Does GTA 6 use AI? What Take-Two's own filings say

Does GTA 6 use AI? What Take-Two's own filings say

Mihai Perdum
Mihai Perdum
Author

11 min readNovember 22, 2026

Key takeaways

  • Take-Two CEO Strauss Zelnick said on 3 February 2026 that generative AI has zero part in what Rockstar is building for GTA VI. In the same interview he said the company runs hundreds of AI pilots, studios included.
  • We counted the phrase artificial intelligence in twelve Take-Two annual reports. It appears 0 times from fiscal 2015 to 2023, then 2, 4 and 7 times in 2024, 2025 and 2026, mostly in the risk factors.
  • The fiscal 2026 report warns that staff may paste confidential or regulated information into third-party AI services. Any team with an AI assistant next to its Jira and Confluence carries the same risk.
  • None of the three Take-Two patents on character movement and road pathfinding names GTA VI, and none claims machine learning. The locomotion patent's text has 0 hits for machine learning or neural network.
  • The post saying Claude Code finished GTA 6 in 12 hours came from an account created 52 days earlier. It names no model and links no log, and the game shipped only on PS5 and Xbox Series X|S.

Does Grand Theft Auto (GTA) 6 use AI? Not generative AI, says its publisher. Nothing we found contradicts that, but the denial covers generative AI in this one game. Take-Two CEO Strauss Zelnick told GamesIndustry.biz in February that generative AI "has zero part in what Rockstar Games is building". In the same breath he said the company's products "have always been built with machine learning and artificial intelligence". Take-Two's own annual reports used the phrase "artificial intelligence" 0 times from fiscal 2015 to 2023 and 7 times in fiscal 2026, by our count of twelve filings. Claude Code is Anthropic's AI coding agent, and it can also control a computer's screen. A viral post says it finished the game in twelve hours. The post comes from a seven-week-old account with no log, about a game that is not on PC.

Grand Theft Auto VI (the sixth main game, GTA 6 for short) came out on 19 November 2026 on PlayStation 5 (PS5) and Xbox Series X|S. On launch day its English Wikipedia page had 750,227 views, more than any other article that day (only the Main Page and the search page ranked higher). On 10 November it had 64,375.

I run a small consultancy that builds Atlassian apps and AI tooling. So I care less about the game than about how a publisher this size talks about AI to investors, in documents its lawyers sign. I read twelve of Take-Two's annual reports and three of its patents, and then I traced the viral post back to its source. A lot of it applies to any team putting AI into its own work.

Three things called AI in this story

The question mixes up three things. The answers differ.

  • Generative AI produces content: text, images, voices, 3D models. This is what Take-Two says is not in GTA VI.
  • Machine learning is any model trained on data, from an upscaler to a spam filter. Take-Two says its products have always used it, without naming where.
  • Game AI is the old meaning: rules that designers write so that a pedestrian runs from a gunshot or a police car follows you. Every open-world game has it. None of it has to be trained.

Four findings: one game, three filings, no patents, no evidence

  1. The "no generative AI" answer covers one game. The denial begins "Specifically with regards to GTA 6" (section below).
  2. AI went from zero to a risk factor in three annual reports. None of the reports from fiscal 2015 to 2023 uses "artificial intelligence"; fiscal 2024, 2025 and 2026 use it 2, 4 and 7 times.
  3. The patents fans call "GTA 6 AI patents" never mention GTA 6 and claim no machine learning.
  4. The Claude Code story has no evidence behind it. It traces back to one post, and a near-identical post from another account appeared the day before.

Does GTA 6 use generative AI? Take-Two says no, for this game

The main source is Zelnick's interview with GamesIndustry.biz, published on 3 February 2026:

"Specifically with regards to GTA 6, Generative AI has zero part in what Rockstar Games is building. Their worlds are handcrafted. That's what differentiates them. They're built from the ground up, building by building, street by street, neighborhood by neighborhood. They're not procedurally generated, they shouldn't be."

The denial starts with a qualifier. So it covers one game. The timing matters too. GamesIndustry.biz notes the interview came after a week in which Take-Two's stock fell to an 11-month low following Google's announcement of Project Genie, which GamesIndustry.biz calls an experimental AI "playable world" tool. On the earnings call the same day, Take-Two president Karl Slatoff said Genie "is not a game engine". To him it looked "more like a procedurally generated interactive video at this point" (transcript).

