After GTA VI: Will GTA VII Be Built by AI? The Future of AI-Generated Open Worlds

Start ten years ahead.

An NPC you robbed three weeks ago recognises you in a bar and reacts — not with a barked line from a pool of twelve, but with a specific reference to what happened. A side mission generates around the state of your world: the businesses you burned, the people you alienated, the part of the map you never visit. A designer types a description of a coastal town and gets a playable one back.

Every element of that is describable with technology that exists in some form today. None of it is close to shippable in a game like GTA, and the reasons why are considerably more interesting than the demos.

For what is actually confirmed about the game shipping in November, rather than the games that might follow it, start with the GTA VI complete guide.

1. What "AI-built" would actually mean

"AI-generated game" collapses at least five separate capabilities that have almost nothing to do with each other:

These have different economics, different failure modes and different legal exposure. A studio could adopt three and reject three. Treating them as one thing is why most "AI will make games" arguments go nowhere.

2. LLM-powered NPCs

The most-discussed and the most misunderstood. A properly generative NPC needs four things, and only the first is easy:

• Dialogue generation — a model producing in-character speech. Solved, in the narrow sense.
• Personality persistence — the same character sounding the same way an hour later, and next week. Hard, because prompts drift.
• Memory — recalling your prior interactions specifically, not vaguely. Hard, and expensive to store per player.
• World-state grounding — knowing what has actually happened in the game so it doesn't invent events. Hardest, and the one demos always skip.

That fourth item is the real barrier. An NPC that hallucinates a bank robbery you never committed hasn't added immersion — it has broken the contract that the world is consistent. And unlike a chatbot, a game cannot say "I may be wrong about that."

3. AI agents: NPCs with goals

This is the direction I find genuinely interesting, and it's a different idea from talking NPCs.

Today, an NPC executes a script:

NPC: follow predefined mission script
     → wait at marker
     → deliver line when player arrives
     → drive to next marker

An agentic NPC would instead pursue a goal:

NPC goal: earn $50,000 before the month ends
        ↓
Observe environment (what's available? who's around?)
        ↓
Plan (rob? deal? borrow? partner up?)
        ↓
Find allies — including, possibly, the player
        ↓
Act, and react to what the player does
        ↓
Modify plan when it fails
        ↓
Remember the outcome, and who helped or betrayed
        ↓
        └──────── loop ────────┘

That's an agent living inside Vice City — and it's the same architecture pattern now being built for the web, where agents observe, plan, act and remember against real sites. I've written about the web-facing version in the agent-readiness guide; the loop is identical, only the environment changes.

Why this is more promising than chat: goals and planning can run on cheap symbolic systems. Goal-oriented action planning has been in shipped games since F.E.A.R. in 2005. You don't need an LLM to give an NPC a goal — you need one only if you want the NPC to explain the goal in natural language. Split those, and most of the value arrives without most of the cost.

Why it's still hard: emergent systems are hostile to authored narrative. If an NPC can genuinely fail at their goal, your story has branches you didn't write. Most studios respond by constraining the agent until it's a scripted sequence again.

4. Generated missions and procedural narrative

Procedurally generated quests have existed for years, and they have a reputation problem earned honestly: they tend to be structurally identical with different nouns. Go here, kill that, return.

What generative systems could add is contextual specificity — a mission built around your actual history rather than a template. The interesting version isn't "generate a mission", it's "generate the reason for a mission, from a world state that already exists."

The unsolved part is quality control. A designed mission is playtested. A generated one is played for the first time by the customer. For a studio whose reputation rests on mission design, shipping an unplaytested experience is a strange trade — and the failure isn't a crash, it's tedium, which no automated test detects.

5. Voice, animation and text-to-3D

Voice synthesis

Technically the most mature item on the list and legally the most contested. Performer unions have negotiated specifically over synthetic replication, consent and compensation. The technology being ready is not the constraint; the rights framework is. Expect this to be settled by contract law rather than by capability.

Animation

Machine learning is already normal here — motion matching against large captured databases is standard in modern games, and learned motion models are an active research area. This is a case where "AI in games" is already true and uncontroversial, because it operates on captured human performance rather than replacing it.

Text-to-3D

Useful for concepting, weak for shipping. Generated meshes tend to arrive with topology no animator wants, UVs that don't unwrap cleanly, and geometry budgets nobody asked for. The realistic near-term use is a starting point an artist rebuilds, which saves less time than it appears to.

6. World models

The most spectacular demos and the biggest gap between demo and pipeline.

