Game studios have bought into AI, but players still aren’t sold

(By Caixin journalist Guan Cong)

In early August, sweltering heat in Shanghai did little to keep visitors away from ChinaJoy, China’s largest gaming expo. Crowds packed the halls, queueing for demos, collecting merchandise and posing with cosplayers. Most were in their 20s. Their enthusiasm, however, masked a fundamental question the industry can no longer avoid.

For the first time in its 23-year history, ChinaJoy adopted an artificial intelligence (AI) theme: “Play with AI”. Among the 1,000 games on display, a dedicated zone featuring 70 AI-powered titles drew broad attention. Yet reactions were mixed. Some players were impressed by AI features that exceeded expectations; others came away underwhelmed by what they saw as a lack of genuine innovation.

“AI only increases the combinatorial outcomes of skills, rather than creating new gameplay,” one developer told Caixin after trying a game with AI-generated abilities.

Over the past year, AI has been rapidly adopted to populate multiplayer games with more intelligent bots and lifelike virtual teammates, helping extend user engagement. But keeping players grinding longer is not the breakthrough many in the industry are looking for.

Instead, gaming is entering a generative AI-driven content boom similar to shifts seen in film and television. In the first half of 2026, two editions of Steam Next Fest featured more than 8,400 playable demos, including over 1,000 explicitly labelled as using AI. In China, more than 4,000 of the over 5,000 new games launched on distribution platform TapTap during the same period were AI-generated minigames, contributing to a tenfold year-on-year expansion in new titles.

That surge in low-cost creation, however, has also triggered a crisis of trust among players, echoing earlier periods when markets were flooded with low-quality copycats.

Even so, many in the industry remain optimistic. They argue that gaming, with 680 million users in China and a mature content ecosystem, is one of the most promising proving grounds for real-world AI applications. In that view, foundational models combining generative and interactive capabilities could become essential infrastructure for the sector.

Defining the AI game

China’s gaming industry has been given a policy boost after being recognised as an independent cultural product in the country’s 15th Five-Year Plan. In 2026, AI-focused games began receiving official publishing licences, drawing fresh venture capital interest.

An emerging industry test for whether a game is truly “AI-native” is straightforward: remove the AI and see whether the core gameplay still works.

One example is the AI-driven detective game Sherlock Holmes: Night Stalker, which received a publishing licence in June. The game has no fixed narrative branches. Instead, players uncover clues through interactions with AI-powered non-player characters, or NPCs.

“Some NPCs really like hearing nice things, while at times you may need a tougher tone to get help,” Wu Yunyun, co-founder of developer Gamercury AI, told Caixin, adding that the AI ultimately shapes the outcome. During testing on Steam, players stretched a 40-minute demo into hours by repeatedly talking with NPCs to draw out clues. The game uses ByteDance’s Doubao and Google’s Gemini large language models, along with MiniMax for voice generation.

Wu said the 15-person team spent a year and a half taking the game from development to completion, with nearly a year of that devoted to securing a publishing licence. Regulators have taken a more cautious approach toward emerging AI-native titles, he said. Reviews now explicitly examine how a game behaves once connected to model application programming interfaces (APIs) and how its NPCs respond dynamically.

Integrating AI models is also reshaping in-game economics. Interactions with NPCs consume tokens. Pricing is set around the expected token use of a standard playthrough, while players who exceed their free allocation can buy additional bundles.

Some creators are leveraging “one-sentence game generation” engines. In late March, 25-year-old developer Wu Tong used his proprietary text-based AI engine, FunloomAI, to produce three distinct games. Without marketing, one title, Chongzhen Simulator, attracted 25,000 players organically. The performance helped him secure tens of millions of RMB in Pre-A funding in May, valuing his startup at 200 million RMB (US$29.6 million).

“Past technological leaps mostly improved games visually, with ever-higher rendering quality,” Wu said. “The real value AI brings is freedom and the democratisation of creation. That points in a very different direction from the traditional idea of improving game quality.”

The innovation dilemma

Despite such indie successes, investors remain cautious. Truly AI-native projects are still scarce, and expectations for AI in gaming remain lower than the surrounding hype might suggest. Yoyo Yang, an investor at Makers Fund and a former Bandai Namco executive, said technology may accelerate supply, but player attention and spending power remain limited.

The mismatch reflects a broader problem in the global games market: faster, cheaper production tools are filling platforms with derivative content rather than producing the high-quality intellectual property many players increasingly want.

Large studios, facing slower growth, are already adjusting away from the traditional logic of high-cost, long-cycle development. In the first half of 2026, major companies including Epic Games, Sony Interactive Entertainment and Tencent Holdings shut several studios, including some focused on big-budget AAA titles. In July, Microsoft’s Xbox division laid off more than 3,000 employees and spun off four teams after four straight quarters of declining revenue.

In China, the mobile-heavy gaming market generated 350 billion RMB in 2025, but growth is still led by top legacy titles. Overall user growth has hovered at about 1% for three consecutive years. Players have become more selective, while their attention has been fragmented by minigames and short-form video on social platforms.

