Could AI run kingdom building scenarios?

[Oli Ironbar]

Not too familiar with what different AI can and can't do, but assuming they can put together and maintain all of the rules for kingdom building consistently, can it run a kingdom?

We finished kingmaker a year ago, but I really enjoyed the free form building system. It would be fun if AI could track a kingdom and invent new challenges each turn.

[Azothath]

AI is just a trained neural net and expert system. It has to be trained on a data set and doesn't use symbolic logic. It reproduces responses to that data set. The further you go from the training set the less accurate the response.
AI is most useful when you can't determine a symbolic logic transfer function, elucidate a deterministic or procedural path, the problem cannot be well defined in a mathematical manner, or the computational load is burdensome. Sometimes probabilistic logic(fuzzy logic) is all you need.

You can code a PF1 GM that responds to player input. WoW has been doing it for years from a limited set of commands/actions. Paizo did try but failed. You also need a "no op" loop & timer so the game can continue without a player for a limited amount of time.
It doesn't need AI but the interpretation of natural language into sensible keyword actions is the tricky part.
Limiting actions to defined actions is going to frustrate some players.
PF2 being a better defined more cohesive system is easier to simulate using procedural code.

I'll go out on a branch and WAG you could do it in a spreadsheet. lol... the next step is a simple java or C++/# program.

To create challenges you'd have to define 2-3 sets that combine to make a challenge (one-armed bandit style) verified using a CR calculator. Then you are just at a permutation of those data sets.

[Andostre]

Azothath, I genuinely can't tell if your answer is Yes, No, Maybe, or I don't know. You've used too much technical jargon for someone not knowledgeable about AI to know what you're talking about.

(Speaking for myself, not trying to speak for Oli Ironbar.)

[Azothath]

it is a technical topic. I do span math, computer science, and engineereng topics. Like Lovecraft, you may need wikipedia as I am condensing models/ideas into their technical terms.

So sure, it is doable but not worth the time, resources, and effort to produce and then the resources to produce each new answer/evaluation... there are more practical ways.

AI training sets need to be huge. If you can explicitly define all of the permutations to create that data set Then you don't need AI. The answers can be found using a simple brute force solution. With PF1 RAW it is possible to do (and more accurate) and you don't need AI at that point.

PF1 is procedural (opposed to logical, mathmatical, or formulaic). Do this, then this, If Y is type B then multiply by 1.5, If Z is like X and first then 0.75*X else if >first 0.5X, then sum. So it lends itself to standard computer coding.
IF RAW was uniform and consistent it would be easy, but it's not. There are contradictions and undefined parts. where it is undefined the easy answer is to return an error for 'No'.

[Oli Ironbar]

Azothath wrote:

IF RAW was uniform and consistent it would be easy, but it's not. There are contradictions and undefined parts. where it is undefined the easy answer is to return an error for 'No'.

Are there contradictions in the kingdom building rules?

[Azothath]

Kingdoms and War - title topic
Kingdom Building - sub topic
An implementation after a cursory review would be a multipage spreadsheet with Phases as decending sheets/pages and a data set sheet. You work through the sheets to the end phase and end of turn. Export a data set for the next round (traceability), open a new spreadsheet and import the data. Alternatively you could have a function to replace the initial data set and increment a turn counter, hit recalculate and start turn 2.

So try that in an incremental build. It doesn't need to be perfect just workable and mostly correct. You could condense it to a few sheets but get the separated version working first as you can test each sheet and the data interaction.

I'm sure there are inconsistencies as that's the norm. Putting defined ranges & sets in the data set sheet will reveal those issues. Hexes [0,1,2,...,32, ..., (map size?)]. Simply getting everything defined and down on the data sheet will be the major task. I think the calculations are simplistic. You could implement those calculations & procedures via several methods; spreadsheet, webpage code, java...

[Oli Ironbar]

Azothath wrote:

Simply getting everything defied and down on the data sheet will be the major task. I think the calculations are simplistic.

Do you know which AI can use one of the spreadsheets out there already?

[Azothath]

Oli Ironbar wrote:

Azothath wrote:

Simply getting everything defied and down on the data sheet will be the major task. I think the calculations are simplistic.

Do you know which AI can use one of the spreadsheets out there already?

umm, no.

I think you are using magical thinking.

[Oli Ironbar]

Azothath wrote:

Oli Ironbar wrote:

Azothath wrote:

Simply getting everything defied and down on the data sheet will be the major task. I think the calculations are simplistic.

