The AI thread

[AEthelraedUnraed]

On 6/21/2026 at 3:37 PM, BraveSirRobin said:

I’m also a little surprised to see a programmer who thinks AI programming might be easy.  During my 28 years of programming I’ve never talked to a person who thought AI programming was easy under any circumstances.

To clarify, I do most definitely not think AI programming is easy in general.

Neither do I think fully modeling the human spotting behaviour and capabilities with all their ins and outs is easy.

I do think it is relatively easy to program a believable spotting AI. That is, an AI that behaves closely enough like a human to make its behaviour plausible to an external observer. Yes, there will be certain cases where a human would spot better or worse (e.g. camouflage matching the background). But no two humans spot alike either, we cannot view from the AI's cockpit, and we don't know where the AI pilot's "head" is directed towards. As long as the AI doesn't do too outrageous things (e.g. instantly spotting an enemy on your tail if he's 1% in the AI's viewport like now), I think it's perfectly acceptable if he occasionally misses an easy target or spots a hard one. It's what happens to us humans as well, after all.

I know that with modern neural nets, it is possible to reach and in increasingly many cases even surpass human performance in a specific task. I'm not suggesting those kind of nets are the best option here - they're very bulky and not geared towards gaming. But for relatively confined tasks, a tiny net might take just a few ms of runtime while still providing good generalisation and believable behaviour. We aren't there quite yet, but I think we'll see machine-learned AI incorporated into more and more games, including flight sims.

[AndreiTomescu]

[Avimimus]

12 hours ago, KC23 said:

Those are some great questions. Back in the day a very old scripted missions combat flight sim called, Chuck Yeager Air Combat had some pretty darn challenging AI to fight against. Mig Alley by Rowans was another. Of course don't talk to me about the fidelity of their flight models. No doubt IL-2 Korea will blow almost any sim out of the water (not to mention the graphics and damage model). 

The original Red Baron was pretty neat in this regard: Unlike the later Red Baron II/3d, the original Red Baron had different aces that would fight very differently. A lot of them were quite challenging.

Of course, the way they did this was changing the flight performance of the Aces aircraft (e.g. some of them could sustain tight turns). Basically, using the flight model to cheat.

[BraveSirRobin]

On 6/24/2026 at 5:58 PM, AEthelraedUnraed said:

I do think it is relatively easy to program a believable spotting AI. That is, an AI that behaves closely enough like a human to make its behaviour plausible to an external observer. 

Then why has no one managed to do it in a combat flight sim?

[AEthelraedUnraed]

7 hours ago, BraveSirRobin said:

Then why has no one managed to do it in a combat flight sim?

Like I already said, whether or not it has already been done doesn't mean it's hard. (And hasn't it been done? I don't know all combat flight sims.)

As for IL2, the current spotting conforms pretty well to what they stated their desired AI behaviour was. Still, I think it comes close; my only real beef with it is that it seems to look everywhere at once all the time. And that should be pretty easy to solve by introducing some directionally dependent spotting chance.

[Leifr]

Well, my early experiences with the AI aren't particularly glowing.

Fighting two MiGs (Ace) against a flight of four Thunderjets (myself included); wingmen don't appear to want to do anything positive, one decides to leave for landing, one MiG lawn-darts itself in to the terrain and the final MiG begins the eternal right-hand bank and turn with me, it didn't break this loop for several minutes. 🤐

[Jaegermeister]

9 hours ago, BraveSirRobin said:

Then why has no one managed to do it in a combat flight sim?

 

2 hours ago, AEthelraedUnraed said:

Like I already said, whether or not it has already been done doesn't mean it's hard. (And hasn't it been done? I don't know all combat flight sims.)

The part that is missing from your ongoing conversation is that good AI programmers working on autonomous flying combat logic would make way more money working for a government contractor in areas like the Replicator Initiative than they would with a civilian flight simulator. Not to mention all the programmers being occupied with military flight simulators which are replacing a large part of pilot training hours to save real world training costs.

