Dev Diary - Ai is a “frame perfect button masher?”
ST4RL0RD08
Member Posts: 39 ★
Hey Kabam,
Thanks for the Dev Diary - it’s good to see some transparency around AI. But one thing still isn’t clear:
You mentioned the current AI is a “frame-perfect button masher,” but something I’ve not heard anyone talk about is what that means from a speed perspective…
In laymen’s terms, it means it reacts every 16.7ms at 60fps. That’s INSANELY fast. Even at the old 30fps, it was still reacting every 33.3ms, which is still far faster than any human can respond. For context, a pro player reacts around 150–200ms. So the AI currently reacts 9-12x faster than a Pro Player?! In what world is that remotely fair or a good player experience?! When you lay just that one element out, no wonder we are all feeling like the Ai is completely cracked beyond belief!
That brings me to the main question:
Is AI 2.0 still going to react frame-perfectly?
Or are you actually introducing input delay, reaction buffers, or human-like limits?
Because no matter how “smart” or “strategic” the new AI is, if it can still react inhumanly fast, it’s always going to feel unfair, not fun. Blocking specials mid-combo, intercepting with frame-perfect accuracy, and holding onto specials doesn’t feel like difficulty, it feels like fighting the engine.
I mean, most successful and competitive games (especially in the fighting genre) implement human-like limitations in their AI to keep things fair. Games like Tekken, Mortal Kombat, Street Fighter, and even God of War (in its boss design) introduce intentional input delay, reaction windows, or AI behavior patterns to simulate real opponents and give players meaningful counterplay. These games know that challenge should come from mechanics and mastery, not inhuman reaction speed. That’s what keeps fights skill-based, satisfying, and rewarding - not oppressive or cheap.
Would really appreciate a clear answer on this.
Thanks.
Thanks for the Dev Diary - it’s good to see some transparency around AI. But one thing still isn’t clear:
You mentioned the current AI is a “frame-perfect button masher,” but something I’ve not heard anyone talk about is what that means from a speed perspective…
In laymen’s terms, it means it reacts every 16.7ms at 60fps. That’s INSANELY fast. Even at the old 30fps, it was still reacting every 33.3ms, which is still far faster than any human can respond. For context, a pro player reacts around 150–200ms. So the AI currently reacts 9-12x faster than a Pro Player?! In what world is that remotely fair or a good player experience?! When you lay just that one element out, no wonder we are all feeling like the Ai is completely cracked beyond belief!
That brings me to the main question:
Is AI 2.0 still going to react frame-perfectly?
Or are you actually introducing input delay, reaction buffers, or human-like limits?
Because no matter how “smart” or “strategic” the new AI is, if it can still react inhumanly fast, it’s always going to feel unfair, not fun. Blocking specials mid-combo, intercepting with frame-perfect accuracy, and holding onto specials doesn’t feel like difficulty, it feels like fighting the engine.
I mean, most successful and competitive games (especially in the fighting genre) implement human-like limitations in their AI to keep things fair. Games like Tekken, Mortal Kombat, Street Fighter, and even God of War (in its boss design) introduce intentional input delay, reaction windows, or AI behavior patterns to simulate real opponents and give players meaningful counterplay. These games know that challenge should come from mechanics and mastery, not inhuman reaction speed. That’s what keeps fights skill-based, satisfying, and rewarding - not oppressive or cheap.
Would really appreciate a clear answer on this.
Thanks.
19
Comments
Besides, there’s rarely any time you’d have to contend with the AI’s reaction time. The only times it is relevant is when you get light or special intercepted. Playing defensively will eliminate those both (not saying that they’re okay though, they gotta go)
Other huge fighting games like Street Fighter, Mortal Kombat, God of War, Elden Ring, Tekken, Injustice Mobile etc, all build in delays or patterns to make the AI feel beatable and fair. MCOC doesn’t. It’s literally the only fighting game I can think of that allows the ai to react “frame perfectly.” The AI doesn’t make mistakes, it doesn’t hesitate, and it can block, punish or change its decision mid-animation without warning. It shouldn’t be able to do that, and that is clearly the reason for all the bs to do with the ai reaction speed. (The whole “heavy spam” or passive stuff is a separate issue — this is purely about reaction speed.)