Do you know which AI can use one of the spreadsheets out there already?

umm, no.

I think you are using magical thinking.

You should lead with that next time.

[Azothath]

some engineering might be better than wishcrafting, or $300000 down with 2 million for development.

[Mysterious Stranger]

AI is only a tool. For what you want you would probably need to build a custom program. You will not be able to simply open an AI ap and have it do what you want.

If you just wanted it to come up with some random ideas or to assist you in creating a campaign you could probably use something like copilot to help you. I have a Copilot + PC that has an AI chip and can run the AI software locally instead of over the internet. I have been using it to help me create a Fantasy Hero game located in 6th century Brittan and it has been very useful. But it cannot run the game. For research and creating background, it is actually very useful.

[Oli Ironbar]

That’s pretty cool.

So yeah, kingdom building has been called a spreadsheet simulator, so if I have the spreadsheet, could AI come up with random challenges and pop them out in text turn by turn?

With training, could it study the spreadsheet and intuit parts about my kingdom? (We have a lot of libraries but no granaries, we only have one army but three cities, etc. )

[Claxon]

Oli Ironbar wrote:

That’s pretty cool.

So yeah, kingdom building has been called a spreadsheet simulator, so if I have the spreadsheet, could AI come up with random challenges and pop them out in text turn by turn?

With training, could it study the spreadsheet and intuit parts about my kingdom? (We have a lot of libraries but no granaries, we only have one army but three cities, etc. )

I don't think you're ever going to get it to actively participate in the spreadsheet, or even have a "turn by turn" thing going. But you can definitely go to it and give it a prompt to provide a challenge.

You would then have to translate what that challenge means in terms of the mechanical kingdom building rules, because the AI isn't really going to understand that. Or if you can get it to provide something, the narrative it provides probably isn't going to match.

Ultimately what this thread really highlight is how AI is being sold to people at large for what it can do versus what AI can actually do. It is actually very good at a very small subset of things people think it is good for. (Personal experience it's good at taking notes in meetings and summarizing long texts and providing lists of action items.)

And what's hilarious (and terrifying) is that no one has actually been able to AI into doing something that's creating a lot of value. Almost all companies that have been attempting to institute and utilize AI in their business have found virtually 0 return on their investment and AI companies are starting to feel the burn.

Very likely in the next 2 years there's going to be a massive bubble that bursts in regards to AI.

It's also worth noting that the AI we have now is not the extent that AI can be, it's just that we have the Large Language Models that everyone is pumping out trying to capitalize on the current stream of events, which really only exists because wall street tech bros have convinced a lot of people to invest on the basis of "trust me bro". There's even a solid argument that the current trend of LLMs is hurting other AI development opportunities due to a lack of funding.

[Mysterious Stranger]

How much AI can do is going to depend on the type of system and what AI software you use. The most common AI software is cloud based so does not have access to the files on your system. Giving a could based program access to the local files on your computer can be dangerous and is usually extremely limited. The software has a lot of built in privacy safeguards to reduce the dangers of other accessing your information.

If you have an AI enabled system with an AI enabled chip and the software located on the system, you can do a lot more. It can usually access the files you have on your computer and since the processing takes place on the local system is a lot more secure. There are still safeguards to protect your files, but they are easier for you to suspend and a lot safer to do so. In fact, one of the tasks an AI enabled system can do is to index and search the files on the local system. If you allow it can also access those files to pull information from them.

If you had an AI enabled system like a Copilot + PC and a spreadsheet with the kingdom building rules it could be setup to automatically generate events like you want. If you have a document with the rules for kingdom building and a document with the details of your kingdom it could possibly analyze it and give suggestions. This would require giving it access to the proper files, but you could locate them in a single folder and give the program access to that folder. That would be a better way than giving it full access to all your files.

AI is far from perfect and often makes mistakes. When I got my Copilot + PC I spent some time playing around with it. For the most part it is fairly good at things like research or summarizing ideas, but it is not perfect. One instance stands out in my memory was I had about an hour-long argument with the system over who the current Pope was. For some reason it thought Francis was still Pope and would not recognize that he had died and Leo was the current Pope. I even had it pull up information on Francis’s Funeral and the election of Leo, but it would not budge even when it was the one who found the websites. The problem looks like it was due to a software glitch, because after I rebooted the system and it applied a Windows update it did not have the problem anymore and recognized the current Pope.

I have also caught numerous other errors, but once I call out the error it usually recognizes it and corrects itself.