Developers capable of producing truly compelling games will remain rare, regardless of technological progress, Yoyo Yang said.

Yang Wenfeng, a gaming industry veteran and founder of dollar fund BBX Ventures, said AI could eventually create value in two main ways: by increasing playtime through more varied generative experiences, and by using player data to tailor difficulty and play style. Even so, he said many current visions of world models in gaming miss a basic point: compelling design depends on human understanding of players, not just more efficient tools.

Enter world models

To push the frontier further, technology companies are betting on world models — AI systems designed to simulate the physical world. In early 2026, startups founded by AI scholars Fei-Fei Li and Yann LeCun each secured US$1 billion in funding, raising expectations that simulated environments could help advance artificial general intelligence.

Chinese startups are also moving quickly into the field. In January, Beijing-based AISphere released its real-time world model, PixVerse R1, allowing users to interact with a video stream through natural language and generate subsequent actions continuously. The company is initially targeting simulation and romance games, genres that align well with the model’s ability to instantly generate visuals based on player instructions. In July, AISphere completed a 2.98 billion RMB Series C funding round.

However, traditional game developers remain sceptical of fully AI-generated gameplay. Aishi Technology executive Li Qiushi noted that while AI firms want to create novel, AI-native game categories, established studios prefer to stick with highly profitable, mature formats. Li added that current models struggle to replicate the nuanced interactive “feel” that game studios have spent years perfecting, and they still need to improve instruction response times and memory continuity.

Instead of replacing core gameplay, studios are using AI to streamline asset creation. VAST founder Song Yachen said the gaming industry’s adoption of AI 3D generation accelerated in early 2026 after processing times dropped from weeks to seconds and costs fell to mere cents. VAST focuses on generating 3D assets to establish stable physical rules — such as tracking a bullet’s precise trajectory — that traditional games require, rather than relying on the visual approximations of video-based models.

Major tech companies are also tailoring their AI outputs to fit existing industry workflows. Tencent Holdings Ltd. released its open-source Hunyuan 3D world model in April, allowing developers to import generated assets directly into standard game engines like Unity and Unreal for manual editing.

Despite these advancements, single generative models cannot yet manage the complex engineering of a full video game. Unity China AI executive Gu Shenhua said current world models only produce basic spatial assets and lack the deterministic logic, operational stability, and precise control necessary for modern game development.

Trust and copyright risks

As AI reshapes production, it is also altering the relationship between consumers and content. The technology has already become deeply embedded in development workflows, shifting many programmers’ responsibilities toward prompt design, testing and infrastructure.

At the same time, stronger models are creating new dependencies. Language models still struggle with spatial reasoning in interactive systems and often rely on approximation. Multimodal capability matters because an agent that cannot interpret screenshots or video must depend on another model to “see”, increasing complexity and reducing reliability.

Those changes are driving demand for infrastructure. According to a report from Gamma Data, about 73% of the top 50 Chinese game companies by first-half revenue had external demand for cloud and computing-power partnerships, while 57.6% had already connected to mainstream AI models.

At a campus recruitment event in May, Mihoyo said it planned to invest as much as 100 billion RMB in AI over the next three years, regardless of outcome, with the aim of building highly personalised games tailored to individual users. The company said AI-generated real-time game content could emerge within three years.

Chinese game companies have the financial capacity to make such bets. Of the world’s top 50 listed gaming companies by revenue in 2025, 32 were Chinese, the highest share of any country. Among China’s listed game companies, 62% posted positive operating cash flow that exceeded net profit attributable to shareholders. Many are also investing directly in AI companies: Mihoyo is a shareholder in MiniMax, while NetEase and Tencent participated in DeepSeek’s first funding round.

Outside of China, however, the industry faces severe backlash. Gamers and developers warn that the technology could strip the sector of its creativity. A Game Developers Conference survey showed the proportion of developers viewing AI negatively surged from 18% to 52% between 2024 and 2026.

Rampant cloning justifies these fears. High-quality titles are frequently reverse-engineered and mass-produced using AI, flooding digital storefronts with cheap knockoffs and undermining consumer trust.

Major console platforms are intervening to clean up their ecosystems. Nintendo recently altered its eShop ranking algorithms to prioritise short-term revenue over download volume, aiming to bury artificially boosted clones. Sony has also repeatedly banned developers known for publishing low-quality AI games on the PlayStation Store.

Industry groups are now trying to tackle the intellectual property risks associated with AI training data. Nick Poole, CEO of the UK Interactive Entertainment Association, said the organisation is developing a pilot marketplace to licence creative content for AI models. The initiative aims to ensure creators consent to and are compensated for their work, offering a commercial solution to mounting copyright disputes.

This article was first published by Caixin Global as “In Depth: Game Studios Have Bought Into AI, but Players Still Aren’t Sold”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.

Related: [Big read] Chinese games rake in 100 billion RMB from overseas users | Can Wukong blaze a new trail for the Chinese gaming industry?