Today’s Wonder News covers Sega’s Sonic campaign AI-training terms, PUBG Ally’s live AI teammate beta, the latest Steam AI-disclosure argument, miHoYo’s AI companion, General Intuition’s gaming-data funding, coding-agent provenance research, and playable-game benchmark papers.

The strongest new game-specific item is not a model launch. It is a campaign signup flow: GamesRadar+ reported that Sega’s Sonic Chaos Emeralds hunt pointed users through Community terms that allow information to be used for AI-model training and improvement.

What Changed Overnight

  • GamesRadar+ reported that Sega’s Sonic campaign asked participants to accept Community terms before joining the Chaos Emeralds hunt, with AI-training language appearing in the signup path.
  • Community’s own terms, last updated May 21, 2026, say member messages and related information may be used to create, train, and refine machine-learning or analytical models.
  • TechRadar reported that PUBG Ally’s AI teammate mode is playable in beta on Steam through the end of June, using Nvidia ACE technology.
  • Epic CEO Tim Sweeney again criticized Steam’s AI disclosure labels, while prior Steam reporting keeps the line focused on AI-generated content players consume.
  • PC Gamer reported that miHoYo has released BSide: Olivia Lin, an AI companion on Steam in China, as the studio also works AI into future games and tools.
  • New and recent papers on coding agents and game generation keep pushing evaluation toward provenance, human review load, executable projects, and browser playtesting.

Sega’s Sonic campaign turns AI terms into the lead issue

GamesRadar+‘s report says Sega’s Sonic Chaos Emeralds hunt asked fans for signup information before online or city-based participation in Los Angeles, San Diego, Chicago, and New York. The campaign was operated through Community, a messaging platform, and the report points readers to Community terms tied to AI-model training and improvement.

Community’s terms are a useful primary source because they are broader than one campaign. They say Community may use member messages and related information, including demographic and location information on aggregated, de-identified, or anonymized bases, to improve technology and train or refine machine-learning and analytical models.

For game companies, this is a different disclosure problem from whether a Steam game used generated art. A fan campaign can collect phone numbers, messages, location-adjacent information, and participation data before the user ever reaches a game. The relevant question is not only “was AI used?” It is also what player or fan data is collected, who processes it, and whether AI training is clear before someone joins.

Steam’s AI label fight is still unresolved

GamesRadar+ and PC Gamer both covered Epic CEO Tim Sweeney’s latest criticism of Steam’s AI disclosure labels. Sweeney argues that the label can make it harder for a developer to reach players, especially when the AI use is part of production efficiency rather than the player-facing product.

The older Steam disclosure update is still the practical anchor: PC Gamer reported that Valve’s form focuses on AI-generated content consumed by players, not every behind-the-scenes efficiency tool. That distinction matters for AI-game creators because generated NPC dialogue, player-facing art, quests, and live behavior are different from using an agent to clean up build scripts or organize a backlog.

The Sega item shows why disclosure cannot be reduced to one store label. Storefronts, campaigns, community tools, chat companions, and creator platforms can all carry different AI and data terms.

AI Characters And Live Tests

PUBG Ally is now being tested with real players

TechRadar reported that PUBG Ally’s Ally Duo mode is playable on Steam through the end of June. The mode uses Nvidia ACE and gives players an AI teammate, Ella, that can respond to spoken or typed instructions while participating in the match.

The Verge’s earlier Nvidia ACE coverage described the pitch: AI characters that can perceive, plan, and act inside games rather than only deliver chatbot dialogue. The public beta is important because it moves that claim into a live shooter setting where timing, noise, navigation, loot, and combat decisions are testable.

TechRadar’s early impression was skeptical. The reported doubts were about whether the companion feels useful and human-like in practice, not whether the demo concept is interesting. That is the right bar for AI game characters: they need to help the player inside the loop, not only answer in a convincing voice.

miHoYo ships an AI companion on Steam in China

PC Gamer reported that miHoYo has released BSide: Olivia Lin on Steam in China. The software presents Lin Li as an AI companion who can respond through letters, music, and personal-story prompts.

The item belongs in the game newsletter because miHoYo is not a generic chatbot company. It operates live-service games with large content pipelines, and PC Gamer connects Olivia Lin to the studio’s broader AI investments and upcoming titles, including Petit Planet and Genesis.