The promise looks like this:

Prompt: "A Florida-style coastal town with a marina,
         a nightlife district, and hurricane damage."
        ↓
Terrain
        ↓
Buildings and streets
        ↓
Road network and navigation
        ↓
NPC population and businesses
        ↓
Physics and collision
        ↓
Stories and conflicts
        ↓
Playable environment

Note what happens as you move down that stack. The top is geometry — plausible to generate. Halfway down you hit systems: navigation meshes that must be traversable, collision that must be watertight, businesses that must be enterable, an economy that must be coherent. By the bottom you're generating meaning, and meaning is what the player is buying.

There's also a market truth that gets lost in the excitement: a plausible town is not an interesting one. Vice City is not compelling because it resembles Miami. It's compelling because of a thousand specific authored choices about what to exaggerate. That is the exact thing Take-Two said differentiates Rockstar — and it was said in response to a world-generation demo. Confirmed

7. Could one person eventually build a GTA-sized world?

Probably, for some definition of "GTA-sized". Almost certainly not for "GTA-quality".

Scale and quality are different constraints. Generating a hundred square kilometres of plausible terrain and buildings is a compute problem, and compute problems tend to yield. Making a hundred square kilometres worth exploring is an authorship problem — thousands of decisions about what's funny, what's surprising, what rewards a detour.

Where I'd expect this to land is not "one person makes GTA" but a small team makes something that would previously have needed sixty people. That's the historically normal shape of a tools revolution: the floor rises, the ceiling stays where it is, and the number of viable studios goes up. Game engines did exactly this in the 2010s.

8. The problems nobody has solved

Any honest version of this article spends as long here as on the possibilities.

Rights and law

• Copyright. Training-data provenance remains contested. A studio shipping generated assets inherits whatever the courts eventually decide.
• Artist rights. Whose work trained the model that replaced the concept artist?
• Voice rights. Actively negotiated, not settled. Synthetic replication of a performer's voice is a labour dispute with real leverage behind it.

Economics and engineering

• Compute cost. Per-player inference against a one-time purchase price. Either the model runs locally — competing with rendering for hardware that is already fully committed — or it runs remotely, at a recurring cost per player.
• Latency. Conversation is real-time. Remote inference introduces a delay players read as broken.
• Offline play. A single-player game that requires a connection to talk to NPCs has changed what it is.

Quality and safety

• Moderation. A system that can say anything eventually says something quotable in a headline.
• Ratings. An ESRB or PEGI rating is granted against content the publisher can enumerate. Runtime generation is structurally hard to certify — I'd argue this is the single hardest obstacle to shipping generative dialogue in a mainstream retail game.
• Player safety. Open-ended systems in the hands of a large audience get pushed toward every boundary that exists.
• Consistency and hallucination. An NPC describing a mission that doesn't exist wastes the player's evening.
• Bugs. You cannot QA an infinite output space. Testing shifts from verification to sampling, which is a different discipline with weaker guarantees.

The creative risk

• Homogenisation. Models trained on existing games produce content resembling existing games. An industry that adopts them broadly converges on a shared average — the opposite of what makes any of these worlds worth visiting.

9. What I actually expect

My honest forecast, clearly labelled as forecast: Speculation

The pattern in the left column: production tooling, invisible to the player, verifiable output. The right column: things the player experiences directly, where a failure is a bad review rather than a slow sprint.

Which is exactly the line Take-Two drew for GTA VI — automate the process, not the advantage. The forward-looking question isn't whether AI enters game development. It already has. It's whether the line moves, and how far.

10. Frequently asked questions

Work with me

The agentic shift is already happening on the web

I'm Jayant Solanki — an SEO, GEO and automation strategist. The agent loop in section 3 — observe, plan, act, remember — isn't hypothetical outside games. It's already navigating websites, comparing products and completing checkouts on customers' behalf.

What that means for a site today:

• Semantic markup and an accessibility tree an agent can actually read
• Bot and WAF rules that don't block your own customers' assistants
• Product data that agrees across feed, schema and page
• Measurement that separates agent sessions from human ones

Book a technical review → Is your site agent-ready?

Ranked #1 for "metal buildings" · +30% YoY organic traffic · Evidence-labelled research

Sources

• TechSpot, Take-Two CEO on generative AI having "zero part" in GTA 6 (4 February 2026)
• The Next Web, GTA 6 uses zero generative AI, Take-Two CEO confirms
• Notebookcheck, Take-Two stock falls after Project Genie generative AI tool debuts — context for the statement
• web.dev, Build agent-friendly websites — the same agent loop, applied to the web

Jayant Solanki

AI-Ready SEO, GEO & AIO strategist based in Indore, India. Part of a GTA VI cluster written to a stricter evidence standard than most games coverage — with forecasts labelled as forecasts.

More about Jayant →