It’s not whether it can be done or has been done, it’s who has it and will it filter into the game industry. I would think that sort of behavior model would be very valuable to certain companies right now. With the current race to dominate the escalating AI drone conflicts around the world, who knows?

 

[AEthelraedUnraed]

1 hour ago, Jaegermeister said:

The part that is missing from your ongoing conversation is that good AI programmers working on autonomous flying combat logic would make way more money working for a government contractor in areas like the Replicator Initiative than they would with a civilian flight simulator. Not to mention all the programmers being occupied with military flight simulators which are replacing a large part of pilot training hours to save real world training costs.

It’s not whether it can be done or has been done, it’s who has it and will it filter into the game industry. I would think that sort of behavior model would be very valuable to certain companies right now. With the current race to dominate the escalating AI drone conflicts around the world, who knows?

It's missing because it's another matter entirely 🙂 For autonomous flying, you don't want to mimic humans. You want to do the best you can, and in fact exceed human performance. A military spotting AI has no value to a historical combat flight sim (or indeed most of the gaming industry in general) and vice versa.

Secondly, don't worry, there's more than enough programmers left who don't go for the money 🙂 Speaking for myself, having a fun assignment and the self-fulfilment I get from my job are equally important. Also, I don't know about the situation in the US or Russia, but here in the Netherlands there aren't that many programming jobs in the defence sector. I know way more people working in the medical sector or semiconductors.

[Aapje]

5 hours ago, Jaegermeister said:

Not to mention all the programmers being occupied with military flight simulators which are replacing a large part of pilot training hours to save real world training costs.

Do military flight sims have AI? My impression is that they just tend to use real pilots, since they need to train a bunch of them anyway, and they would never be content with some generic AI stuff, but would instruct the Red Team to act out a specific scenario, or dogfight really well, in a way that a modern AI cannot beat (at least compared to a highly trained and very talented fighter pilot).

[LukeFF]

6 hours ago, Aapje said:

Do military flight sims have AI?

Oh, absolutely. There have been a lot of episodes about this for instance on the Fighter Pilot Podcast. Especially here in the US, where "live" range time is becoming more complex and difficult to arrange, robust flight sim trainers with realistic AI enemy behavior are in high demand. 

[Dash,Polder]

Where the boat is being missed is focusing resources on pushing a much greater ground war under AI, it would do tons more good there and it's made for that kind of limited movements and structured dynamic.

I have my doubts it'll ever be capable of thinking on a human level in a mono on mono fight, it doesn't think, it's just very fast at reacting to what it see's in a data set.

[BraveSirRobin]

15 hours ago, AEthelraedUnraed said:

Like I already said, whether or not it has already been done doesn't mean it's hard.

lol.  Yes, it actually does.  The idea that developers are not making obvious easy changes that everyone wants because ??? Is ridiculous.  And insulting to the developers.

[BraveSirRobin]

12 hours ago, Jaegermeister said:

The part that is missing from your ongoing conversation is that good AI programmers working on autonomous flying combat logic would make way more money working for a government contractor 

 

It’s not being missed at all.  According to @AEthelraedUnraed the needed AI changes should be easy to make (lol).  So “good” AI programmers are not necessary.

[BraveSirRobin]

15 hours ago, AEthelraedUnraed said:

 I think it comes close

It absolutely does not.  One of my favorite things is to watch an AI gunner sit down after an enemy aircraft moves out of sight.  I know the aircraft is still nearby.  The gunner should know it, because I’m yelling at him that the enemy aircraft is still close. But he’s just sitting there without a care in the world. Even fixing that without overloading the processor is probably relatively difficult. BTW, that just one example.  I’m sure there are many others.

 

15 hours ago, AEthelraedUnraed said:

my only real beef with it is that it seems to look everywhere at once all the time. And that should be pretty easy to solve by introducing some directionally dependent spotting chance.