So when the game forces you to be aggressive because of a degen timers, unblockables etc etc, and you’re up against that kind of AI, it just feels rigged. It’s not about difficulty, it’s about it being actually fair like basically every other competitive fighting out there…
So that’s what I’m trying to query: is the ai 2.0 going to actually have its reaction time modified to be comparable to a human, or is it going to be the same cracked BS but just rolled in a bit of glitter? That’s all I’m asking…
AI will always react faster than you. It’s not human, in most fighting games it can literally read the frame data of the attack you’re using and respond with something faster. If you ever beat an AI in any fighting game, it’s because it let you beat them. This game is no different, it’s not a game thing but an AI thing. Computers have extremely fast reaction speed, that’s just the way they work.
The problem isn’t that MCOC’s AI “never makes mistakes,” it’s that when it does respond, it does so faster than humanly possible, and without any animation commitment. That’s the real issue.
If we’re being forced to play aggressively with degen timers, unblockables, etc., then yeah — the AI needs to play fair too. Otherwise it’s just frustrating, not challenging.
That’s why I’m asking if AI 2.0 is going to address reaction speed, not just behavior. Becahse that’s what would actually change how the game feels.
The AI IS playing fair, to a degree. They purposefully code bad decisions into the AI to give players a chance. That’s how the playing field is leveled. Your comment about how the AI replies to moves extremely quickly is applicable to all fighting game AI’s. I’ve been in situations where the AI just chooses to block my attacks sometimes and not block them other times. I’ve been in situations where a very frame heavy move lands sometimes and other times the AI responds with something faster and I eat a full combo. This isn’t any different between MCOC and other fighting games.
The best argument that you can make, from my perspective, would be saying that actions in MCOC have fast frames, making most moves near impossible to react to as a player but trivial for an AI. But the thing about that argument is that it’s not an AI issue, it has nothing to do with the AI’s reaction speed. It’s an issue with the fundamentals of the game itself and how fast the game’s moves are. Adding delays into the AI’s moves wouldn't make it respond slower, it would just make it respond less frequently. Which means you would get situations where the AI just stares at you and lets you hit them, which is what the “easy” AI in the game is already doing.
Kabam literally said in their dev diary that the AI is a “frame-perfect button masher.” At 60fps, that means the AI can act or change its decision every 16.7 milliseconds. That’s 10–12x faster than a pro fighting game player, and closer to 15x faster than most average players. It doesn’t need to read inputs, it doesn’t need to cheat. If it sees a window (even just a single frame) it reacts. That alone breaks the concept of fairness in any fighting game.
And no, that’s not the standard. Go play literally any other major fighting game. Street Fighter, Tekken, Mortal Kombat, Injustice Mobile, even Souls-like bosses like Elden Ring or God of War. Their AIs don’t respond with cracked reflexes. They’re coded with reaction delays, pattern reads, or limited inputs — specifically to keep things fair and beatable. Why? Because AI that behaves like a robot isn’t hard — it’s frustrating. It turns timing, prediction, spacing — the fundamentals of fighting games — into pure RNG.
That’s why we constantly see:
• Specials getting blocked mid-animation after a full combo
• The AI breaking out mid-special and instantly punishing you
• Dash-ins and MLLM combos no longer being reliable because a single long frame gives the AI enough time to react
• AI parrying or intercepting in windows that are physically impossible for a human
• Defensive recovery into a punish before we’ve even finished a hit
This is literally why players say “you can’t end with a medium anymore” — the AI has enough reaction speed to block the special that follows it. That’s a reaction time issue. Same with combo chains or dash timings getting punished by backdraft intercepts. This stuff doesn’t happen because we messed up. It happens because we’re not fighting something humanly fair.
