Use AI to Build Your Game. Just Keep It Out of the Part Players Feel.
Iron Nest shipped with zero AI, Dragon Snacks put its AI in the build pipeline, and Valve’s own policy already explains why both of them won
TL;DR
Two people, $0 marketing, zero AI disclosure. Iron Nest: Heavy Turret Simulator finished #1 in wishlists gained and #2 in unique players at Steam Next Fest in June 2026, and it still closed the week at 99% Overwhelmingly Positive.
53% fewer reviews is what Game Oracle’s controlled study found for high-potential Steam titles that disclose AI use, and Iron Nest never had to find out what that number does to a launch.
Dragon Snacks Games used AI too, just never anywhere a player could see it, and the payoff was a piece of networking architecture their engineer would not otherwise have had time to build.
Valve wrote this exact distinction into policy in January 2026, and Epic’s CEO is now publicly fighting it, because Unreal Engine 6’s whole pitch depends on the line Valve just drew not existing.
A founder called me the week after Steam Next Fest had ended, wanting to know if he should put “built with zero AI, ever” on his Steam page like a badge of honour.
I asked him what he thought that badge would actually buy him, and he did not have an answer beyond “it feels honest.”
It might be honest, but it is also not the point, and Iron Nest is the cleanest evidence I have seen all year for why.
Iron Nest and Dragon Snacks Games solved the same problem from opposite directions in the same six months, and both of them were right.
The Game That Never Had Anything to Disclose
Nick Talmers and Dominik Latos are a two-person team with twenty-plus years of combined experience between them. By their own account they built a game about crewing a five-thousand-ton dieselpunk artillery platform in an alternate-history 1920s Spain on a zero-dollar marketing budget.
There is no minimap and no targeting reticle. You take coordinates off a typewriter, chart bearing and range on a paper map, run the numbers through a ballistic calculator, load the correct shell by hand, and rotate the turret yourself. Kotaku’s reviewer put it simply:
“Every operation in Iron Nest feels tactile and absurdly meticulous.”
The Steam page describes the game as the first step in a larger universe the two of them intend to keep building. By April 2026 they had 227,816 net wishlists on a two-person, six-month, zero-marketing run, and by June they had finished the week #1 in wishlists gained and #2 in unique players at Next Fest, with that 99% Overwhelmingly Positive demo score still intact.
Here is the detail that matters for this piece: there is no Steam AI disclosure on this page, and under Valve’s January 2026 rules, that is the operative line. Nick has never publicly claimed AI use anywhere in the pipeline, which does not rule out a coding assistant humming away under the hood somewhere, since Valve’s efficiency-tools exemption means he would have no obligation to mention it if there were.
What it does rule out is anything generative touching the layer a player can feel, because nothing in the finished game gives a player reason to wonder who, or what, built it.
That absence traces back to how the game found its own setting, not just its marketing. In a preview interview with TheSixthAxis, Nick and Dominik explained that the dieselpunk world was not a pitch decision made up front.
Nick found that the more automation and digital systems he stripped out of early prototypes, the more engaging the gameplay became. That discovery is what pushed the setting toward a dreadnought-era alternate history. Iron Next is built around analog systems and massive naval guns, in a world where the arms race and old European power structures never got interrupted by a First World War.
The tactile design did not get bolted onto a marketing hook after the fact; removing automation was the prototyping process that produced the game.
A game whose core discovery was that less automation makes it better has nothing to gain from quietly automating the part a player touches; since every disclosed asset would have been a small tear in the exact fabric the game is selling.
Iron Nest did not win Next Fest by having an AI strategy; it won by having nothing in that column at all.
The Studio That Used AI Correctly
Dragon Snacks Games is a small studio run by Jen MacLean , a former Xbox executive, and she has made a specific and unusual public promise: players “will not see generative AI” in her game. That is not really a policy so much as a commitment, one kept entirely away from anything player-facing.
Where MacLean’s studio uses AI heavily is on the engineering side, on what she calls, in her interview with Chirag Vadhia’s newsletter The Side Quest, “the tasks that no one wants to do.”
Automating the build process freed her lead engineer from infrastructure maintenance, and that freed time went into building peer-to-peer networking for sixteen-player sessions that only has to store a world state rather than run a server for every game instance.
The AI efficiency gain did not just save a few hours of tedium; it funded an architectural decision that lowered cost and improved the actual player experience at the same time.
MacLean is just as clear about where she thinks the money is wrong.
She calls generative gameplay AI “so nascent” and warns that people “confuse environments with gameplay,” since a model can generate a town to walk around, but as she puts it, “walking around a town is generally very not very fun” without the designers who make it worth the walk.
