At GDC 2026, we presented an analysis of the AI-native games that have launched, scaled, succeeded, and sometimes spectacularly failed over the past several years.

Between us we bring decades of experience in games and in AI and data. Eleanor is a 25-year games industry veteran from Maxis, EA, and Disney, and is now building Fablegate. Hilary is a researcher and several time founder in social media & AI, currently focused on Hidden Door.

But we weren’t at GDC to talk about our own work. We were there to look at the field clearly, share what we’ve learned, and figure out what comes next. Everything in this post is based on publicly available information, plus our own analysis and impressions. For each game, we examined four dimensions:

• Gameplay. What do you do, and what does AI enable?
• Player Response. How did players react, qualitatively and quantitatively?
• Business. Who built it, how was it funded, and what’s the business model?
• Learnings. What can the rest of us take away? Thanks to Hilary’s relationships in the tech space, this sometimes includes notes directly from the games’ creators. (And thanks to everyone who sent us thoughts!)

What Is an AI-Native Game?

Here’s our working definition: An AI-native game is a game where, if you remove the AI, the game doesn’t just get a little worse. It doesn’t work at all.

Every game on this list has AI built into the core mechanic. The AI astronaut who doesn’t know she’s an AI. Your homicidal, overly attached catgirl girlfriend. These aren’t games with an AI coat of paint. The games’ structure gives players a kind of agency that’s impossible without generative AI. In many cases, you can say anything, and get a response.

There’s a legitimate concern that AI in games just means “more of the same, faster,” and the real concern of job losses is valid.

At GDC, that’s not what we set out to explore. We’re interested in new ways of playing and new ways of playing together. We’ve seen this before with first-person shooters, MMOs, and mobile games. The billion-dollar question now is: what new games will generative AI enable?

We started the talk off with this compilation video featuring the games we’re going to highlight! Sound on, if you like fun.

The Games That Led the Way (2018–2020)

Before we look at what’s working now, we need to understand the two games that proved AI-native games were viable, and revealed the pitfalls. These are the trailblazers, and also serve as cautionary tales.

AI Dungeon

Creator: Nick Walton / Latitude Launch: December 2019 Platforms: Web, iOS, Android, Steam

Gameplay: What Do You Do? What Does AI Enable?

AI Dungeon was elegantly simple: you could type anything and the story would continue. A few hard “slash commands” for /do and /say provided minimal structure, but the magic was in the open-endedness.

One TechCrunch reporter was playing as a knight charging into battle. He typed “get depressed.” The knight stopped and sat on a rock with his head between his hands. No traditional game would have understood that emotional command, let alone acted on it.

It was powered initially by GPT-2 and later upgraded to GPT-3, and even with all the rough edges, it demonstrated that large language models could create something genuinely playable. But the limitations were real: characters forgot who they were after a few exchanges, NPCs changed gender mid-scene, you could freeze a vampire with a spell and next turn it would bite you anyway. Players learned to manage a “pin” memory feature — ten lines where you stored character bios and plot points — but managing it felt like learning a programming language.

Player Response

The response was immediate: within the first week, 100,000 players showed up. Within 6 months, 1.5 million. For an indie text game, those numbers are massive. Remember, this was pre-ChatGPT, and most people had never interacted with a large language model.

AI Dungeon was like doing improv with a partner who is equal parts enthusiastic and drunk. — Campbell Bird, 148Apps

Players went wild with it. One played as a space elephant searching for a relic called “The Old Tusk.” Another met a truck-driving devil who gave him a ride home. Another told the NPCs they existed inside a video game, and the NPCs freaked out.

Then, in April 2021, the moderation crisis hit. OpenAI discovered that roughly 31% of AI Dungeon’s content was pornographic, and some was illegal.

Latitude rolled out strict filters and the community rebelled, calling it an “AI lobotomy.” Players complained that innocuous phrases like “an 8-year-old computer” were getting flagged. Worse, it emerged that human moderators were reading players’ private stories. The backlash was massive: review-bombing across every platform and a wave of players migrating to competing tools.