And no, it’s not just animation speed — it’s about how fast the AI is allowed to react inside those animations. You never see a God of War boss suddenly block mid-animation or parry mid-combo do you? Why? Because they’re coded to feel fair, not perfect.
So again, the question is simple:
Is AI 2.0 still going to have that frame-perfect 16.7ms reaction speed baked in? Or are we finally going to see the AI capped to a human-level reaction window, like literally every other modern fighting game already does?
Because if not, all the behaviour polish and glitter in the world won’t matter. It’ll still feel rigged — and players will still burn out on it.
I have to wonder if I’m November when 60fps was introduced the AI was limited to decisions at a 30fps rate which is why that month felt fairer than normal. I dunno, this is the stuff that has me in a wait and see.
The problem is the current AI system has no memory. It can only react to what it sees. So by definition, it can only react instantly. It cannot do something now based on what it saw a half second ago. That's what makes the AI so "twitchy."
However, the AI doesn't "react" to what you are doing at all, at least not in the way you probably think. The game doesn't know what you're doing, because it doesn't anticipate anything. It doesn't remember the past, and it also doesn't predict the future. It sees what's happening now, and acts on what it sees happening now. That can *sometimes* seem like reaction, for example if it randomly decides to throw an attack the AI can validate that at the moment it is out of range to throw that attack, and the AI never chooses to "whiff" so that random decision gets thrown away. Later, when you move into range the same random decision now works, and the attack happens. It can look like it is reacting to you moving toward it, but in fact it is not reacting to you moving forward, it is acting at the first moment such an action is possible.
It is the combination of the action state tables that define which options the original AI designers programmed the AI to choose from as action options for each AI state, combined with the current set of environment inputs the AI can observe, combined with an action validator that decides if an action is valid to attempt at that instant in time that combine to form what the AI exhibits as behaviors. It does this because that's all it can do. AI 2.0 is intended to broaden the nature of behaviors the designers can program the AI to follow, and a lot of that involves giving the AI a better understanding of time - how combat flows, how to regulate the pace the AI follows, and other things necessary if you want to tune the AI to be more or less "good" at fighting.
So sure, maybe the AI is twitchy or basic, but its reaction time is robotically fast — and that’s the problem.
It’s why we see things like:
• AI breaking out mid-special animation (e.g. Spiral’s SP1)
• Specials being blocked mid-combo, even when the first hit landed
• MLLM into special chains failing and being blocked, so having to resort to MLLL with certain champs to land a special because the M animation took too long and the AI was able to block in a single frame
• Relics somehow getting blocked on hit two or three
• Instant punishes the moment you miss timing by a single frame
Even if the AI isn’t “smart,” the fact that it reacts frame-perfectly at 60fps means it still:
• Blocks or dodges mid-input
• Exploits and/or punishes tiny/single frame animation gaps with inhuman precision
• Can seem to cancel out of animations after they appear to have committed to one (e.g. a dash forward becoming a backdraft intercept)
That’s not about AI intelligence, it’s about reflexes.
And that is an AI issue. A big one. Because no other fighting game allows that.
Games like Street Fighter, Tekken, God of War, Elden Ring, and Injustice all add artificial human-like delays, pattern reads, or animation commitments to make AI feel fair. They limit reaction time on purpose. MCOC doesn’t.
All I’m saying is that right now, the AI’s frame-perfect reaction time allows it to do things no human ever could. That’s the problem. I’m not asking for an advantage, just a level playing field. Just like in other games, the AI shouldn’t be able to pull off actions that we, as players, physically can’t.
So my query/potential concern is, if AI 2.0 still runs on frame-perfect decisions, then all the cool new “behavior trees” won’t matter. It’ll still feel busted.
The issues you mention above with Spiral, and specials and other stuff have nothing to do with reaction time. We can do it too. Changing AI reaction time will not fix this.
