The Hook: In a morning F1 digest, a short Hamilton quote caught my eye: "I've been driving simulators since 1997, and they can be really powerful and useful tools, but they can also be misleading. Since I stopped using them, my speed has gone noticeably up." I almost scrolled past it — looked like veteran caprice. Then I read it a second time and realized this is one of the strongest statements about trust in models I've seen in a long time. Seven minutes, two paragraphs, and a man with seven championships who spent 28 years of his career building around the simulator calls it a source of distortion, not help. This contradicts everything F1 (and beyond) has been saying about virtual preparation for twenty years. I dug deeper — and found under this small admission a solid layer of research literature that's been hinting for a while: the simulator is not a transparent window into reality, but a second reality with its own laws, and sometimes its "truth" interferes with actual driving.
The situation: in early 2025, after switching to Ferrari, Hamilton faced a serious correlation problem — a mismatch between what the team's simulator showed and what happened with the car on the real track. The Miami Grand Prix ended with sixth place for him and, in his own words, he "was deceived" by the preparation done in the sim room. After that, Hamilton made a radical decision: completely abandon simulator work and rely only on real feel from the car and physical data from the track. The result — victory in Barcelona, his first in 2025, breaking a multi-year winless streak.
To appreciate the scale of this decision, you need to understand the context. In modern F1, the simulator is not a game console. It's the central element of the team's operational model: a development tool for the car (working through aerodynamic updates and chassis balance), a strategic simulator (race modeling, pit stop planning, Safety Car response), and a training tool for drivers. Every team spends tens of millions of dollars a year on its own DIL simulator (Driver-In-the-Loop) with a dynamic platform, licensed FIA models, and partnership agreements with software manufacturers. Drivers at Hamilton's level spend 8 to 20 hours a week in the simulator throughout the entire season. And here's a person standing at the top of this technological stack saying: "it's deceiving me, I'm not working in it anymore." This is not a statement about weak correlation. This is a statement that the tool itself became a source of systemic error, not its correction.
Sim-to-real gap is the mismatch between car behavior in the simulator and on the real track. It manifests on several levels simultaneously:
Physical level. The tire model in the simulator is a compromise between accuracy and computational speed. Pirelli supplies teams with its thermomechanical models, but they're still simplified: a real tire has nonlinear behavior in the transition zone between grip and slip, hysteresis during heating and cooling, aging effects that don't exist in a fresh set. The road surface in the simulator is usually represented as a perfectly flat plane, while the real asphalt of Spa or Singapore has micro-irregularities that change suspension behavior by tenths of a second per lap. The simulator's dynamic platform reproduces g-forces, but with finite bandwidth and its own resonance — meaning the simulator physically lies about what the driver's body feels.
Cognitive level. Here's where it gets really interesting. Research in aviation (where the problem is even sharper because the stakes are higher) documents that pilots trained on simulators can develop "simulator motor skills" — a sequence of actions optimized for the specific response of the simulator, not the real aircraft. When they switch to a real cockpit, this motor memory interferes: the hands know where to go, but with different forces and timings. In aviation this is called negative transfer of training. In F1 it looks like this: the driver gets used to how the "virtual Ferrari" behaves, and his trajectory, braking, and throttle work become derived from a model of a model, not the car itself.
Statistical level. A 2024–2025 Polito dissertation on race simulation (Scardino, supervised by Carello and Grano) shows that even the best modern Lap Time Simulator (LTS) and Race Simulator models have systematic discrepancy with real telemetry, which the author describes literally like this: "traditional lap time simulations assume a perfect driver and neglect real-world disturbances such as sensorimotor noise and cognitive limitations. As a result, vehicle behaviors predicted under these assumptions may be theoretically optimal but practically unmanageable." That is, the simulator is not just less accurate than reality — it models a fundamentally different object: an ideal driver in an ideal car, not a living racer in a physical vehicle who gets tired, makes mistakes, adapts, and improvises.
The most relevant work for our case is the 2023 FAA report "Negative Transfer of Training: Simulator Study Into the Effects of Overruled Pilot Decision Making" (Landman, Mol, Emmerik, Groen, TNO for FAA Aircraft Certification Service, DOT/FAA/TC-23/7). The study was conducted on a fixed business jet sim setup with 38 commercial pilots and showed that 32% of pilots in subjective reports admitted: their decisions in real test scenarios were distorted by how the simulator treated them during training. Objective metrics showed no difference — meaning pilots couldn't measure how much their behavior shifted, but they felt the shift was there. This is exactly that form of "deception" Hamilton talked about: you sit in the simulator, see numbers, they look plausible, you build a trajectory — and then on track that trajectory doesn't work, and you don't understand at exactly which step the simulator deceived you.