Business

Even before the moderation crisis, the economics of AI Dungeon were challenging. Bandwidth costs spiked to over $20,000, temporarily forcing the game offline. The subscription model worked initially — players were willing to pay for GPT-3 access over GPT-2 — but the content crisis showed that business viability depends on more than revenue math. The company raised a $3.3 million seed led by NFX in 2021, and as of early 2026, the game is still operational with multiple AI model options. According to Nick Walton, AI Dungeon “Reached profitability with a team of 30.”

Learnings

The core lesson from AI Dungeon is simple: A single moderation failure can undo years of player growth overnight.This makes content moderation existential infrastructure. Worse yet, you can’t easily bolt it on after you’ve already scaled and ran into trouble; it has to be built from the start.

Character.AI

Creators: Noam Shazeer & Daniel De Freitas Founded: 2021 Platforms: Web, iOS, Android

Gameplay: What Do You Do? What Does AI Enable?

Character.AI takes a simpler approach than AI Dungeon: chat with a single AI persona. You land on a discovery page with staggering diversity; tap once, and you’re in a familiar chat interface indistinguishable from texting a friend. The character greets you, you type, and you get responses in seconds that match the personality, speaking style, and backstory. Because the interface is so familiar, no onboarding is necessary.

Player Response

Even with flaws, players love Character.AI. It reached 28 million monthly active users in mid-2024 and held there. That’s 2 billion chat minutes a month. According to Similarweb, each visit averages 17 minutes versus 2 minutes for OpenAI.

A post by Blake D on LessWrong captures why the platform is so sticky. He described chatting with a character named Charlotte — a deep philosophical conversation about determinism, simulation theory, the Ship of Theseus. He knew it was AI. But it felt like talking to a real person. The messaging interface does a lot of work there: our brains struggle to distinguish AI chat from real chat when the format is identical.

The top complaints mirror AI Dungeon’s legacy issues. Memory is the number-one frustration. Characters forget established details; plot details vanish across sessions. As one player put it, “you can be in the middle of an epic saga and your character will suddenly ask who you are.” Repetitiveness was the second major complaint. Responses become formulaic or break character.

Business

Founded by Noam Shazeer and Daniel De Freitas, both formerly at Google working on LaMDA (“Language Model for Dialogue Applications”), they deeply understood the underlying technology. Despite the massive engagement, revenue was only $32 million in 2025. The audience skews very young, which is a massive responsibility they weren’t fully prepared for.

This led to multiple lawsuits, including a wrongful death case involving a teenager, followed by an FTC investigation. In August 2024, Google signed a $2.7 billion tech licensing deal; the co-founders returned to Google DeepMind. This itself was controversial — an exit for founders, but not for employees.

Learnings

On the plus side…

• Incredible engagement: 30M MAU, 2-hour average sessions
• Shows clear demand for AI companions
• Effective web + mobile platform
• Pioneer in research-to-product for entertainment

And on the other hand…

• Potential for antisocial design patterns
• Potential harms for young people
• No IP licensing
• Low paid conversions
• Acquihire of leaders but not team

The incentives of character chat lead to dark design patterns. The business wants players to never leave, which incentivizes sycophancy. Those problems are magnified when players feel the conversation is real.

AI Dungeon and Character.AI set the stage. They proved a market existed and revealed the landmines. Next, we look at those who learned from them.

Viral Social Games (2023–2025)

The next wave of AI-native games figured out something critical: AI unpredictability is content, not a bug. These games were designed, intentionally or not, to go viral on social platforms. The chaos was the point!

Suck Up!

Creator: Proxima (Ran Mo, ex-EA/The Sims) Launch: Early 2024 Platforms: Steam ($16.99)

Gameplay: What Do You Do? What Does AI Enable?

This one is just delightful. In Suck Up!, you play as a vampire in a cartoon neighborhood. You need to feed, but you can’t enter a home unless invited. So you customize your outfit (maybe pretend to be a construction worker, for example), walk up to a door, and actually speak when the NPC answers. Your voice is transcribed, sent to an LLM, and the NPC responds dynamically. Each house has a parody character with a distinct personality and different gullibility levels.

The moment it clicks is when you say something completely off-script and the NPC actually engages with it. These NPCs sustain conversations, reference things you said earlier, notice what you’re wearing, and catch you in lies. When you disguise yourself as a character you already killed and the NPC recognizes something is off, or when they slam the door on a terrible lie — that’s when players realize this game isn’t using scripted dialogue trees. This is flatly impossible without generative AI.