90% of the problems you listed have more of an issue to deal with hit stun as opposed to anything else. Hit stun is the time after you hit the AI where the AI can’t make any inputs, it’s what allows combos to happen in the first place. If an AI is blocking you mid combo, or dodging special attack hits after the first few have already landed, the problem lies in the fact that the hit stun of the previous attack was too short and needs to be fixed. Yes, the AI is able to exploit those design flaws with their reaction speed, but the original problem isn’t due to their ability to find openings quickly, it’s the fact that the design flaw exists in the first place.
In regards to the points you brought up about AI canceling out of animations, like turning a dash attack into a backdraft intercept, do you have any clips to show this? I’m not saying you’re lying, I’ve just never seen it so I’m trying to visualize it in my head and I’m not really sure what that would look like.
Totally agree that hitstun, frame windows, and animation commitment are all part of the equation. You’re right to point that out. But the key thing here is: the AI is only able to exploit those flaws because it reacts faster than a human ever could.
If there’s a 1-frame gap between attacks, a human wouldn’t even see it — but MCOC’s AI can, because it validates and executes actions every 16.7ms. That’s why we see:
• Specials being blocked mid-combo
• Medium finishers into specials getting blocked/punished
• Relics being somehow blocked on hit 2 or 3
• The AI being able to parry us way more since the 60fps launch (although it happened a lot at 30fps too — but it’s way more prevalent now)
• Or the infamous “backdraft intercept” animation cancel (which I’ve linked below from Vega’s breakdown,) although there are loads of examples on streams & vids from Trappy, Lefty, KT1 etc too.
Other games, like Tekken, Injustice etc (examples above,) have those same frame vulnerabilities — but the AI can’t abuse them, because it’s intentionally coded to react like a human. That’s the difference.
So yeah, MCOC has hitstun quirks — but without the AI’s cracked speed, most of them would go unnoticed. That’s why it is an AI issue.
Here’s Vega’s video:
Vega fully breaks down where Serpent cancels a dash forward into a backdraft intercept — mid-animation. That’s the kind of stuff you’d never see in Tekken or Street Fighter, because their AI can’t react that fast.
Also just tagging @average_desi — just to clarify: we actually can’t do that.
Players deal with input delay (even minor ones on high-end phones), and our average reaction time is ~180ms. The AI gets to act every 16.7ms with zero delay, fatigue, or hesitation. Even with the “coded mistakes” from Kabam it’s still reacting 12-15x faster than we can.
So when it pulls off a frame-perfect punish, blocks mid-special, or cancels into a punish, it’s doing something no human ever could.
This isn’t about how “smart” the AI is — it’s about the speed at which it’s allowed to make decisions/respond.
Here's a video of me backing out after getting by a special attack.
Even if there's no ability to "roll the right decision" it only takes a few frames to roll the right decision. A simple example is how many times the AI is special intercepting now. It used to be much safer to tap block bait a special; it's basically something you cannot do any more and this changed in the last 6 months.
OP, see DM for advise in keeping this thread open.
Here are a couple of videos to show what I’m talking about, in case I’m completely wrong and Vega’s situation was completely different:
With the existence of Parry, light intercepting, medium intercepting, corner dexing, and backdraft intercepting, dash attacks are actually one of the worst possible moves you can make in this game purely because of how unsafe they are and how easy it is for your opponent to get an opening off of one. Think about how difficult it would be to get an opening against the AI if it never dashed at you ever. If the AI continues to have the ability to counter dash attacks in any way other than blocking, the game’s going to get to a point where dash attacking is just a non-option, which is extremely unfun and bad design overall. All that being said, this was a very long-winded way of saying that I agree with your points that the AI should absolutely not be able to do these things. The AI isn’t breaking the laws of the game in any way by doing them, players could and have done this too, I just don’t think it’s fair for the AI to be able to do it as well, because if it could then realistically what are we supposed to do against that?
