[agubelu]

Happy you asked! Here are a few interesting cases that the respective ATCs probably hadn't seen before or will ever see again:

- Plane loses radio communication on final approach and ignores go-around orders, lands on occupied runway as ATC gets others out of the way [1]

- Guy steals a plane and takes off without permission [2]

- Someone left some cones on a runway at JFK [3]

- Door blows off aircraft [4]

- Some dude runs across the runways [5]

- Pregnant woman giving birth on flight [6]

- Student pilot freaks out, ATC calms her down and gets her instructor on frequency [7]

- Earthquake [8]

- Pilot hits a deer [9]

- Bomb threat on board, pilots decide to evacuate on the tarmac after no help arrives in ~1h [10]

Some situations have probably happened before, somewhere. Some others are completely new. I would highly recommend that you watch these videos (they're all relatively short) and genuinely ask yourself whether our current state-of-the-art AI models would be able to successfully handle these situations in the short timeframe required to do so. Let alone the fact that by AI we mean text models, so I'm not sure how they would integrate with terrain information, real-time radar data, arrival/departure routes, etc.

I think you have a fundamental misunderstanding of what control towers do. They're not there to simply observe planes do their thing and intervene if they get too close. They actively handle the traffic, and this task requires human pattern recognition and cross-domain reasoning skills in a matter of seconds, and the technology to replace this is simply not there yet. If you still disagree, I'd love to learn which technologies you'd apply to this problem and how they would compose with each other in order to achieve the same outcomes as in the cases I linked.

Of course, this is not to say that ATC shouldn't be made safer and more automated wherever possible. Particularly in the US, where equipment is severely outdated and some dubious regulations allow their ATCs to handle runways and give clearances in a way that would not be allowed in Europe and have already resulted in more than a few close calls. These are all valid concerns, but IMO they can't be extrapolated to "all ATC services can and should be automated".

I fly often, and I for one feel safer (at least, with our current technology) knowing that there are humans in the cockpit and in the control tower who can react and take control when unexpected stuff happens.

[1] https://www.youtube.com/watch?v=cXNWwKx9c1o

[2] https://www.youtube.com/watch?v=4LLmF9tZoEE

[3] https://www.youtube.com/watch?v=YmywjMQDbos

[4] https://www.youtube.com/watch?v=0ma0JzO43Ig

[5] https://www.youtube.com/watch?v=VZviKoEKAaw

[6] https://www.youtube.com/watch?v=5pZ3VOPlarw

[7] https://www.youtube.com/watch?v=dgc2Wh4cOgo

[8] https://www.youtube.com/watch?v=o80cNJ_XhX0

[9] https://www.youtube.com/watch?v=x04kRUIgXpQ

[10] https://www.youtube.com/watch?v=aAW2JbqxLRM

[CamperBob2]

Plane loses radio communication on final approach and ignores go-around orders, lands on occupied runway as ATC gets others out of the way [1]

Trivially handled by video game-grade AI. Is an aircraft somewhere it's not supposed to be, or doing something it's not supposed to be doing (most likely because a human has screwed up somewhere?) Route everything else out of the area, signal authorities. Human intervention not required.

Guy steals a plane and takes off without permission [2]

See above. What can ATC possibly do about this, besides alert other aircraft to the situation and signal law enforcement?

- Someone left some cones on a runway at JFK [3]

Foreign object detected or reported on runway, an everyday occurrence. Alert other aircraft and facilities personnel. What else is the controller going to do, go downstairs and pick up the cone?

Door blows off aircraft [4]

What is the controller supposed to do about this, other than accept the crew's request for an emergency landing and (guess what) route everybody else out of their way?

Some dude runs across the runways [5]

Like other 'surprises', two-legged FOD might have surprised the Wright brothers, but that's about it. Alert law enforcement and warn aircraft.

Pregnant woman giving birth on flight [6]

Just another routine 'surprise' covered by standard procedure. Handled primarily by the air crew rather than ATC. Handle emergency landing permission if requested by crew, alert medical personnel to meet the aircraft when it lands, done deal.

Student pilot freaks out, ATC calms her down and gets her instructor on frequency [7]

Honestly not sure what should be done to automate this kind of situation. I suppose an LLM could handle it as well as anyone else. :-P

See also the edge cases where the pilot(s) are incapacitated and an untrained civilian needs to be talked through the landing. How often does that happen in real life (and how often does it actually work?)

Earthquake [8]

Not really something ATC would be involved with except at the purely-tactical level. Ground outgoing traffic and reroute incoming flights until all-clear given by officials.

Pilot hits a deer [9]

Reroute traffic and alert medics.