Chris Nemcosky (Nemco), Dragon Snacks’ co-founder and CTO, added more detail on this in a direct message to me. Several other systems came online through AI tools too:
The backend that stores user-generated content, specifically the in-game paintings players create
Unit test generation for core systems like inventory
Tracking down what he called some very nasty bugs
The thing about Nemco is he knows what he’s talking about.
A grizzled technology veteran: VP of Engineering at Backflip Studios on Dragon Vale, then VP of Studio at Scopely, where he helped bring Monopoly Go to market, one of mobile gaming's biggest hits this decade. He knows what he's talking about.
Dragon Snacks is not the only place this shows up at a useful scale.
The distinction between AI coding tools has its own practical shorthand worth keeping in mind: autocomplete-style assistants save time on tedious, well-bounded tasks, while assistants built for whole-codebase understanding save the cognitive overhead of holding an entire system in your head during a refactor.
For a two-person team, that is not a nice-to-have; it is the difference between shipping and not shipping at all.
Valve Already Drew This Line
This is not just a design philosophy that two teams happened to arrive at independently. It is now Steam’s actual policy.
Valve rewrote its AI disclosure form in January 2026, and the rewrite did one specific thing: it split “efficiency tools” from “generative content that ships with your game and is consumed by players.” Valve’s own language:
“Efficiency gains through the use of these tools is not the focus of this section.
Rather, it pertains to the application of AI in generating content that is included in your game and experienced by players.”
Code assistants, build automation, and QA scripting are exempt; while AI-generated art, voice, narrative, or anything live-generated during play, whether pre-rendered or produced on the fly, must be disclosed publicly on the store page.
Iron Nest and Dragon Snacks are not two studios who happened to guess the regulatory future correctly. They are the demonstration that the line Valve drew was already the line that mattered commercially, months before Valve made it official.
The Counter-Example Worth Taking Seriously
Not everyone agrees the line should exist at all, and the person disagreeing loudest has the least neutral seat in the room.
Epic’s Tim Sweeney called Steam’s disclosure requirement a “Scarlet Letter” in a June 2026 interview, arguing that any developer who discloses AI use risks “a hater community trying to kill the game,” and calling the policy “really irresponsible of Valve.”
Sweeney is not wrong that the stigma is real, but his argument gets complicated once you look at the timing. Epic showed off Unreal Engine 6 at State of Unreal in Chicago that same month, with Claude and Gemini built in as core architecture rather than as a plugin.
Live demos furnished a virtual apartment and re-lit an entire city scene from a Claude prompt; Epic’s own framing was that “for UE6, we see LLMs, generative AI models, and tools like Claude and Codex playing a central role in helping you build content faster.”
That is exactly the category Valve’s policy requires disclosure for, not the exempt one. If a UE6 title ships AI-furnished scenes or AI-relit environments that a player experiences directly, that is generative content under Valve’s own definition, whether or not Sweeney would prefer it classified as an efficiency tool.
Sweeney is not only defending small developers from an unfair stigma; he is defending the commercial viability of his own engine’s flagship feature, which happens to sit on the disclosed side of the exact line he wants erased.
The debatable version of this is that AI’s value in games has nothing to do with how much of it you use; it is entirely about whether any of it touches the layer the player can feel.
What the Data Actually Says, and Doesn’t
Ross Burton, a data analyst at Game Oracle, built a causal model across 9,879 Steam releases from January to October 2025, controlling for publisher backing, developer experience, genre, and release timing.
Games disclosing AI use received 53% fewer reviews than comparable non-AI titles.
Median review score came in at 84.6% for AI-disclosed titles, against 88.3% for non-disclosed ones.
For established studios specifically, the modelled sales impact runs 40 to 60% down.
Here is the nuance most coverage of this study drops: Burton himself is careful to say the effect is not proven causation.
“Heavy AI use often correlates with other decisions that lead to a poorly crafted game,” as he has been quoted, meaning the disclosure tag might be a symptom rather than the cause.
The stigma is also not really a small-studio-versus-big-studio story so much as a quality-and-potential story, since Burton’s own framing is that “for low-quality games, AI makes no difference. But for high-potential games, the AI Stigma is real and severely punishes developers who otherwise would have succeeded.”
That is the population Iron Nest belongs to: a small team with a genuinely high-potential game that, had it disclosed anything, would have been exactly the kind of title Burton’s data says stood to lose the most.
It disclosed nothing, because there was nothing generative in the pipeline to disclose.
Mapping This Back to Automate, Augment, Protect
This all maps cleanly onto the framework from Where AI Belongs and Where It Doesn’t: automate the mechanical layer freely, augment the judgment-intensive layer carefully, and protect the irreplaceable layer fiercely.
