If Copilot and other AI are real, why isnt "vibecoding" taking over modding?

[chaosarah]

I often have discussions with colleges and friends about all this AI-hype. But i have been wondering for a while now, wouldnt modding be the first thing AI-coding should take over, if it worked?

We have mostly moderatly professional people with often limited to medium programing skills, takes years of hard work to create mods. Shouldnt "AI" be the perfect tool to make this a every persons game?

 

So why havent we seen a flux of AI-mods, no AI-generated pictures but real adaptations of base games? 

So far i am not aware this "revolution" was or is happening. Shouldnt things like animation mods or UI mods or general game mechanic mods be easy IF this "vibecoding" would work? Shouldnt it attract alot of modders, which wish to sharpen thier AI-skills? Wouldnt modders be able to build up thier own library of trained models, which make them spit out mods and later even games at a much accelerated pace?

 

 

So am i missing this AI-explosion or is this just not happening and AI vibecoding is basically bullshit?

[value_undefined]

There's a few factors at play here impacting it. The models are, while impressive, not nearly as good as the companies invested in them want you to believe they are. They cannot just one shot a complicated mod, and require someone who is already familiar with programming to steer and guide it through the process. It's not really a "program me a mod" level of function which is going to limit the number of people who can actually make use of it.

 

Similarly, modding still requires a certain amount of creativity. At the end of the day the concept and design of how the mod will work still depends on the human in the mix. You can have a model scope some of this out and even suggest a general workflow, but if you are trying to build something engaging that people want to use it's going to depend on the skills of the person involved.

 

I do have a conspiracy theory that it actually has had a large impact on the modding scene though. The number of SKSE mods on Nexus has ballooned recently ranging from anything as simple as "fix this one specific quirk" to full on remakes of older, script heavy mods. For most of these mods the author's tend to ignore the standard practice of releasing the source for the SKSE mods and get very defensive when asked about it. Personally I think they just don't want it to be too obvious that the mod is largely vibe coded.

Edited May 19 by value_undefined

[Papersword]

10 minutes ago, value_undefined said:

I do have a conspiracy theory that it actually has had a large impact on the modding scene though. The number of SKSE mods on Nexus has ballooned recently ranging from anything as simple as "fix this one specific quirk" to full on remakes of older, script heavy mods. For most of these mods the author's tend to ignore the standard practice of releasing the source for the SKSE mods and get very defensive when asked about it. Personally I think they just don't want it to be too obvious that the mod is largely vibe coded.

 

That's actually a very interesting observation. I don't think I clocked that myself despite noticing a huge up-tick in SKSE mods.

[Pamatronic]

49 minutes ago, chaosarah said:

vibecoding is basically bullshit?

Pretty much.

Speaking as someone who is a professional software developer (both C and object orientated languages), I can tell you wihout hesitation that ai tools currently are in no shape to genuinely replace human programmers or mod creators in any meaningfull capacity. 
I do personally use copilot integration in Vscode when working on skyrim mods, but even here its mostly limited to rather tedious tasks like for example adding another slider to an MCM. Nothing which requires much thinking. 

I have tried to use it on some more complicated tasks, but it always failed miserably because it seemingly doesn't understand papyrus and always wants to use methods which simply dont exist here, but only in other languages. And this completely ignores the fact that 90% of skyrim coding is knowing the quirks of the existing functions when actually used ingame, which an ai obviously doesn't because it cant play the game...
And then there's the other issue of the Creation kit, which you at least need some cursory knowledge of to even get your scripts compiled and into the game. Also something the ai massively struggles with.

And even in "real" languages, ai always hast this issue that it breaks down the second you want it to do any real work. it either hallucinates something faulty, or something so messy that cleaning it up would require more work then just doing it yourself to begin with. So no one in our company uses any paid model or IDE integrations.

TLDR: Ai cant do papyrus and is generally less competent then their creators want you to think.
 

 

25 minutes ago, value_undefined said:

I do have a conspiracy theory that it actually has had a large impact on the modding scene though. The number of SKSE mods on Nexus has ballooned recently ranging from anything as simple as "fix this one specific quirk" to full on remakes of older, script heavy mods. For most of these mods the author's tend to ignore the standard practice of releasing the source for the SKSE mods and get very defensive when asked about it.

cant explain the defensive part, but personally I would argue most modders simply forget to include the sources for SKSE plugins because users and other authors simply wont need it.
 

[Reginald_001]

1 hour ago, chaosarah said:

I often have discussions with colleges and friends about all this AI-hype. But i have been wondering for a while now, wouldnt modding be the first thing AI-coding should take over, if it worked?

We have mostly moderatly professional people with often limited to medium programing skills, takes years of hard work to create mods. Shouldnt "AI" be the perfect tool to make this a every persons game?

 

So why havent we seen a flux of AI-mods, no AI-generated pictures but real adaptations of base games? 

So far i am not aware this "revolution" was or is happening. Shouldnt things like animation mods or UI mods or general game mechanic mods be easy IF this "vibecoding" would work? Shouldnt it attract alot of modders, which wish to sharpen thier AI-skills? Wouldnt modders be able to build up thier own library of trained models, which make them spit out mods and later even games at a much accelerated pace?

 

 

So am i missing this AI-explosion or is this just not happening and AI vibecoding is basically bullshit?

Because modding is a hobby, not a job.

 

You don't say 'I wanna learn to play violin', and then go to an AI site to prompt violin music.

It's really that simple.

 

Has nothing to do with the (complete and utter lack of) quality of vibecoding right now. And like Pamatronic said, as soon as you want to do real work, AI falls apart in hallucinations.  