Bomb threat on board, pilots decide to evacuate on the tarmac after no help arrives in ~1h [10]

Not an ATC issue except for the need to (guess what) reroute traffic.

They actively handle the traffic, and this task requires human pattern recognition and cross-domain reasoning skills in a matter of seconds, and the technology to replace this is simply not there yet. If you still disagree, I'd love to learn which technologies you'd apply to this problem and how they would compose with each other in order to achieve the same outcomes as in the cases I linked.

You could still have a valid overall point regarding the need for humans in the real-time loop, but I disagree that most of the scenarios you mention support that point. All of those cases can be (and are) anticipated, all of them have happened before except possibly [10], and all of them can be handled by computers at least as well as humans. Except possibly the situation with the pilot needing real-time psych support.

Let alone the fact that by AI we mean text models, so I'm not sure how they would integrate with terrain information, real-time radar data, arrival/departure routes, etc.

I definitely don't mean LLM-style text models. As I suggested in the first answer, nothing on this list except (again) #7 would have flummoxed a game programmer ten or twenty years ago. If some people want to try an ELIZA-like LLM to deal with #7, fine, but that would be a research problem. Nothing else on your list requires any new research.

[agubelu]

You have listed the correct actions to take in each scenario, but it's still not at all obvious to me how each case would be fully automated, start to finish. I would like more specifics, besides simply

  if(conesOnTheRunway) {
      closeRunway();
      rerouteAircraftOnGround();
      rerouteAircraftOnAir();
      pickUpCones();
      reopenRunway();
  }

Since I took the time to compile a list of examples that you sadly didn't find surprising enough, I would appreciate it if you returned the courtesy and provided a more concrete design of a fully automated ATC system using current technology.

> Trivially handled by video game-grade AI.

I think it's fairly likely I have played more flight simulator games than you have. If you know of a single one of them where the ATC AI isn't utterly stupid, please do let me know, I'd love to try it. There's a reason VATSIM exists and remains very popular :)

[CamperBob2]

First of all, to the extent controllers are "surprised" by any of those events, you're making my point for me.

Second, I wasn't referring to ATC AI in flight sims; as you suggest I have no experience with that. I doubt anyone ever put any serious effort into flight sim ATC AI compared with AI for realtime strategy or even FPS games. Years ago, a primitive neural net dragged a 9-dan Go master up and down the ban, so I imagine our current tech can handle a few planes.

I'm not paid to redesign the ATC system, or qualified to do so (although that won't stop Musk, I'm sure.) But as I've made clear, I believe it can and probably should be done. In practice it would probably look more like

    if (AircrewOrEmployeesReportForeignObjectOnRunway) 
    {
    closeRunwayAndRedirectTraffic();
    alertFacilityPersonnel();
    leaveRunWayClosedUntilSomeoneInAuthorityReopensIt();
    }

So many of the things you mention are simply not in ATC's wheelhouse to begin with. Their job is to keep things moving (or not) while other people in authority deal with those situations. That part wouldn't change, as I see it.

[agubelu]

Thanks for replying. I won't continue this conversation as it seems to me you don't actually have any specific evidences of your claim besides "I believe it can be done", and don't actually know much about the responsibilities and areas of authority of ATC.

[CamperBob2]

I'd say the burden of proof belongs with those who argue that it's impossible to automate ATC while maintaining at least the current safety record. That's an extraordinary claim at this point.

I asked for examples of situations that couldn't be handled through automation, you provided some that I consider invalid or inapplicable, and... well, there we are. GG

[pas]

I enjoyed the thread, and thanks for compiling the list, but I have to add, I haven't found them surprising either.

Automation doesn't have to mean "level 5+: ATC AI on, let's go find an extended happy hour"

I think it's very important to separate the software engineering (and systems engineering, and safety and process design, and other disciplines involved on the object-level) from the challenges at the meta-level (politics, legal liability - insurability, scaling and economics, and procurement issues, avoiding yet another too big to fail boondoggle, and so on).

One obvious problem is that by definition someone sitting there doing their shift has a very holistic view, and asking them what do they need to do their job better might not worth it economically. (The faster horse problem. Though sending a few enthusiastic designers there, also crunching the numbers of the past near-misses and other issues would likely reveal gaps in the current procedures and tools, and ... and of course this all then runs aground because changing procedures and tools is hard, hello FAA, etc.) But, but, of course doing the top-to-bottom design naively is almost a surefire way to burn a few quick billion bucks for nothing. (So, I think this should be something like an ongoing challenge, like the DARPA Grand Challenge for driverless cars.)