Should the AI be tuned to human reaction time? Absolutely not. It's supposed to be faster. In fact, I'd argue that a few minor tweaks aside, the AI is really good. In almost every situation it's clear, to me, that I made a mistake. The AI special intercepting isn't a problem either because that's something the players do all the time.
The issues with the AI are:
1. because it can make frame perfect rolls, you don't want it to take actions at spacing that a good player cannot react to. If the AI is able to walk forward into light range then immediately throw a light, you're not blocking that. The fix for that is to stop the AI from rolling a light when it enters what @KabamDORK used to call "action distance". This is what they appear to have changed and what needs to be fixed. The AI being able to throw a light so close to the player is just not fair (especially since it can hit-confirm into a combo). The AI is even going for lights on standup when you miss-time a dash in now...
2. Baiting specials is a real nightmare especially when XL champs are involved.
The only option is to dash backwards, get them to light attack and then punish them with a counter attack. You can't beat the reaction time and get a hit in first anymore....
It is also "frame accurate" in a manner similar to MCOC. And I do not believe it directly incorporates human reaction time. Rather, just like with MCOC, it executes its action scripts with some randomness, making it seem like it isn't seeing and doing everything instantly. Which is not to say there aren't fighting games that explicitly incorporate human reaction times. However, this is not trivial to figure out just by observing it, because random state engines can behave similarly.
Unfortunately, the technical details of most fighting game AI is a bit of trade secret magic, so you don't see too many public descriptions of specifics (which is why I had to find one for a game released during the Triassic period), but the basic technology comes in mostly just a few flavors. behavior trees and scripts, state engines, and neural networks (and sometimes hybrids). SF2 is more the first kind, MCOC is I believe the second kind, and the third kind is out there but uncommon. And to be honest, most of the interesting stuff I've seen recently involving human reaction time hasn't been in the AI side of the fence, but rather on the PvP side of the fence. Because as weird as it sounds, games where all the opponents are humans still need to consider how to deal with human reaction times, because most real time shooters use predictive logic to account for network lag.
I think what’s been really valuable is that even in some of the technical breakdowns or semi-counterpoints, we’re actually all circling the same conclusion: whether it’s hitstun, spacing, or decision frequency — the AI’s ability to act every 16.7ms is what enables those quirks to become actual problems.
That’s the heart of it. The AI isn’t “cheating,” but it’s so fast that it’s able to exploit every frame of vulnerability in a way no human ever could — and that’s why it feels off.
Love the callout from @HungaryHippo and @Toyota_2015 about the “action distance” stuff — that’s bang on. The second an AI steps into range, it can throw a light in the same frame, and unless you’re already blocking, there’s nothing you can do. That’s not skill-checking — it’s speed-checking.
I also really appreciate DNA’s insight into how other games appear to simulate human reaction without always technically limiting frames — that was super useful context. But again, the takeaway from that still stands: other games are deliberately built to feel fair and simulate human reactions — MCOC isn’t. Yet.
And that’s really all I’m hoping AI 2.0 addresses. No need to dumb things down or make it passive — just build in some human-like constraints so timing, spacing, and reads actually matter again. Because right now, it’s too easy to feel like you’re being punished for doing the “right” thing… just one frame too slow.
Thanks again for keeping this civil and insightful. Genuinely a great discussion.
Watch night thrasher and silver centurion fights at the start. AI can hit and dash back at the same time
Of course almost none of us are experts, but you compare the "easy" early game ai with the late game ai, the "harder" ai is much more enjoyable to play cause its actually reactive, the early game ai is definitely slower, with champs just standing with their block down, u can heavy them with no repercussions, but the result just infinitely slower gameplay that borderlines on boring. So im not sure slowly the ai down is really the solution here.
im pretty sure that even if the ai could remember earlier frames, but make decisions only a "human" level", if it was coded to just win, we'd never be able to touch the opponent. just the existence of being able to parry and intercept highlights that the ai is not a perfect player.
