[0.914][AI] ai casting curse on units with zero defense

Bad dog, may thou' be cursed forever and ever, may your treats taste like dry grass and may eating your food feel like liking on trees, you shall forever be damned to never know where you are and may you have trouble walking!

Now I cursed your dog too!

Sincerely
~ Kongdej

Unfortunately, I don't think the AI actually knows what "curse" does. All it really knows is that it has a skill (perhaps of a certain type - in this case, debuff), with a certain priority attached to it. So if the opportunity presents itself, it'll use it (or uses the highest priority one). This is done because they want skills to be moddable. So you could add new skills with all kind of silly effects, and you'd still be able to get the AI to use it, despite having no real idea of what it does.

IMO, the system could use some kind of condition based priority. Like if enemy def is high, priority is high (in case of "curse"), etc...

That's pretty much how I imagined that battle proceeding

This problem applies more generally to buff/debuff spells.

In a previous game, I saw a champ waste actions casting protection from fire on himself and his minions while facing my Fire1/Air1/Water5/Death5 Ceresa. Sure, I could give them an attractive tan with my burning hands, but I also have no-brainers like graveseal, drain life, touch of entropy, not to mention a L13 ice elemental. The champ's time would have been spent much more productively hitting me with a stick. Besides, I only had 2 fire shards but 7 water and 7 death; I know this much about my enemy, so they should too.

Stoneskin is another one; AI loves casting it even when the damage output from the opposing side is heavily skewed towards blunt.

Sure, I can see how one could arrive at this, but the definition of the curse spell includes the modifier it applies to the target. At the most basic level, priorities aside, one could check whether the modifier would have any effect at all on the target. They are both just numbers.

EDIT:

I think a good way to think about this is in terms of opportunity cost of actions. Knowing only the damage type and modifier applied by a skill/spell, I can evaluate its maximum benefit versus opportunity cost of spending that action doing something else. For example, when deciding whether to cast protection from fire, which reduces fire damage by x%, I determine that at best I can hope to prevent m fire damage from my opponent out of my M total hitpoints, over t actions. On the other hand, I can deal n/N damage to my opponent right now by hitting them with a stick. If m/M is small compared to n/N, my time is much better spent sprinting towards them.

Similarly, I can calculate my opponent's potential damage output over the next t actions for each damage type. If they are capable of dealing 10x more blunt damage than piercing damage, spending an action protecting myself from piercing isn't a viable option - just defending might be better.

Exactly how prohibitive this becomes when considering the number of possible actions is left as an exercise for the devs; I count fewer than ten distinct types of damage in the game, so maybe it could be done for the direct damage dealing/absorbing spells.

Spells with indirect effects, such as slow or graveseal, are obviously trickier, including for the reasons Kalin mentioned.