Rockstar's own answer came second-hand. IGN reported in August 2026 that Kinda Funny's Greg Miller asked Rockstar "point-blank" about generative AI in GTA 6. The team told him there is "no AI" used in the game. That is a journalist's account of a verbal answer at a preview event, and I found no written Rockstar statement saying the same.

Did GTA 6 use AI in development? Take-Two says machine learning, without saying where

The same interview has the line most headlines skipped: "This company's products have always been built with machine learning and artificial intelligence ... right now we have hundreds of pilots and implementations across our company, including within our studios. And we're already seeing instances where generative AI tools are driving costs and time efficiencies." On the earnings call he said generative AI "is already moving into the category of efficiency".

So the company's line is simple: AI goes into the tools that build the game, and the game itself, the streets and the people in them, is made by hand. It never names the tools. It never says whether Rockstar is one of the studios running pilots either, and I found nothing public that does.

The voice actors are part of the same picture. The Screen Actors Guild (SAG) and the American Federation of Television and Radio Artists (AFTRA) merged into one US actors' union, SAG-AFTRA. It struck video game makers in 2024-25 over "fair compensation and the right of informed consent for the AI use" of performers. A rule from an older agreement exempted games that started production more than a year before the strike, and GTA VI was one of them. The union's chief contracts officer, Ray Rodriguez, called that legacy rule "insane, but it's there" (PC Gamer). The deal that ended the 320-day strike passed with 95.04% of votes and requires consent and disclosure for digital replicas (TheWrap). No source says which agreement GTA VI's performers worked under.

Does GTA 6 use AI upscaling? Yes, on PS5 Pro and Xbox Series S

I think upscaling is the most concrete answer to the question. It is the one place a neural network is publicly known to touch the game, the step that turns a lower-resolution frame into a sharper one. On PS5 Pro, Rockstar says it uses Sony's PSSR (short for PlayStation Spectral Super Resolution) in its second version, PSSR2 (as Rockstar writes it), to upscale 1440p to 4K. Sony's Mark Cerny describes it as "an AI library that analyzes game images pixel by pixel", built with AMD (the chipmaker behind the console's processor). Digital Foundry's tests, as reported by Dexerto, found the PSSR output on the Pro "convincingly passes for native 4K".

A one-eyed red and green robot holds a magnifying glass up to a big screen where a blocky, pixelated sunflower on the left becomes a sharp sunflower on the right, with small chips marked with stars under each, on a cobalt-blue background.
Illustration, generated locally with ERNIE-Image. On PS5 Pro, Rockstar says it upscales with Sony's PSSR2, a neural network that turns a lower-resolution frame into a sharper one.

Rockstar's technology leads said in a statement to Wccftech that the Xbox Series S version starts at 720p and reaches 1080p "using AMD FSR Upscaling 4". FSR, short for FidelityFX Super Resolution, is AMD's upscaler. AMD calls it "ML-based FSR Upscaling" (ML, short for machine learning) and says FSR 4 is its earlier name. On PS5 and Series X, Rockstar upscales 1440p to 4K with its own TAA (temporal anti-aliasing) upscaler, which it calls its "in-house TAA solution". It does not mention machine learning for that one.

Take-Two's 10-K reports: from zero to a risk factor in three years

Interviews are what a CEO wants you to hear. A 10-K is the annual report a US-listed company files with the SEC (the US Securities and Exchange Commission). Its risk factors section lists what the company's lawyers think could hurt the business. So we downloaded twelve of them from EDGAR, the SEC's filing database, for fiscal 2015 to fiscal 2026. Take-Two's fiscal year ends on 31 March. A 20-line Python script strips the HTML and counts the terms. The part that matters is the patterns:

python
1terms = {"artificial intelligence": r"artificial intelligence",
2         "generative AI": r"generative (?:AI|artificial)",
3         "machine learning": r"machine learning",
4         "AI (word)": r"\bAI\b"}   # \b = whole word, so "said" and "AIM" never match; AI must be upper case

The other three patterns ignore case.

fiscal year

words in the 10-K

"artificial intelligence"

"AI" as a word

"machine learning"

2015 to 2023 (nine reports)

51,874 to 69,424

0

0

0

2024 (filed 22 May 2024)

71,076

2

5

0

2025 (filed 20 May 2025)

70,918

4

14

0

2026 (filed 22 May 2026)

68,257

7

24

0

The jump came in 2024, but the nine reports with no mention at all say more to me than the seven mentions in 2026: until fiscal 2024, Take-Two did not name AI to investors anywhere in its annual report. The reports stay between 51,874 and 71,076 words. Length did not drive the count. And the term was not always absent. EDGAR's full-text search finds "artificial intelligence" in two much older Take-Two reports, for fiscal 2005 and 2006. There, "expanded artificial intelligence capabilities of new platforms" is one of the things that would make people buy new consoles. About twenty years ago AI was a selling point of the hardware. From fiscal 2024 it sits mainly in the risk factors, and the count has grown every year since.