>Die
>Ai gets rooted
>6 gorillion heavy attacks in 5 seconds
The point to AI 2.0 is not to make the AI slower, and it won't. The point - which Kabam itself addressed in various places - is to give the AI system more tools to mimic human-like behaviors. This requires giving it features and tools it does not currently have, for example the current AI system has no memory (beyond the implicit memory of states). It cannot act on what it saw, only what it sees. It has no ability to predict, so it cannot strategize, so it cannot ever employ a faulty strategy. And it doesn't seem to contain a strong notion of scripted actions - the idea that the reason you do X is because you intend to do XYZ. It seems to do that sometimes (it can complete full combos, for example), but not at others.
The point to AI 2.0 is actually counter-intuitive: it is to make the AI smarter. Because only if the AI is smarter can players outsmart it. You can't outsmart dice. Specifically, if the AI was smarter, the devs could then tune how smart it was.
Human reaction time is on the order of 100ms, about a tenth of a second. It doesn't matter if the AI reacts ten times per second or twenty or thirty or sixty or a thousand times per second. Once it is faster than us, it is faster than us. There's no benefit to be had from being even faster in terms of outperforming us.
Incidentally, it occurs to me that you might be conflating how fast the AI is, and how often it makes decisions. The AI can get more or less twitchy by reducing how often the AI chooses to take action. But if we "slow" the AI down and force it to make decisions only, say, every tenth of a second instead of thirty times per second, it will still be "frame accurate" and still be acting instantaneously. It will still wake up ten times per second, see what's happening right now, and then act instantly to react to what's happening now. How fast the AI can react, and how often it chooses to make decisions, are two different things that affect the behavior of the AI differently.
Kabam literally stated the AI is a “frame-perfect button masher” — meaning it can make and validate decisions every 16.7ms at 60fps. That’s 10–12x faster than a pro player, and up to 15x faster than your average player. It doesn’t predict, it doesn’t anticipate, it just sees a gap — even a 1-frame gap — and acts instantly. That’s not intelligent. That’s just cracked.
And this one design choice explains so much of what feels off in high-end content:
• Specials getting blocked mid-combo
• MLLM into special chains being blocked while MLLL works
• Dash-ins being intercepted before your animation even completes
• Relics being blocked after landing 1–2 hits
• Parry “fails” that probably aren’t actually input drops — just the AI reacting a frame faster than us
The issue isn’t about how “good” the AI is. It’s that it’s allowed to act faster than any human ever could. That’s not something you can outplay. That’s not about spacing or memory. That’s a system-level imbalance.
And to be blunt, as I said before, no other modern fighting game does this. Street Fighter, Mortal Kombat, Injustice Mobile, Tekken, even God of War and Elden Ring — they all build in human-like reaction caps, or buffer windows, or intentional AI delays. Not to make things easy — but to make them fair. Kabam’s AI doesn’t have that. Never has. And with 60fps, it’s now reacting twice as fast as it used to.
DNA, your point about smarter AI is spot on. Smarter is better. But smarter + frame-perfect = even worse. Because now you’ve got an AI that not only reacts faster than us, but makes smarter decisions while doing it. That’s a disaster waiting to happen.
And let’s not forget — Kabam said in their dev diary that they don’t want to change that “frame-perfect button masher” mechanic. That’s what scares me. Because it means that no matter how smart AI 2.0 is, it’ll still be reacting at 16.7ms. That’s the bit we, as a community, need to question right now — before it’s fully built, implemented and rolled out. Because once it’s done and in the context, it could be another 10 years before it’s changed again…
IF it’s ever changed again.
So this isn’t about behavior or difficulty. It’s about speed. And the fact that the AI operates on a system that no other competitive game allows. That’s not something we can adapt around forever.
Thanks again for the thoughtful responses. Even where we disagree, I think we all want the same thing — a game that challenges us, but doesn’t cheat us.