Dragon Quest 4 on the Famicom (and its English version Dragon Warrior 4) do this with the ally AI.

In Chapter 5 (the final chapter, and by far the longest), you control the legendary hero (whose name and gender you choose at the start of the game), and while the main characters of the previous chapters join, you don't get to control them directly. Rather, they are controlled by an AI; you set a tactic for the AI (like Offensive or Use No MP), and the AI will act according to that setting. (For those familiar with later DQ games, note that there's no "Follow Orders" tactic in this game, a decision that is quite controversial among players.)

Here's the interesting part: The AI *learns*. For example, when fighting a boss, Cristo (Kyril in modern translations) may attempt to cast Beat (Whack) on a boss, which isn't going to work because bosses are, of course, immune to instant death. After a few tries, however, he will stop trying, and I'm pretty sure this knowledge is retained even in the event of a party wipe. (So, store your money in the bank (so that you don't lose half of it when you die), attempt to fight the boss, lose, don't reset, then try again, and Cristo should behave better on the next attempt. There's even a "Try Out" tactic that's meant to get the AI to learn. (I was able to train the AI to cast BeDragon on Metal Babbles (Liquid Metal Slimes in modern translations) by having the hero Transform into Mara (Maya) and then cast that spell; it only does 1 damage, but consistent 1 damage on metal slime type foes is actually quite good, as any DQ fan would attest.)

(Note that this post does not apply to the remakes of Dragon Quest 4, which use the AI setup of later games, including the "Follow Orders" tactic and DQ7's option of setting separate tactics for each character.) Yes.

An AI algorithm might work like this: For each possible move, give the move a score, based on some heuristics. An action that produces a more favorable result (to the AI), like finishing off an enemy unit, is given a higher score. Optimally, the AI would choose the best scoring action, and perform that.

Now that we have this algorithm, we can scale the intelligence down in a simple way; don't just choose the best scoring action. Instead, only choose between the top X% of the available actions. You now have a knob that you can tweak, allowing the AI to be dumb at first and smarter once you've killed most of the enemies.

Puzzle Quest's difficulty setting works like this. On Hard, the AI always chooses the best move; on Normal it chooses between the top 10% (which isn't much different from Hard; they should have made it 25%), and on Easy it chooses from the top 50%.

Bastet (Tetris clone that gives you the worst pieces possible) has a scoring system like this when it chooses which piece to give you. One could tweak this algorithm to make a Tetris clone where, instead of speeding up, the game gives you worse pieces as you reach higher levels.

I haven't gotten around to Dragon Quest yet. I probably would not have noticed the AI's behaviour. That was very informative, thanks!

I remember a phrase I heard a developer use - "artificial intelligence is easy. artificial stupidity, on the other hand..."

Since that AI controls your party, and it's mandatory in the Famicom/NES version, you likely would notice its behavior over time.

There's one thing I forgot: The ally AI in DQ4 (and all other DQ games with ally AI) cheats. Specifically, while you choose your actions at the start of the round, the AI does not. I know this because I've seen instances where a character would be killed by the enemy, successfully revived (with the spell that only works 50% of the time and only restores 50% HP), and then healed; the player would not have known that that would happen.

Prey's typhon are supposed to be similar, with a hive mind. Unfortunately, while they can adapt, it's nothing like what it should be.

I've never played it, but I remember seeing the trailer for Echo:

Here's an article
[url=https://arstechnica.com/gaming/2017/08/echo-hitman-preview/#:~:text=Unsurprisingly%2C%20given%20Ultra%20Ultra's%20staff,like%20playing%20chess%20against%20yourself]Link[/url].

Here's a trailer
Link

If you're thinking of something like that then the chances are it could have been done 10 or 20 years ago.

Examples of tech like that would be speech recognition or VR.

Proper physics would be another one.

Now that I think about it sneaky developers already copy players actions to add into their bots A.I and there was dynamic difficulty with A.I players in games from the past.

It's possible to do this to some degree.

Max Payne (the original) did this a bit. The systems learned you play style and made the AI stupid to your preferred style (which is what AI should do -- make the game more fun for you). If you liked to camp and kill, the AI would rush you. If you preferred they hide so you have to blitz them, they would. Again, to some degree.

Precisely what you're asking for can be pretty challenging. AI can take a lot of processing, memory and, it still only learns what it's told to learn unless it has lots (AND I MEAN LOTS) of samples. How many geth do you kill over the course of a game? With current AI techniques, it'd have to be thousands to make any meaningful impact. It's easier probably to code it specifically, a la Max Payne or just some numbers inflation.