The 2024 version was short. AI "may present operational and reputational risks", the law on intellectual property around AI is unsettled, and third-party AI could bring copyright claims. By 2026 it was longer. The heading reads "The development, use, and incorporation of artificial intelligence ("AI") into our products and within our industry may present operational, reputational, financial, and competition risks." It adds "uncertainty regarding copyright ownership of AI-generated assets". An older sentence about competitors imitating Take-Two's best-selling games gained one clause in 2026: "including by using AI to do so".

I'd put this sentence in front of any team rolling out AI assistants, in games or anywhere else, because I think it names the likeliest way an AI tool ends up hurting a company:

"our employees, contractors, vendors, or other partners may use AI tools in ways that are inconsistent with our policies or expectations, including by entering confidential, proprietary, personal, or regulated information into third-party AI services, and some AI providers may have limited operating histories or governance processes"

The same filing says Take-Two is "investing in artificial intelligence ("AI") tools and technologies" to "enhance game play". So the 2026 report holds both positions at once: invest in AI, and warn that your own people may leak data into it. A risk factor says what "may" happen. This sentence reports no incident.

GTA 6 NPC AI: 600,000 animations and patents that never name the game

Search "gta 6 ai npc patent" and you get fan theories. Each NPC (non-player character, one of the people in the street) is where game AI lives. Here is what is on the record.

Edge's cover story, quoted by PC Gamer, says Rockstar set up a Los Angeles team of writers and directors for its crowds. The studio "art-directed every [NPC], creating around 600,000 animations". Each pedestrian gets personality traits that help decide how it reacts: temperament, aggression, where it is, what it has seen and how you treated it. That reads like classic, hand-authored game AI. No source says how the traits work. Nobody says whether any of it is trained.

The patents are Take-Two's. I read their claims (the numbered parts that say what a patent protects).

US11620781B1, "System and method for virtual character locomotion", was filed on 23 October 2020 and granted on 4 April 2023. Claim 1 describes "motion animation blocks" and "archetype blocks" that are "functionally independent from the core game logic". Claim 7 says the blocks can be defined in XML meta files. I searched the page text for "machine learning", "neural network" and "trained model": 0 hits.

US11071916B2, "System and method for virtual navigation in a gaming environment", was filed on 24 April 2019 and granted on 27 July 2021. It builds a coarse graph from the detailed road network and plans a route on it, so the game loads fewer detailed nodes. US11684855B2 is a later patent on the same navigation idea, granted on 27 June 2023.

None of them names GTA VI or any other game. I don't think they prove anything about it. Kotaku described the locomotion patent's first-listed inventor, Tobias Kleanthous, as a former Rockstar lead AI and gameplay programmer, and said it had asked Rockstar for comment. The article prints no reply.

Take-Two AI: the layoffs and the decompilation worry

Two months after the "zero part" interview, Game Developer reported on 3 April 2026 that Take-Two had laid off its head of AI, Luke Dicken (in the role since January 2025), and at least four others from the AI team. Take-Two declined to comment. So the "hundreds of pilots" interview and the end of its central AI team are two months apart, and I can't tell how the two fit. I checked the 10-K for that fiscal year, filed seven weeks after the layoff report, and none of its AI sentences mentions the team or the cuts.

Ten days before launch, on the 9 November call, an analyst asked about AI-assisted reverse engineering and decompilation of games. Zelnick said "it's going to become more of a challenge" and that the company is "utterly focused on enhancing our technical defenses and enhancing our legal activities" (transcript). Rockstar has paid for leaks before: it said the 2022 hack that leaked unfinished GTA VI footage cost $5 million to recover from, plus thousands of hours of staff time (BBC).

Claude Code and GTA 6: the 12-hour claim, checked

On 19 November, launch day, the X account @0xGoldTooth posted "Claude Code just finished GTA 6 in 12 hours while I was at work". The post says it "learned the controls in 15 minutes" and went "through the game files on its own". GameGPU wrote it up as news. The post had 3,763,124 views when we read it on 21 November.