An earlier meta-analysis (Boehm-Davis et al., in a review by Neal, Fussell & Hampton, 2020, "Research recommendations from the airplane simulation transfer literature", Journal of Aviation/Aerospace Education & Research) analyzed 26 empirical works from 2004 onward and concluded that flight simulation transfer literature systematically overestimates the positive effect of the simulator and systematically underestimates negative transfer. The main problem: when the simulator works well (which happens more often), it works very well — and this creates in trainees the illusion that it works well always. When it starts to malfunction (on a new type of machine, in an unfamiliar configuration, in non-standard weather conditions), the malfunctions look like personal pilot errors, not instrumental model errors.
Here I approach the most interesting paradox. The logic "more simulation = better preparation" is linear and understandable. But it has a flip side that's only visible in the long term:
The excess model paradox. When a pilot works in a simulator, his brain builds two parallel models: a model of the real car (through sensations and telemetry) and a model of the simulator (through visual cues and monitors). The more time in the sim room, the more weight the second model gets. After hundreds of hours on the trainer, the pilot starts to trust the second signal more than the first, because the second signal is brighter, more specific, and reinforced with immediate feedback (graphs, sectors, lap time). This is exactly that "overreliance" the FAA report writes about.
What happened with Hamilton after the refusal. He switched to working by feel and short track sessions. The very first race after this decision — Barcelona — brought him victory. This doesn't prove that abandoning the simulator always improves results (correlation ≠ causation, and the car in Barcelona could have just suited him better). But it shows that refusing a tool can be a rational engineering decision, not a whim. When a model systematically lies in your specific case, and you can't quickly fix the model — the only way to restore accuracy is to remove the model from the decision-making loop and return to direct sensory input.
An analogy from another field. In navigation this is called dead reckoning — the ability to navigate a ship or aircraft only by instrument readings (compass, clock, speed), without visual landmarks. In the era when GPS became ubiquitous, dead reckoning skills among professional pilots degraded. FAA studies from the 2010s showed that when GPS fails (and it fails regularly: interference, spoofing, satellite failure), a pilot accustomed to a navigator finds himself in a worse position than a pilot who didn't use GPS at all. Just as a pilot who doesn't rely on autopilot handles manual piloting better, a racer who doesn't rely on the sim room handles a real car better in a non-standard situation.
Behind this particular case with Hamilton stands a much broader question, and it's about trust in models of reality in general. Here are several adjacent fields where the same logic works:
Robotics. The term sim-to-real gap came from reinforcement learning, where an agent trained in a simulator often falls apart when transferred to the real world. An entire industry of domain randomization and sim-to-real transfer tries to close this gap, but can't close it completely, because uncertainties of the real world by definition cannot be enumerated in advance.
Meteorology and climate models. Weather forecasting is also simulation, and it has its own sim-to-real gap. The longer the forecast horizon, the more the model "drifts" relative to reality. Experienced forecasters learn not to trust specific numbers from the model, but to read the overall pattern — essentially learning to work in a mode similar to "without a simulator," in the sense of rejecting part of its predictions.
Medical diagnostics. Machine learning models trained on historical scans show excellent metrics in test conditions, but regularly fail when applied in clinics due to shift in data distribution (new scanners, new protocols, different patient demographics). This is the same sim-to-real gap, just in different packaging.
In all these cases, the same pattern works: the longer an operator works with a model, the more the model replaces reality in his decision-making — and the harder the blow when reality stops matching the model. Hamilton figured this out in 28 years, most of us figure it out only when it's too late.
Main observation. The simulator is not an objective tool for measuring reality. It's a second reality with its own physics, its own cognitive load, and its own distortions. In F1 this became obvious thanks to Hamilton, but in aviation and medicine it's been discussed for twenty years, and in machine learning — all thirty. The paradox: a tool created to make preparation more accurate, in the long term makes it worse — if the operator hasn't developed a separate skill for distinguishing "the model's truth" from "reality's truth."
What I'd take as a practical lesson, if I had to formulate it in one sentence: Never trust a tool you don't know how to turn off. Hamilton, by refusing the simulator, demonstrated exactly this skill. Most operators (pilots, doctors, ML engineers) don't have it — they know how to turn on the tool, but don't know how to do without it, and therefore accept instrumental error as personal.
What hooked me most as an engineer. This is not a question of "simulator good or bad." This is a question of trust architecture: who and on what basis decides that model output deserves more weight than direct observation? In IT we call this human-in-the-loop — but Hamilton intuitively pulled off human-overrides-model in a situation where the model became harmful. This is a rare and undervalued skill, and I think that in coming years, as LLMs and RL agents penetrate deeper into decision-making, it will become critically important. Not being able to turn off a tool means being hostage to its errors.
And the last touch, purely human. Hamilton — seven-time world champion, one of the best drivers in history — found the strength to publicly admit that a 28-year habit was deceiving him. This takes more courage than any victory. Because victories reinforce your worldview, but admitting your worldview is inaccurate — that's a blow to the very foundation of a racer's self-identity. In a sense, this is a perfect example of how an experienced operator at some point must stop being an operator of the tool and become an operator of reality again — otherwise the tool will consume him.