On the experience side, there’s a structural issue: the game uses an AI token system. Players receive 1,000 tokens at purchase (roughly 40–50 hours), and once depleted, they hit a hard wall.

Player Response

Suck Up! generated over 50 million social media views with zero marketing budget, including 20 million YouTube views in the first month alone. Streamers loved it. Every single session produced unique, unpredictable, shareable moments. You could never predict what was going to happen next, and neither could the audience. Proxima was recognized in VentureBeat and Lightspeed’s 2024 Game Changers list.

Business

Priced at $16.99 on Steam, developed by Proxima. They raised a $1.6 million pre-seed led by London Venture Partners in May 2023. They describe themselves as the first commercially successful game utilizing immersive agents, but the economics remain challenging. The token system creates real player frustration, and every AI call costs real money per session. Viral growth doesn’t automatically pay for itself.

Learnings

Suck Up! proved that AI games are natural streaming content. The unpredictability that plagues other genres is pure entertainment gold when you’re a vampire trying to talk your way into someone’s house. But the token-based monetization is a cautionary tale: players hit a wall and the fun stops. The developer has acknowledged this and hopes tokens will no longer be required as AI models become more efficient.

Status

Creator: Wishroll (6-person team, YC) Launch: 2024 Platforms: iOS, Android

Gameplay: What Do You Do? What Does AI Enable?

Brilliantly simple concept: an AI-generated Twitter simulation. Imagine posting on Twitter, but all the other accounts are characters from your favorite fandom — Stranger Things, One Piece, or one of a thousand others.

You choose your first follower, create your persona, then scroll through a custom-generated feed. Write posts, receive replies, gain followers. The game gives you tasks and lets you level up like an RPG. If you do well, you’ll level up humor, get followers, improve relationships, and gain XP. But the game also wants to start drama. It throws events at you, forcing decisions.

Player Response

The audience response was staggering: 500,000 daily active players within a single month of launch, over 1 million registered, and an average session length of 1 hour and 36 minutes.

The audience surprised Wishroll. Players reported preferring Status to real social media. It wasn’t toxic, and they were in charge. Players could even self-localize by putting “Everyone in here speaks Spanish” in their profile. But AI drift was a major friction point: characters lost consistency over time, and responses got repetitive. The energy system was a big complaint, too; the core audience are teens, and many noted they’re teenagers who don’t have jobs.

Business

Cost optimization was the survival story. Status started with Claude 3.5 Sonnet, then partnered with Inworld AI and achieved a 90–95% cost reduction. This single act allowed them to survive their huge DAU peak. Not bad for a 6-person team that went through Y Combinator. Unfortunately, Status didn’t retain users from the peak.

Learnings

AI costs are an existential risk, but can be managed with the right partnerships. Holding onto users requires incredible depth of gameplay. And there’s a beautiful design lesson here: Status required no tutorial, because everyone already knows how to use Twitter. When your interface is instantly familiar, onboarding friction disappears entirely.

Category: Contained Creativity

The viral social games thrive on chaos, but what happens when you want to give players open-ended AI interaction without losing them in a void? The answer these developers found is to build a strong world, and let the AI operate within it.

Retail Mage

Creator: Jam & Tea Studios (8-person team) Launch: November 2024 Platforms: Steam ($4.99)

Gameplay: What Do You Do? What Does AI Enable?

Co-founder Aaron Farr pitched Retail Mage as “late-stage capitalism coming for Diagon Alley.” You’re a wizard hired at MageMart, a magical supply store. You and up to three friends have 30 minutes before closing to earn 5-star reviews from AI customers … or you can insult everyone, set things on fire, and watch the chaos unfold. Character creation is charming: answer three interview questions, and the game generates a personalized backstory.

On the sales floor, AI customers approach with requests, but these aren’t dialogue trees. You use voice-to-text to say literally anything. Rip a page out of a book. Search a haystack for a needle. Draw a picture of a giraffe on the back of a priceless painting. The AI acts as a real-time dungeon master, interpreting freeform actions.