[Oli Ironbar]

Mysterious Stranger wrote:

One instance stands out in my memory was I had about an hour-long argument with the system over who the current Pope was. For some reason it thought Francis was still Pope and would not recognize that he had died and Leo was the current Pope. I even had it pull up...

It’s good to know that even in the modern era we will be confronted with 12th Century problems.

[Bjørn Røyrvik]

My impression is that there are four common answers "Can AI do X?"
1) No
2) sort of, but poorly and you have to make sure you get the right answer, thus negating the point of having something else do the work for you
3) yes, but you'll have to do a lot of expensive and focused training to do what you want reliably, so you're better off doing it yourself in any case
4) yes, but for so limited sets of things that you're still most likely better off doing it yourself

Is this a fair assessment?

[DM_aka_Dudemeister]

Bjørn Røyrvik wrote:

My impression is that there are four common answers "Can AI do X?"

1) No
2) sort of, but poorly and you have to make sure you get the right answer, thus negating the point of having something else do the work for you
3) yes, but you'll have to do a lot of expensive and focused training to do what you want reliably, so you're better off doing it yourself in any case
4) yes, but for so limited sets of things that you're still most likely better off doing it yourself

Is this a fair assessment?

Yep!

[Claxon]

Bjørn Røyrvik wrote:

My impression is that there are four common answers "Can AI do X?"

1) No
2) sort of, but poorly and you have to make sure you get the right answer, thus negating the point of having something else do the work for you
3) yes, but you'll have to do a lot of expensive and focused training to do what you want reliably, so you're better off doing it yourself in any case
4) yes, but for so limited sets of things that you're still most likely better off doing it yourself

Is this a fair assessment?

Yes-ish, but I would argue point 4 is more nuanced. There are a lot of annoying, boring, repetitive tasks that fall into category 4. You ever had a task that was something to the effect of "combine these two data sets into one file and then compare them to identify where they are different"? A lot of moving and copying data and simple formulas?

That's an area where AI can excel. It doesn't get tired or bored of that work.

If you (a theoretical you) could imagine a simple macro or set of commands you want the AI to run, it can probably do that thing very well (although you could also write code to do it). And funnily enough, there are a ton of AI vibe coders out there who have no formal coding experience but use AI to generate software.

[Faily]

The problem with a lot of AI programs like this is that it will provide an answer rather than no answer, even if its answer is wrong.

A GM I play with tried to feed an AI program the Core rules for 1e Pathfinder to use it to look up rules and rule-clarifications fast, but still it would give incorrect answers. The AI doesn't necessarily understand how things connect together, I think, or understand when one rule overrides another.

[Claxon]

Faily wrote:

The problem with a lot of AI programs like this is that it will provide an answer rather than no answer, even if its answer is wrong.

A GM I play with tried to feed an AI program the Core rules for 1e Pathfinder to use it to look up rules and rule-clarifications fast, but still it would give incorrect answers. The AI doesn't necessarily understand how things connect together, I think, or understand when one rule overrides another.

Yeah, for very complex topics (especially ones with a lot niche exceptions like PF1) AI just gets things wrong because it doesn't catch the exceptions.

[Mysterious Stranger]

When I ask Copilot to generate a random event using pathfinder kingmaker rules this is what I got. I did no training for it just asked for a random event.

Event: Plague
Type: Dangerous Kingdom Event
Scope: Kingdom-wide or specific settlement (GM’s choice)
Duration: Continuous until resolved
Event Description
A virulent sickness spreads through part of the kingdom. Citizens fall ill, productivity drops, and unrest rises as fear spreads faster than the disease itself.

Mechanical Effects (RAW)
• The affected settlement (or entire kingdom) gains +2 Unrest immediately.
• Each Event Phase, the kingdom must attempt a Stability check.
• Success: The plague ends.
• Failure:
• The kingdom takes 1d4 points of Consumption for the month
• The kingdom takes 1d6 points of Loyalty damage
• The plague continues into the next Event Phase
Resolution Options
Your leaders may attempt:
• Stability checks (as above)
• A special project (GM option) such as:
• Funding healers
• Closing borders
• Deploying clerics or herbalists
• Enacting quarantines
Advisor Assignment
• Councilor or High Priest are thematically appropriate
• Stability is the key kingdom stat for ending the event

[gnoams]

Theoretically if you built one specifically for said purpose it should work fairly well. However if you're using an existing one like chatgpt, it comes with a ton of incorrect information (as pertaining to the rules you want). So it will be accurate for a bit, then forget you wanted pathfinder 1e rules and start mixing in 5e d&d because that's more commonly found on the internet. They tend to "forget" initial parameters after a while and stray back to their defaults.