I cannot stress this enough.  If it’s something that you think is easy, and it’s something that everyone wants, but it has not been implemented, then it almost certainly is not an easy change. 

[AEthelraedUnraed]

6 hours ago, BraveSirRobin said:

lol.  Yes, it actually does.  The idea that developers are not making obvious easy changes that everyone wants because ??? Is ridiculous.  And insulting to the developers.

We've already discussed this. It doesn't. There are tons of reasons why something easy can be very difficult to implement in certain situations. If you really haven't ever suffered from bad initial design choices and the "buzzwords" technical debt, spaghetti code and feature creep, then you truly are an extraordinary programmer. And that doesn't even begin to touch yet on the business and management side of things.

5 hours ago, BraveSirRobin said:

It absolutely does not.  One of my favorite things is to watch an AI gunner sit down after an enemy aircraft moves out of sight.  I know the aircraft is still nearby.  The gunner should know it, because I’m yelling at him that the enemy aircraft is still close. But he’s just sitting there without a care in the world. Even fixing that without overloading the processor is probably relatively difficult. BTW, that just one example.  I’m sure there are many others.

...which is still directional dependence.

5 hours ago, BraveSirRobin said:

I cannot stress this enough.  If it’s something that you think is easy, and it’s something that everyone wants, but it has not been implemented, then it almost certainly is not an easy change. 

We keep going in circles here, so I'm not going to again address the fact that even for easy things there can be reasons why they're hard or unprofitable to implement.

Also, I'm not sure better spotting is actually that high on the list of "what everyone wants" compared to other AI issues.

6 hours ago, BraveSirRobin said:

It’s not being missed at all.  According to @AEthelraedUnraed the needed AI changes should be easy to make (lol).  So “good” AI programmers are not necessary.

You're really that low that you start playing the man now? That comment was as unnecessary as it is dishonorable.

Since you're apparently no longer able of having a civilised discussion, this conversation is over.

[FodderMonkey]

21 hours ago, Jaegermeister said:

 

The part that is missing from your ongoing conversation is that good AI programmers working on autonomous flying combat logic would make way more money working for a government contractor in areas like the Replicator Initiative than they would with a civilian flight simulator. Not to mention all the programmers being occupied with military flight simulators which are replacing a large part of pilot training hours to save real world training costs.

It’s not whether it can be done or has been done, it’s who has it and will it filter into the game industry. I would think that sort of behavior model would be very valuable to certain companies right now. With the current race to dominate the escalating AI drone conflicts around the world, who knows?

 

AI developers working for government contractors are NOT what you want here.  An AI flying for a defense contractor always sees the enemy to the absolute limits of its sensor platforms.  It flies the aircraft to the very edge of its envelope; the AI is omniscient, lethal and unforgiving.  But that's not what you really want here, is it?  Of course not.  You want an AI that's "good", but not perfect.  You want a challenge, but you still want to win.  You just want to feel like you've earned that win.  A military-grade AI isn't ever going to give you that.  It's going to kill you 10 times out of 10, fighting dirty and exploiting every loophole, weakness and tweak it can find.  And that's what makes this so difficult.  Flawless is easier than, "really, really good but I can still beat it".  I'm confident the devs can do "flawless", but we're really asking them to find that golden standard of imperfection which, honestly, nobody has ever really been able to find yet.

[MajorMagee]

There are decent implementations of AI being done in sims today by small development teams.

Quote

Perhaps the most basic issue with GHPC's previous AI system was often summed up as: "They lack object permanence." Even backing down a hill for a few seconds was enough to make the enemy gunners give up trying to hit you and resume scanning the horizon like nothing had happened. This very obviously needed to be remedied at some point.