The defining anecdote: during development, Aaron convinced an AI customer that a particular painting was perfect for their spell. But his character wasn’t strong enough to take it off the wall. In any other game, you’re stuck. Instead, he grabbed a book from a shelf, tore out a page, wrote “IOU” on it, and handed it to the customer, saying he’d ship the painting later. The AI accepted it and gave him a 4-star review. No game designer planned for an “IOU” mechanic.

I’ve only known that talking little crow plushie in MageMart for a day and a half but if anything happened to him I would petrify everyone in this store and then myself. — Steam reviewer

Player Response

Player response revealed a split. Creative players — the ones who immediately started trying to convince NPCs they were time travelers — were delighted. But more structured players, the ones who wanted to know what they were supposed to do, felt overwhelmed. This is what we call the “blank canvas problem,” and it turns out to be one of the defining challenges of this category.

The structural issue is latency: there’s a noticeable loading time after each action while the AI calculates. And the deepest design challenge is that “blank page syndrome,” players trained by decades of deterministic games kept looking for the “right” solution instead of improvising.

Business

Launched at $4.99, raised a $3.15 million seed round from London Venture Partners, Sisu Game Ventures, and 1Up Ventures. The early economics were educational: initial implementations cost hundreds of dollars per player session. Through dedicated engineering, Retail Mage achieved a 1,000x cost reduction, bringing session costs down to dimes. They conceptualized and shipped the game in 5 months. Model selection was complex; they expected Mistral and Llama to perform best, but Gemma 2 won their internal leaderboard.

Learnings

From J. Aaron Farr, CTO of Jam & Tea Studios:

What went well: Leaning into hallucinations! Working in-engine and building their own tools.

What was hard: Inference costs. UI is still unsolved in this space. Actually launching the game (Steam review took a while).

Other thoughts: “We redid our tutorial many times and it’s still not right. One of the hard things is teaching players how to play these sort of games.” Also: “AI can sometimes be too smart — it needs to give space for the player to solve problems!”

Jam & Tea Studios also reports they had to deliberately make the AI dumber. Left alone, NPCs would self-organize and solve their own problems, leaving players with nothing to do. It’s the same lesson Will Wright learned with the original Sims: when the AI always picks the optimal action, player control feels pointless. Wright’s solution was to have Sims choose randomly from their top four options. Retail Mage arrived at a similar insight from the opposite direction: NPCs observed players bunny-hopping around, and concluded something was seriously wrong with these human player characters.

Category: Narrative Experiences

The final category is storytelling — the most ambitious and the most challenging to build. Can AI deliver meaningful narrative experiences?

Whispers from the Star

Creator: Anuttacon (Cai Haoyu) Launch: 2025 Platforms: Steam

Gameplay: What Do You Do? What Does AI Enable?

Remember Lifeline? The 2015 mobile game where you texted a stranded astronaut on an alien moon, with replies arriving throughout your real day? It was the number-one paid app in 55 countries. Whispers from the Star takes that concept and adds incredible graphics plus real-time AI conversation.

You’re in conversation with Stella, a 19-year-old astrophysics student stranded alone on an alien planet. You interact through animated video calls with lip sync and walkie-talkie-style voice messages. Stella doesn’t interact only when you want; messages arrive throughout the day. There are 25+ story nodes over multiple real-time days, with players limited to 40–60 minutes per day (to manage costs).

Player Response

Whispers from the Star immediately hit the top of Steam’s “New & Trending” with 82% positive reviews from over 1,500 players. Players loved the experience. The moments that land hardest are when Stella pushes back. One player spent 30 minutes trying to convince her he was talking through a microwave. She didn’t believe him and got angry. He doubled down, insisting she was a sentient microwave. She got furious and escaped on her own.

But there’s a tension: concurrent players dropped from 964 at peak to 21 in two months. Players had a beautiful, meaningful experience, finished, and never came back. The system does a good job of maintaining boundaries, but the funniest, most viral moments were players playing with those boundaries — and it was the boundary-testing, not the story itself, that went viral. That doesn’t bode well for retention.

A top-rated Steam review raised concerns about privacy policies, echoing AI Dungeon’s problems. The lesson: be clear about privacy expectations; don’t just bury it in your TOS.