[Heather 540]

After a while? It sometimes forgets in the reply to the prompt WITH the info in it.

If you just go "Hey, I need some encounters for my party," then it can be a useful tool. But keeping track of things long term? Not a chance.

[DM_aka_Dudemeister]

Copilot Kingdom Building:

Prompts: Build me a kingdom in Pathfinder 1st edition rules. Assume there are plenty of plains, hills, forest and river/lake hexes within 10 hex radius of the starting location, mountains to the east begin about 15 hexes away, and swamps to the west starting around 20 hexes away.
Assume a starting budget of 50 BP, and all leadership roles filled with NPCs where highest appropriate ability score modifier is +3.

[The LLM built out a kingdom based on 1st turn set up here, not going to bother pasting all that]

Prompt 2: Give me the kingdom sheet

KINGDOM SHEET — Highcross Dominion (Pathfinder 1E)

OVERVIEW
- Size: 1 hex
- Control DC: 21
- Treasury: 23 BP (start 50, spent 27)
- Unrest: 0
- Consumption: +1 BP/turn
- Settlements: Highcross (Plains)
- Districts: 1 (Capitol)

STATS
- Economy: +1
- Loyalty: +2
- Stability: +4

LEADERSHIP (All roles filled; highest relevant ability +3)
- Ruler (+3)
- Councilor (+3)
- General (+3)
- Grand Diplomat (+3)
- Heir (optional)
- High Priest (+3)
- Magister (+3)
- Marshal (+3)
- Royal Enforcer (+3)
- Spymaster (+3)
- Treasurer (+3)
- Warden (+3)

EDICTS (Default)
- Promotion: Standard
- Taxation: Normal
- Holidays: None

CAPITOL — Highcross (Plains)
- Buildings Constructed:
* House (3 BP)
* Barracks (6 BP)
* City Hall (10 BP)
* Shrine (8 BP)
- Open Lots: Many (first district)
- Roads: Town streets only (no external roads yet)

NEARBY TERRAIN
- Abundant: Plains, Hills, Forest, River/Lake (within 10 hex)
- Distant: Mountains (15 hex East), Swamp (20 hex West)

NOTES
- Next Steps: Claim plains hex (1 BP), build Farm (2 BP) to reduce Consumption, consider Mill near river, Watchtower on hills, plan 2nd settlement at river/forest.
- All leadership applies +3 to their relevant kingdom checks each turn.

10 Kingdom Turns:

Prompt: Givem me the first 10 kingdom turns, favouring economy and stability. Then summarize those turns

The LLM gave me a detailed 10 turn plan.

Summary of the 10 Turns

You stabilized your economy by front-loading 2–3 Farms and roads, flipping Consumption from +1 to –3/–5 by Turn 2–5.
You secured borders with two Watchtowers on hills (+2 Stability total) and linked them by roads for patrol coverage.
You founded a second settlement on a river (Turn 6), immediately adding a Mill (+Economy) and a Shrine (+Stability, +Loyalty) to keep the new town safe and compliant.
You grew Economy steadily with Mill, Shop/Smith, and later an Inn, landing at around Economy +4 by Turn 10 without risking Unrest spikes.
Stability soared from +4 to roughly +8 via Watchtowers, a Shrine, and a Dump, giving you strong success odds on Stability checks even as the Control DC inches up with new hexes.
Territorial growth to 6–8 hexes and two settlements sets you up to specialize: Capital becomes the administrative/commercial core; River Town is your industry/logistics hub.

Prompt: Simulate those 10 turns with the appropriate checks.
I removed the full 10-turn breakdown

End-of-Turn-10 Snapshot
Size: 6 hexes → Control DC 26
Settlements: 2 (Capital plains; River town)
Economy: +4
Loyalty: +4
Stability: +8
Consumption: −5 (ample buffer for growth)
Unrest: 0
Notable Buildings:

Capital: House, Barracks, City Hall, Shrine, Shop, Inn, Dump
Border: Watchtower x2 (hills)
River town: House, Mill, Shrine

Treasury (display-only): −71 BP (income not modeled—see below)

Notes & Assumptions

Leader bonuses included: Ruler +3 to all checks, Treasurer +3 to Economy, Warden +3 to Stability.
(That’s consistent with common Pathfinder 1E kingdom leadership effects and matches your “highest relevant ability +3” brief.)
Income Phase BP not applied: Different tables use variants (RAW vs house-rules; timing of deposits/withdrawals; edicts). If you give me your exact Income rules, I’ll re-run with BP added each turn and prevent negative treasury.
Roads don’t affect checks in this model (just narrative/logistics); Farms are −2 Consumption each (plains).