That does not make Olivia Lin a game by itself. It does show a major game company testing companion behavior, music interaction, and character framing on a distribution surface familiar to players.

Creator Tools, Funding, And Studio Workflows

General Intuition keeps gaming data in the AI-lab funding story

Axios reported that General Intuition raised a $320 million Series A at a $2.3 billion post-money valuation. The company’s bet, according to Axios, is that gameplay video and player inputs can help train world models and large action models faster or more cheaply than other approaches.

This remains useful background because gaming data is being pitched not only as entertainment content, but as a training substrate for action, simulation, and world modeling.

Studio AI claims remain split

PC Gamer reported EA president of enterprise development Laura Miele’s claim that AI has helped drive a rise of creativity at EA studios. GamesRadar+ published a broader developer-sentiment piece collecting objections from developers who see generative AI as ethically, creatively, environmentally, or economically weak for many game-development tasks. GamesRadar+ also covered CD Projekt Red joint CEO Michal Nowakowski saying fully AI-generated games are coming, while expressing doubts about whether rapid prototype factories are the right path.

Those items should be read separately. EA is talking about internal production gains. Developers in the GamesRadar+ package are describing craft, labor, quality, and trust objections. CD Projekt is discussing the direction of whole-game generation. Together they show an industry disagreement, not a settled conclusion.

Coding Agents And Provenance

Agent work is becoming harder to see

Axios reported that Codex usage is accelerating, citing a report from OpenAI, Columbia, Duke, and the University of Pennsylvania. The sampled user thresholds are model-estimated, and Axios notes the user base is still small, but the pattern is important: delegated coding work is moving from demos into regular tasks.

“Detecting AI Coding Agents in Open Source” adds a supply-chain warning. The paper studies more than 180 million repositories and reports that bot-account lookup alone recovered only 3.3% of one Claude Code commit snapshot. That means visible bot identities are a weak proxy for how much agent-written code is entering open-source projects.

“Augmentation with Dilution” looks at what happens after agent adoption. Across 11,097 GitHub repositories from January 2023 through May 2026, the paper reports no significant change in the absolute number of human contributors, but lower human contributor density, fewer newcomers as a share of contributors, and a 5.3% increase in review depth.

For game teams, the connection is practical. Game projects contain gameplay scripts, scene files, shader code, UI, audio hooks, physics behavior, and tests. If agents add more code while making authorship less visible, teams need stronger review evidence and executable checks.

Playable-Game Benchmarks

The research package is moving in the same direction as the product news: AI-game systems need executable evidence, not only generated code or polished clips.

GameCraft-Bench evaluates whether agents can build complete Godot games from natural-language specifications. Its strongest reported agent score is 41.46%, and the paper says agents often implement recognizable mechanics while missing content, visual feedback, and coherent presentation.

GUI Agents for Continual Game Generation argues that a game-generation loop needs a player. Its PlaytestArena covers 200 browser-based game-generation tasks, and Play2Code uses a coding agent and GUI playtester in a shared loop.

GameGen-Verifier decomposes game specifications into runtime-checkable keypoints. GameDevBench evaluates 132 game-development tasks with multimodal assets and scenes. OpenGame proposes an open agentic coding framework for web games. AI GameStore uses human games as a broader evaluation space for vision-language models, and SWE-Bench Mobile shows that industrial mobile-app tasks remain difficult even outside games.

The shared direction is execution. A generated game has to run, respond, show the right thing, and keep its rules intact. Static code output is only one part of the evidence.

Watch Next

  • Whether Sega or Community changes the Sonic campaign flow, terms presentation, or AI-training language after the fan reaction.
  • Whether PUBG Ally’s beta produces public gameplay evidence beyond edited demos and early impressions.
  • Whether Steam, Epic, and other stores separate player-facing generated content, behind-the-scenes production tools, live AI behavior, and campaign data terms more clearly.
  • Whether miHoYo expands Olivia Lin outside China or connects companion behavior to its game worlds.
  • Whether coding-agent provenance papers lead to practical repository labels, review workflows, or storefront expectations for generated code.
  • Whether game-generation benchmarks converge on playable traces, runtime assertions, and browser playtests rather than screenshot-only judging.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.