Edited May 19 by Reginald_001

[anon6076]

Also, modding involves a lot of working with instruments that ai is not so good - textures, models, etc. And modding of this site kind is just not allowed for ai. Very haram!

[Vader666]

3 hours ago, anon6076 said:

modding involves a lot of working with instruments that ai is not so good - textures, models, etc.

 

Generating fully textured 3D models out of a single image is not only very much possible, there are models out there which quite frankly do a "decent enough" job at it.

 

5 hours ago, chaosarah said:

wouldnt modding be the first thing AI-coding should take over, if it worked?

 

I dont think so.

Mostly because modding in games comes down mostly to scripting in the engine- / game- specific language instead of stuff like Python or C++.

Since those languages have a much more narrow scope and  are pretty much tied to a given engine or even a specific engine version, creating an LLM that is capable of generating code in those languages just isnt worth the hassle.

 

[Zdun]

LLM models are basically big Neural Networks. They read whole internet and based on that generalize enough to predict next word. The problem with modding is there is not the whole internet of mods. Mods are divided between different games, languages and technologies(code, assets). There is not enough training set for LLM to learn on it.

 

Still LLM can help you to write mod description, create assets or assuming you are moding in language which is popular enough for LLM to generalize solutions to give you answers. Eventually LLM are getting better and better in generalization and there is posiblity that moding scene will create more and more code but we are not there yet.

[Operand]

On 5/19/2026 at 7:47 PM, chaosarah said:

I often have discussions with colleges and friends about all this AI-hype. But i have been wondering for a while now, wouldnt modding be the first thing AI-coding should take over, if it worked?

 

The simple answer is: it already does to some extent. First address libraries for Fallout4 NG F4SE were AI-made.

The honest answer is: there's a platitude of reasons why it doesn't "take over", to name a few:

• AI with decent coding capabilities isn't cheap and can reach $200/month; for something where you hope to make a living of it could be reasonable, but for a passion project like a mod? No way.
• Most of the modding touches things that are proprietary and would clash with various EULAs of those corporations
• Needless to say, adult modding is out of the window completely, with likely some SWAT knocking on your door if you try (okay, this is a joke. for now)
• Even competent AI needs a lot of guidance and it won't complete the design, it is only decent at a direct and orchestrated task and most of the serious modding requires structural and architectural knowledge
• Most of the models were never trained on very narrow specialized field of game internal engines because those are proprietary, so the results are worse in the first place 

"Local AI install" won't solve any of these problems because:

• Only "leaked" old models like GPT4o are available, so it's very degraded/outdated/irrelevant versions of frontier models
• Open-source models are always much worse because let's be honest, corporations steal a lot of data with their scrapers and get away with it, but the open-source cannot afford to do that (see, it's only legal if big money corpos do it, us mortals will get sued to oblivion if caught)
• Most of the models will be censored to the limits and freak out at the slightest issue
• Even abliterated / uncensored models are limited, the un-censoring can only fix inference, not the training
• Most of the models for local installs need to be quantized (how else would you run 256G+ models in memory) and that reduces their output quality a lot
• Most people cannot even dream of the hardware that can run even semi-viable models locally (I own 5090 and I have 96Gb or RAM with X3D specialized CPU as well, my hard limit is 80B quantized model which is just barely a starting point)
• The usefulness is further diminished due to very limited context window. For anything near 70B the most that my config can realistically handle is 8192 tokens in the context. It "technically" works with higher numbers... if I am ready to wait for 10..15 minutes (!) for a single response once it passes the 8-10k tokens. And 8k tokens isn't much at all, it's about 8-10 mid-length responses, so a very brief conversation at most.

 

  

On 5/19/2026 at 8:54 PM, Pamatronic said:

Pretty much.

Speaking as someone who is a professional software developer (both C and object orientated languages), I can tell you wihout hesitation that ai tools currently are in no shape to genuinely replace human programmers or mod creators in any meaningfull capacity. 

 

I would not be so categorical. On a directed task with a well-known technology and strong referential base, frontier Claude proprietary models generally impress me with their results. Caveat - they need a human to check their work, but it's still a lot of help. I'd say, in many cases it saves up to 80-90% of the work. "Many" does not equal "all", of course.

 

So, the main issue is that if you define "vibe-coding" as "someone clueless about engineering to do the engineering work", then your point holds true. Because in my experience, to be successful with AI utilization on coding tasks, one already needs to be an engineer, and a pretty high level one at that, since you'd need to know the cost-benefit analysis of the trade-offs you make, call out AI bs when it happens, write a maintainable code, etc - though, in many cases, yet again, you can instruct AI and it will self-correct.

 

I'm speaking from real experience, though I work in big tech with virtually unlimited access to tokens, so I am free to experiment however I wish on a corpo dime with AI and I managed to achieve good results that save a lot of time.

Edited June 3 by Operand

[Aouregan]

A lot of modding requires knowledge of workarounds or methods of circumventing specific obstacles the game throws at you.

Especially with games that do not support modding out of the box, where you then have to start injecting code at runtime level and all that fun stuff.

I can see AI code being used for parts of the process, not its entirety. I mean, it is not like we have to worry about things like data protection, infosec, and all those goodies when modding the average game. I wouldn't trust AI with writing hardened code.

 

I personally am not a big fan of the whole AI fad. I think it is dangerous. Not because of the quality of the code (albeit that sometimes is also very questionable), but rather because of the resource consumption.

I am also a very old fashioned coder, more or less growing into the job in the 90ies, maintaining mainframes in Cobol and coding with physical manuals next to me on my desk.

 

Someone somewhere put it very nicely:

 

The purpose of AI is to allow wealth to access skill, without allowing skill to access wealth. 

Edited June 3 by Aouregan