I checked what I could:

  • I did not watch the attached video or read the replies, which need an X login.
  • The account is 52 days older than the post. It joined X on 28 September 2026. Its bio reads "AI anomalies & deep web leaks".
  • The post names no model or platform. It links no log or code. Nothing else backs it.
  • X's crowd-written fact-check on it (a Community Note) says "Grand Theft Auto VI is not available on PC at this time". Zelnick said on the 9 November call that the game launches on "2 consoles".
  • A day earlier, @starmexxx posted "Claude Code just played GTA 6 for 8 hours while i was at work ... it learned the controls in about 15 minutes ... i didn't write any code or set up any scripts", with the same structure and phrases.
  • Gary Marcus quote-posted the @0xGoldTooth post on 20 November: "this is fake. AI can't actually do this yet."

Then there is the tool itself. Computer use is the Claude feature that looks at the screen and clicks and types. Anthropic's docs call it a research preview: in the Claude Code CLI it runs only on macOS, and the Desktop app adds Windows. Computer use "is the broadest and slowest" way for Claude to act, so it tries more precise tools first. Every screenshot is downscaled before the model sees it (a 3456×2234 screen goes to about 1372×887). Sony's Remote Play does stream a PS5 to a Mac, but Sony lists a DualSense controller among the things you need to play, and its on-screen panel only gives the PlayStation, Create and Options buttons, while computer use clicks and types. Once connected, Sony says, "the screen of your PS5 console displays on your Mac". Nothing on that page lets the Mac open the console's files. Tom's Hardware notes that PC players can stream the game through Xbox Cloud Gaming. The post does not say it used either service. The claim is a twelve-hour finish of a game rated 99 on Metacritic plus a trawl through "the game files", from a new account. I found no log anywhere. I would not repeat it as news, and I'm surprised GameGPU did.

A blue robot types on a laptop whose screen shows a blurry pixelated red car on a grey road, while a snail crawls past on the desk and a padlocked black treasure chest sits on a shelf behind, its cable hanging loose, on a violet background.
Illustration, generated locally with ERNIE-Image. Over Remote Play, Sony says the screen of the PS5 console displays on the Mac.

The documented Claude-and-GTA story is smaller and more useful. In June, Gizmodo covered Ziwen Xu, a 25-year-old founder who tried to build a GTA-style game by prompting an AI agent instead of coding it himself. He upgraded to Claude Max 20x, an Anthropic subscription, to "beat the real GTA 6 to launch". On day two he wrote that the agent built downtown Los Angeles skyscrapers, "which is a problem, because this is supposed to be Florida". He had also burned 33% of his weekly usage in one day. If you want measured numbers on what these agents finish, our agentic coding benchmark results are where we keep them.

What GTA 6 tells teams that ship with AI tools

Take-Two split AI use in two: tools that make the work faster, and content that ships to the customer. Valve drew the same line on Steam. Valve's disclosure rules ask about "Pre-Generated" content made with AI during development and "Live-Generated" content made while the game runs. The January 2026 update says "efficiency gains through the use of these tools is not the focus". Activision's Steam page for Call of Duty is an example of a Pre-Generated disclosure: "Our team uses generative AI tools to help develop some in game assets."

I think it's the right split, and it works as a policy for a software team too. AI that drafts the Jira ticket, summarises the Confluence page or writes the first version of a test is a tooling decision, while AI output that reaches a customer (release notes, a support answer, code in the product) needs the review and disclosure you would want from a contractor. Write the two lists down. Keep them separate and most of the arguments go away.

The people side is not settled, even in games. GDC's 2026 State of the Game Industry asked more than 2,300 professionals. 52% think generative AI is hurting the industry, up from 18% two years earlier. Still, 36% use it at work. 30% of respondents at game studios use AI tools, against 58% at publishers, support teams and marketing firms.

And the 10-K's data-leak sentence applies just as much to a Jira or Confluence site. Before arguing about whether an AI feature is allowed, check what your plan lets you switch off. I'd start there. We wrote up which Atlassian plans let you opt out of AI training and what stays on when you disable Rovo.

Still unknown: whether any of GTA VI's characters, animation or physics use a trained model, which AI tools Rockstar's own developers use, and whether the patents shipped. Rockstar may answer some of it in a technical talk one day. Has your team written down which of its AI uses are tooling and which ship to customers?

ShareXLinkedIn

Comments

0

Add a comment

Confirm your address using the email we send. You can unsubscribe at any time. Your email stays private and this does not sign you up for the newsletter.