Now, the AI crews have full access to the radio system - the one that's used to send enemy location pings to your tactical map. The exact same communication network simulation is running for the enemy team, too, and now they have the ability to respond to what they know! Platoons will assign a portion of their members to watch for nearby reports, anticipating the detected enemy making an appearance soon. They'll keep watch until enough time has passed that they feel their attention is needed in their original assigned sectors, at which point they'll return to standard mode. But if they hear about a new, closer threat, they'll orient on that one instead.

 

Edited June 27 by MajorMagee

[Aapje]

@MajorMagee

The lack of object permanence is the consequence of an AI that has only two states for its relationship to the enemy, the enemy being known to be there and being tracked (as in: the pilot is looking at the enemy or has done so very recently), or the enemy being unknown and untracked. But a major factor in air combat is the third option, known to be there, but untracked.

With that lack of a third option, you indeed get the issue that the AI entirely forgets the enemy once it loses visual, which would result in the enemy completely stopping all maneuvering during a fight as soon as you fly into a blindspot, which would make for a horrible AI. A way to deal with that without introducing that third state is to give the AI the ability to track your specific location even when it can't see you, which means that it never does a bad maneuver based on an expectation of where the enemy is, although that doesn't match reality.

This is an example of how a design choice causes a lack of realism and how you can then either hack something in to make that less bad, but still unrealistic, or fix it properly, but with much more effort required.

Of course, there is a major limitation that the AI should perform well, so it is also not possible to add a huge amount of complexity.

[Xtremist]

4 hours ago, MajorMagee said:

There are decent implementations of AI being done in sims today by small development teams.

 

That's right! The new AI in GHPC is fantastic, very believable, and scales good with the difficulty. And the entire game is made by maybe 4 guys... 

[AKA_Relent]

Id be happy just to see a little more randomness in the AI behavior, especially when they are defensive.  Instead of just going into a semi-gentle left hand defensive turn, maybe at random points they can try to reverse the turn (to a right hand turn), or roll and dive/climb, which would give some impression of their trying to shake you from their six.

[BraveSirRobin]

13 hours ago, AEthelraedUnraed said:

We've already discussed this. It doesn't. 

Yes, it absolutely does.  The idea that there is something in the code preventing these changes is completely and utterly ridiculous. It's an extremely difficult thing to code.  That's why absolutely no one has done it very well.

[Panzerlang]

1 hour ago, BraveSirRobin said:

  That's why absolutely no one has done it very well.

Pure codswallop, there are a few that HAVE done it very well.

Personally, I've always found the IL2 series AI to be adequate. In as much as I fill its gaps with imagination (I'm an ace, the enemy all suck).

In MP, humans are aces and vets, AI are rookie filler. Kind of per reality (5% of pilots were aces iirc).

Edited June 28 by Panzerlang

[KC23]

Been playing for a while, and I really love this game. Super immersive, great lighting. Damage model for the ages. The ground units, trains. It's all pretty darn good.

[BraveSirRobin]

21 hours ago, Panzerlang said:

Pure codswallop, there are a few that HAVE done it very well.

Personally, I've always found the IL2 series AI to be adequate. In as much as I fill its gaps with imagination (I'm an ace, the enemy all suck).

In MP, humans are aces and vets, AI are rookie filler. Kind of per reality (5% of pilots were aces iirc).

Are you talking about the IL2 GB AI that everyone is currently complaining about or the IL2 1946 AI that blatantly cheated?

[Panzerlang]

2 hours ago, BraveSirRobin said:

Are you talking about the IL2 GB AI that everyone is currently complaining about or the IL2 1946 AI that blatantly cheated?

BoBII (Wings of Victory) and yes, 1946 too, regardless of it 'cheating'. 1946 is one of those cases where the destination is more important than the journey, aka The Player Experience. But I guess you're arguing objective over subjective.

1946, application of imagination: "These aces are able to eke out the last drops of performance; I'm not".

It's why they're called 'simulations', not 'reality'. They *simulate*. Yes, one can debate the quality of the simulation but there's a reasonable limit to that. AI and quantum PCs, that's when the ball game will enter a different arena.