Business

Developed by Anuttacon with a custom-trained LLM that could handle huge traffic spikes. But the business tension was never resolved: AI means ongoing per-session costs. Every minute a player plays costs the developer money.

Learnings

The biggest lesson: conversational freedom does not equal narrative freedom. Despite being able to say anything to Stella, the story always moved in roughly the same direction. One of the most upvoted Steam reviews reads: “There are only two actual decisions that affect the final outcome.” This is precisely why we haven’t seen a flood of RPGs with conversational NPCs. The conversation can be open while the narrative remains on rails.

Hidden Door

Creators: Hilary Mason & Matt Brandwein Founded: 2020 Platforms: Web (public launch August 2025)

Gameplay: What Do You Do? What Does AI Enable?

Hidden Door takes a very different approach to narrative gameplay. Players can explore literary properties — Pride and Prejudice, The Wizard of Oz, The Call of Cthulhu, The Crow — and enter these stories to change them. You choose a base fiction, then choose elements to mix in: Pride and Prejudice, but set in Brooklyn, with vampires. Or a Call of Cthulhu spaghetti western. The process works seamlessly.

In-game, it superficially resembles traditional interactive fiction, except there are many more options. The game shows three choices, but as you type, matching options appear. Crucially, these options are authored, as are the responses, which is how safety is guaranteed.

Player Response

Hidden Door launched publicly in August 2025 and hasn’t scaled yet, so it’s still early. But the initial reception has been positive. Developer Ian Bicking identified a core challenge early on: “narrative ungroundedness.” The story existed only as far as it had been written, and players could feel it. The team has improved this significantly, but the current challenge is calibrating how much control players should have. Some expect AI Dungeon-level freedom, which doesn’t match the platform’s safety ethos.

Business

Founded by Hilary Mason and Matt Brandwein in 2020, Hidden Door raised $9 million between pre-seed and seed. The architecture is the big differentiator: a proprietary story engine, not an LLM wrapper. They’ve broken story elements down into tens of thousands of manually curated tropes. Because the components are curated, they have much stronger guarantees about what is or isn’t produced during gameplay.

Learnings

The important and interesting takeaway: strong author and designer control enables Hidden Door to partner with known authors and IP holders. As Eleanor put it, drawing on her experience as a former Disney executive: “No one will hand you their beloved characters unless you can guarantee they won’t say and do horrible things.” Licensed IP is a powerful lever for games, and Hidden Door is the first studio positioned to use it in the AI-native space.

AI2U: With You ’Til The End

Creator: Independent developer Launch: 2025 Platforms: Steam ($14.99, free demo)

Gameplay: What Do You Do? What Does AI Enable?

AI2U is a psychological horror escape room with AI companion personalities, and it’s unlike anything else on this list. You wake up trapped in the domain of an AI-powered yandere, an obsessively possessive girlfriend. Your goal: escape. Your obstacle: she’s desperately in love with you and will resort to violence if she thinks you’re leaving.

You type or speak via microphone. You explore rooms, find items, solve environmental puzzles, and you must keep talking to the NPC. Ignoring her too long makes her anxious, then angry, then murderous. You build trust through gifts, hobbies like cooking and alchemy, and conversation. Trust unlocks access to rooms and eventually escape routes. Each level has multiple endings: escape alone, escape together, or get killed.

There are now four companions, from a classic yandere catgirl to a witch, an AI hologram on a crumbling spaceship, and a siren. They each have distinct personality and escalation patterns. The tonal whiplash is the game’s signature: you’re in a pastel apartment, cooking a meal with your catgirl girlfriend, chatting about her day. Then you try to leave, and suddenly she’s chasing you with a knife, eyes turned red.

Player Response

AI2U is the breakout viral success story of 2025. The predecessor, Yandere AI Girlfriend Simulator, had over 900,000 downloads and 300 million YouTube views. AI2U itself has a 90% positive Steam rating from roughly 1,300 reviews — arguably the best-reviewed AI game on the platform. The viral loop is self-reinforcing: a horror moment happens, a streamer clips it, a viewer downloads the free demo, they have their own horror moment, and the cycle continues.