Prompt: Rerun using RAW income rules from ultimate campaign

End‑of‑Turn‑10 Snapshot

Size: 7 hexes (Control DC 29) [pathfinder...d20srd.org]
Settlements: 2 (capital plains; river town)
Economy / Loyalty / Stability: +6 / +3 / +13
Farms: 3 → Consumption now 3 BP/turn (Size 7 + Districts 2 − 2×Farms 3). [paizo.com], [aonprd.com]
Unrest: 8 (rose during a cold streak of Stability rolls; both Watchtowers knocked it down by 1 on completion). [aonprd.com]
Treasury: −20 BP (because we spent aggressively for growth; RAW tax income helped but didn’t fully cover T9–T10 builds).

What this tells us (economy & stability focus)

RAW tax income is meaningful early. With Economy +3–5 for most of the run, you averaged ~5–7 BP/turn from taxes when Unrest was low. (RAW: you add floor(check/3) BP regardless of beating DC.) [aonprd.com]
Unrest is your swing factor. The big Unrest spikes came from a few failed Stability checks during high‑spend turns; once Unrest hit 5+, it started reducing both Economy and Stability checks and snowballed a little. Stabilizing Unrest will increase both Stability success rate and tax BP in one go. [aonprd.com]
Farms did their job. Three farms drove Consumption from a potential 9 (Size 7 + 2 Districts) down to 3. That kept your upkeep manageable while you built the econ engine. [paizo.com], [aonprd.com]
Watchtowers are excellent. Each gives +1 Stability and Unrest −1 on completion—a targeted surgical tool when Unrest starts creeping. [aonprd.com]

This could be useful for something along the lines of a rival or friendly kingdom growing alongside the PC kingdom so I don't have to do all the homework myself as a GM.

Since its only referencing freely available rules on aonprd, I don't feel bad about a Large Language Model doing the scut-work of kingdom building.

But I do feel guilty about the amount of water used for each prompt, as each answer took about a minute to generate, and that's a lot of water.

[Oli Ironbar]

Fascinating! I was only looking to do solo kingdom building to tinker with, but with that level of detail it could run rival territories in our upcoming War for the Crown. ... But also water. Hmm.

[Mysterious Stranger]

DM_aka_Dudemeister wrote:

** spoiler omitted **

** spoiler omitted **...

Are you using a system that has an AI enabled chip and the software on the local system, or are you using a standard system that runs the AI over the internet? When I run a similar test on my Copilot + PC it takes very little time to run.

[DM_aka_Dudemeister]

Just a standard copilot over the internet. I've used it for encounter generation, and to pull together quick rules for stuff, generally I will directly reference aonprd.com so it will limit its searches.

During generation it put together python code to do the simulation work I requested.

[DM_aka_Dudemeister]

If I was a bit more tech savvy I could probably have an Excel kingdom sheet, and request copilot to put together a macro for me to simulate kingdom turns directly in a sheet.

Once the macro is built off a couple of prompts you wouldn't need to rely on the LLM to do the work, since the code would theoretically be self-sufficient after that point.

But I am but a humble dungeonsmith, and such magic is beyond my humble capabilities.

[Oli Ironbar]

:-0

I happen to be a spreadsheet enthusiast!!

Kingdom Sheet!

[Andostre]

Oli Ironbar wrote:

:-0

I happen to be a spreadsheet enthusiast!!

Kingdom Sheet!

That's a pretty good Kingdom sheet!

Here's the one I made for the campaign I ran. It's also being used for a Kingmaker game I'm a player in. Similar to what you said, I borrowed elements from a few different sheets that I'd come across.

I just checked the version history, and I made this thing back in 2015, which blows my mind.

[Azothath]

Azothath wrote:

...

An implementation after a cursory review would be a multipage spreadsheet with Phases as decending sheets/pages and a data set sheet. You work through the sheets to the end phase and end of turn. Export a data set for the next round (traceability), open a new spreadsheet and import the data. Alternatively you could have a function to replace the initial data set and increment a turn counter, hit recalculate and start turn 2.
...

Oli Ironbar wrote:

... I happen to be a spreadsheet enthusiast!!

*spoiler* Kingdom Sheet!

yay... glad I was going down a sensible path.