Dragon Snacks is the textbook case. Build automation sits squarely in Automate Freely, and the freed engineering time went into Protect Fiercely territory, the actual multiplayer architecture players would feel the second it worked or did not.
Iron Nest is the more extreme case. There is no visible Automate layer at all, at least none the team has ever pointed to publicly, and every system a player touches, by their own design philosophy every system that exists, sits in Protect Fiercely. That is not a studio leaving efficiency gains on the table out of stubbornness; it is a studio whose entire commercial bet is that the tactile authenticity of the Protect layer is the product, so there is nothing to gain by automating around its edges.
Where the Line Actually Gets Tested
Three more cases are worth holding up against the framework, because they are where it stops being tidy.
Roblox’s Cube model is Automate Freely at platform scale.
4D generation has produced over 160,000 objects during early access.
In Wish Master, the game where Roblox tested it first, players using the feature showed a 64% increase in average play time.
The tool generates a car with wheels that actually spin, or a bookcase with shelves you can configure, but it does not decide what the game is or whether the object belongs there. A human still has to assemble the world around it.
Roblox’s new Planning Mode sits right on the Augment/Protect boundary.
By April 2026, 44% of the platform’s top 1,000 creators were already using Roblox Assistant or third-party tools like Claude, Cursor, and Codex through Roblox Studio’s own Model Context Protocol (MCP) integration.
Planning Mode analyses a creator’s existing code and data model, asks clarifying questions, and produces an editable action plan before touching anything.
The creator reviews and approves the plan before execution starts.
Whether this lands in Augment or drifts toward something riskier depends entirely on whether the developer treats the plan as a draft to interrogate, or a directive to rubber-stamp.
Suck Up! and The Finals are the two examples Ross Burton’s own study names as AI-disclosed successes, and neither is a template a small studio can casually copy.
In Suck Up!, live AI-generated NPC dialogue is not a shortcut around writing. It is the entire design premise, since every conversation is genuinely improvised by the model in response to what the player says.
In The Finals, Embark disclosed AI text-to-speech for in-game announcer lines and survived the backlash, but Embark is a Nexon-backed studio with an established audience already willing to extend it trust before the controversy hit.
The pattern across all of it: the tools that work are the ones a human reviews before a player ever sees the result, or the ones where the AI generation is disclosed and is the game rather than a shortcut hiding inside it.
The first model requires AI generation to be the product itself.
The second requires institutional credibility most small teams have not built yet.
Neither is the same move as automating a build pipeline nobody was going to notice either way.
Iron Nest simply sits at the conservative end of that same spectrum, not in a different category entirely.
If You’re Building Right Now
If you’re a small studio deciding where AI belongs in your pipeline:
Ask what your Iron Nest is
The specific thing a player would feel the tool touching, and;
Build a wall around it before you build anything else.
If you’re evaluating a pitch or a greenlight:
Treat an AI disclosure on a high-potential title as a modelled 40 to 60% sales risk, not as neutral information.
If you’re building on UE6 specifically:
Plan your Steam page copy on the assumption that disclosure will be required,
Do not bet your launch window on Sweeney winning this argument first.
The Reframe
The studios spending 2026 arguing about how much AI to use are answering a productivity question. Iron Nest and Dragon Snacks both answered a different one: which layer of the game is the product, and what happens the moment a tool touches it without anyone knowing.
Next time: the EU’s AI Act Article 50 takes effect August 2.
It is the bloc’s transparency and disclosure rule for AI-generated content, backed by fines up to €15 million or 3% of global turnover.
It is not a games-specific rule, but the Commission’s own draft guidance already names AI-enabled NPCs and professional-only coding assistants as likely candidates for the same obvious-so-exempt carve-out Valve built into its Steam policy. Part II of my Three Machines report already covers the fuller EU compliance picture against the US and China’s competing regimes.
The harder question, whether a Steam page with AI-generated art trips this for a studio based outside the EU, is still open, since the Commission’s final scope guidelines had not landed as of the Code of Practice’s June 10 publication.
Abbas Saleem is a Principal Consultant at Llama & Griffin, advising game studios, streaming platforms, and investment funds across six continents. He writes The Pattern Recognition: gaming industry intelligence 12 to 24 months before it becomes consensus. LinkedIn | Schedule a call
Gaming industry pattern recognition from inside the conversations. How a two-person team out-shipped studios with marketing budgets, why Valve’s policy language is quietly becoming the industry standard, and what the EU’s August disclosure rules mean for every Steam page being written right now.