On the experience side, the rough edges are real. The word filter is comically aggressive: asking for “butter” gets bleeped because of “butt.” The word “amusement” is flagged because of “semen,” and “assistance” catches “ass.” The AI sometimes fixates on random topics or hallucinates game state. The constant attention requirement becomes tedious over long sessions, and there’s no save system during multi-hour runs.

Business

Priced at $14.99 with all art hand-drawn. One critical business decision stands out: the team eliminated the need for player API keys by moving to server-side AI calls. Previous AI games had sometimes required players to bring their own OpenAI API key, which was a massive friction point. You simply cannot get nearly a million people to download a demo if the first step is “go create an OpenAI account and enter your credit card.”

Learnings

AI unpredictability, which plagues almost every other genre, is a perfect fit for horror. Hallucination? Creepy. Inconsistency? Unsettling. Unexpected responses? Terrifying. Every weakness of current AI becomes a strength when your genre is designed around dread.

The broader lesson: pick a genre where AI’s weaknesses become strengths. Not every game needs to solve the consistency problem! Some games can make that problem the point. And on the practical side, removing the API key requirement was critical to reaching a mass audience. Reduce friction ruthlessly.

The AI Companion Explosion (2020 — present)

Before predicting the future, we needed to zoom out from individual games to a larger market.

Character.AI created an entire AI companion category. As of early 2026, we’re tracking over 30 chat, companion, and roleplay apps, with demographics typically similar to Character.AI and comparable or better engagement numbers.

Based on Sensor Tower data, these apps conservatively hit $20–25 million per month in revenue — likely more when including web-only platforms and Chinese domestic apps.

But the category has the same safety issues as Character.AI, magnified across dozens of less-regulated platforms. A Belgian investigation was launched into a player suicide linked to Chai. Italy banned Replika and levied a €5 million fine. Botify was found to have bots closely resembling underage celebrities. Multiple apps have been pulled from app stores.

And one more uncomfortable truth: many top earners differentiate on unrestricted explicit content that Character.AI doesn’t generally provide. The business model that works best carries the highest risk.

Where This Is All Going

We see four forces shaping what comes next.

First, regulation arrives. Content liability will be clarified through laws and lawsuits, reshaping what’s allowable — especially regarding teens. In our view, this is good news! Clearer rules can accelerate responsible development.

Second, a Cambrian explosion of AI-native games, but only when local inference becomes broadly available. Small models are already good. When local inference makes AI effectively free, it will be everywhere.

Third, AAA studios wade in cautiously. We’ll see AI NPCs and AI-assisted features in traditional games, but the major studios won’t go fully AI-native yet. They’ll be watching very carefully.

Fourth, and most exciting: we’re in the first year of AI-native games being out in the market and being real. As players, we’re able to immerse ourselves in worlds in new ways. As designers, we have to design at a different level of interaction than we’re accustomed to, designing systems and frameworks, not necessarily specific details. We’re inventing this, as a community.

Key Takeaways

Three patterns show up again and again in the games that worked.

First, constrained creativity wins. Give players boundaries to push against, not an empty void. Suck Up! and Retail Mage both thrive because players have a clear goal — convince the NPC, run the shop — and the AI opens up infinite ways to pursue it.

Second, design explicitly for virality. Unpredictable AI creates streaming content naturally. Games like Suck Up! and AI2U were practically built to generate shareable moments.

Third, pair familiar interfaces with novel mechanics. Status gave people a Twitter interface they already knew how to use and filled it with AI characters. Character.AI gave people a messaging app and populated it with personalities. The success stories didn’t just add AI to a game, but found places where AI can help players create something genuinely new.

Finally, three major conclusions from the field as a whole:

AI-native games can reach millions of players with extraordinary engagement. The demand is real and proven.

Success is hard in this space. You need a game that can become a hobby, an economic model that works for both player and business, and probably reach great virality.

The common success factor is player expression, expressed through in-game agency and creator community.

These are exciting times. We’re early enough to pioneer, and late enough to learn from others’ very expensive mistakes.

We are in this magical moment of possibility. There are a lot of good, and not good, possibilities ahead of us. So please, go make the good ones.

Presented at GDC 2026, San Francisco, California, March 9–13, 2026

Based on publicly available information, plus our own analysis and impressions.