions are wrong, the learner is being trained, confidently, on a world that does not exist. This matters most in domains where the model was trained on the historical record, because the record contains the situations that happened and omits the ones that were avoided. A simulator built from past cases may systematically fail to prepare anyone for a scenario that has no precedent, which is often precisely the scenario that matters.

There is a further issue of luck and consequence. Human learning in the real world is welded to stakes. The decisions you remember are the ones that cost you something. A simulation can model consequences, but it cannot supply the endocrine reality of a real failure, and it is an open question how much of the learning travels without that weight. Deliberate practice research suggests that what matters is informative feedback rather than physical risk, which is encouraging for simulation. But the evidence on whether simulated consequence fully substitutes for lived consequence is not, as far as I can find, established.

What is demonstrated, what is plausible, and what is speculative

It helps to separate the layers, because the term “AI simulation” covers all three.

Demonstrated, with randomized or well-controlled evidence: surgical simulation improving real operating room performance; simulation with deliberate practice outperforming traditional clinical education on measured tasks; low fidelity matching high fidelity for transfer in most comparisons; self-play producing superhuman performance in a fully specified game; language-model-based agents negotiating at a high level in a structured multiplayer game.

Plausible engineering, consistent with the evidence but not established: adaptive scenario generation that targets a learner’s specific documented weaknesses; AI role-play for negotiation, interviewing, and difficult conversations, where the practice improves performance on similarly structured tasks; synthetic case banks that expand the rarity of pathologies or failure modes a trainee encounters.

Speculative, beyond current evidence: rapid development of broad practical judgment that transfers across domains; compression of tacit, embodied, socially embedded expertise into a short course; AI-generated experience substituting for years of apprenticeship in fields with ill-defined success criteria.

The distinction matters because the seductive version of this idea lives in the third category, and most of the confidence behind it is borrowed from the first.

Aviation’s long experiment with synthetic hours

If simulation could compress experience at scale, commercial aviation would have found out. The industry has logged decades of simulator use, and much of its culture is organized around what simulation can and cannot do.

The evidence there is genuinely mixed in a way that repays attention. Simulators have become the primary place where pilots rehearse emergencies that would be unsafe or impossible to practice in flight, and the FAA emphasizes that such training and proficiency checks are where knowledge and skills are meant to be maintained. At the same time, regulators have had to intervene for the opposite problem. In 2013 the same agency issued a safety alert warning that continuous reliance on automated systems was degrading pilots’ ability to hand-fly an aircraft when the automation failed, and it asked operators to build real opportunities for manual flight into routine operations.

Read together, the two facts describe a boundary. Pilots can rehearse a great deal of the cognitive and procedural content of flying in a synthetic environment, and that rehearsal counts. But the basic manual control that the automation made unnecessary could only be maintained by exercising it, and the agency’s recommended remedy was to build practice back into ordinary line operations rather than assume it would happen on its own. The lesson for AI simulation is that some capabilities are preserved by the structure of the work and cannot be contracted out to a separate training module. If the day job no longer exercises a skill, adding practice hours may not restore it as reliably as expected.

Designing practice that actually transfers

If the compression is real but bounded, the design question becomes specific. Several principles follow from the evidence above.

Score the decisions, not the atmosphere. Fidelity is not where the learning lives. The simulator should confront the learner with the cases that vary in a way that reveals a weakness, and it should tell them, quickly and specifically, whether the call was right.

Reward variety over repetition. The reason a career produces expertise is the sheer range of situations it accumulates. A synthetic system can compress that range by deliberately surfacing the rare case, the awkward variant, and the situation the historical record barely contains. The value is in the coverage of the distribution, not the number of runs.

Build the feedback into the loop. Deliberate practice research is emphatic that practice without informative feedback does not consolidate into skill. A scenario that ends quietly teaches less than one that says clearly where the reasoning broke.

Include the breaking cases. A training environment should deliberately include situations where its own assumptions fail, because that is where the learner must discover that the model is not the world. Meta’s Diplomacy results make the point concretely: the synthetic partner was convincing until the specific moment it was not, and the humans who caught it learned something the designers could not have scripted.

Measure transfer outside the simulation. The only test that matters is whether the person performs better on the real thing. A learner who improves inside the environment and nowhere else has learned the environment.

The larger question

Simulation is best understood not as a replacement for experience but as a second channel for pattern exposure. It supplies the cases that the world delivers too slowly, too rarely, or too dangerously. It does not supply the stakes, the ambiguity, or the open-endedness that make a career of practice so difficult to compress.

The claim worth making, then, is narrower than the headline version. Simulation can manufacture a substantial portion of the pattern exposure that normally takes decades, in domains where the task has clear feedback and where the variation can be represented honestly. It cannot, on the current evidence, manufacture judgment that travels across domains, and it cannot substitute for the part of learning that depends on consequences that actually land.

The practical stance is to use it as a supplement with a high expectation of near transfer and a low expectation of far transfer — and to keep the real thing in the loop, because that is the instrument that measures whether any of it stuck.

Sources and further reading

Discussion

What would you add or question? Add your comment below. A human reviews it before publication.

Loading comments…

Join the discussion

Comments are public after approval. Please do not include links, email addresses, or private information. For one short AI reply, address @AIGuide in your comment or reply to its opening comment. Cloudflare verifies submissions to limit spam. Read our community guidelines.

The wider community forum is also open: Browse article discussions in the forum · Forum home