Most air traffic control games are worked with a mouse, or with typed commands in a fixed format. Talk to Tower is worked by talking, the way the job is actually done. That sounds like a presentation detail. It is not — it changes what the game can be about.
A typed command interface has to define a grammar, and then the player’s job quietly becomes remembering it. Type BAW214 H240 and you are not practising air traffic control; you are practising the notation. The moment your phrasing drifts from the grammar, nothing happens and the game tells you that you were wrong.
Real controlling does not work that way. A controller says whatever gets the meaning across, at speed, with hesitation and self-correction, and the crew works out what was meant. If a game cannot absorb that, it cannot simulate the part of the job that is actually difficult.
Three layers, cheapest first
Speech arrives as text and is handed to the cheapest thing that can resolve it. Only what genuinely needs more understanding gets it.
Deterministic correction
An explicit self-correction is handled before anything else, by code, never by a model. “Turn right — sorry, correction, left heading 240” can only ever execute as a left turn, because only the words after the final correction marker survive. This is safety-critical and is not left to a probability.
Deterministic parse
Ordinary clearances are pattern-matched locally, with no network call at all. “Slow to 180”, “bring it down to 3000”, “steer 240”, “fly heading two four zero” and aviation pronunciations like “niner” all resolve here. So does any question answerable from live simulation state, which is answered from the aircraft itself and can never change it.
Language model
What is left is the genuinely conversational: negotiation, offers, social speech, hesitant fragments, ambiguity. That goes to a model with the selected aircraft’s live state and recent radio history, and it is locked to that one flight so it can never answer about, or act on, a different aircraft.
The ordering matters for feel as much as cost. The common case has no network round trip, so a routine clearance is immediate; the unusual case is allowed to take a moment, which is also true of a real crew working out what you meant.
The model never moves an aircraft
This is the part that matters most, and it is a deliberate boundary. A language model can propose a clearance. It cannot apply one. Whatever comes back is re-parsed and re-validated by the deterministic engine before any aircraft state changes.
So the failure mode of a misunderstanding is a pilot asking you to say it again, not an aeroplane doing something nobody asked for. And when the radio genuinely cannot understand you, the crew tells you in character — “say again”, “confirm heading” — rather than the game showing you an error about a service. A frequency that reports its own infrastructure problems stops being a frequency.
Safety-critical ambiguity is refused rather than guessed. The one thing the radio will not do is choose between 030 and 230 on your behalf.
Nothing is scripted, including whether you succeeded
Understanding your words is only half of it. What happens next is computed, not staged.
An approach clearance is permission to capture two radio beams, not a command to land. Whether the aircraft actually captures them is worked out from its real position, height and heading against the published Mumbai procedure: inside roughly 10 to 14 miles, within 30° of the runway heading, and beneath the descent path. Issue it from the wrong place and nothing happens, because the aeroplane genuinely was not in the beam.
That is why the game can be honest with you about being wrong without ever having to invent a rule. The geometry either worked or it did not, and you can see which on the radar.
Crews that sound like their situation
Every flight carries its own voice, accent, experience and assertiveness, assigned once and kept for the life of that flight. On top of that, how a line is delivered is derived from what the crew is actually experiencing: workload, accumulated delay, how long they have been left without an instruction, and how much fuel they have.
A crew under fuel pressure speaks faster and more insistently. A fatigued long-haul crew drags. A crew that has been ignored will chase you — neutrally at first, then stating what they need, then making it your problem. That escalation is not a script; it is the honest consequence of leaving an aircraft alone, and it is the thing that makes the frequency feel occupied rather than decorative.
One frequency, half duplex, as in reality: key up while a pilot is transmitting and you cut them off.
What this is not
Worth stating plainly, because the phrase “AI” is doing a lot of work elsewhere. The game does not learn from you between sessions. Crews do not remember you next time. There is no adaptive difficulty quietly reading your performance. Radio context lasts for the conversation you are having and no longer.
The intelligence is in understanding what you said and simulating what follows — not in a system that studies you. And it remains a game and a training aid, not certified software, and not for operational use.
