Capcom Bans Gen AI Assets, Pearl Abyss Apologizes, and OpenAI Scraps Sora
Four studios took public positions on AI this week, from Capcom banning gen AI assets to Pearl Abyss apologizing after undisclosed AI art surfaced in Crimson Desert. The disclosure debate that was simmering at GDC broke into the open. On the audio side, IBM integrated ElevenLabs into watsonx for enterprise voice agents, and Mistral launched its first TTS model with open weights but a non-commercial license that blocks production use.
Industry
Capcom drew the sharpest line, stating it will not implement any generative AI assets into its games. A separate report confirmed the company will use generative AI to speed up production, but not for asset creation. The boundary between those two positions broke down at Pearl Abyss. Players spotted AI-generated artwork in Crimson Desert; the studio issued a public apology, attributed the asset to an external studio’s “temporary placeholder,” and committed to a full audit. A Techdirt opinion piece argued the placeholder excuse is losing credibility. A GamesIndustry.biz survey of over 600 industry professionals found 86% believe generative AI use should be disclosed on storefronts, with 51% supporting mandatory requirements.
Six outlets covered one Eurogamer interview in which Owlcat Games PR Manager Katharina Popp confirmed the studio uses generative AI for prototyping and placeholders on The Expanse: Osiris Reborn, with the final product to be “100 percent human made.” Kotaku flagged that claim as a “grey area”, citing the Crimson Desert precedent. Owlcat faced a similar cycle in 2024 when a concept artist job listing required Midjourney experience, and the studio’s position has been consistent since. Most of the current coverage framed the disclosure as a defensive response rather than a restatement, which is not an unfair reading.
Charlotte Cook, a consultant, told AI Gamechangers that AI’s most immediate value sits in pre-production, and that some companies are deploying it internally while publicly distancing themselves. NetEase offered a quiet counterpoint: its mobile game Project Parallel has used AI-generated 2D assets for over a year, and no one noticed. At GDC, Tencent demonstrated genAI animation tools that Game Developer called “technically impressive” but questioned whether they offer anything for developers.
Audio & Speech
Three new TTS models launched in a single week, all positioning against ElevenLabs. Mistral released Voxtral TTS, a 4B parameter open-weight model claiming a 68.4% win rate over ElevenLabs Flash v2.5 in multilingual voice cloning evaluations. It supports 9 languages, clones from under 5 seconds of reference audio, and reports 90ms time-to-first-audio according to TechCrunch. The self-hosting pitch is strong for studios with data residency requirements. The catch: Voxtral ships under CC BY-NC, which rules out commercial game production without a separate license. Smallest.ai followed with Lightning V3, claiming a 3.89 MOS that beats OpenAI and ElevenLabs on conversational benchmarks, with 15-language support and real-time chunked generation at 44.1 kHz. techcrunch.com
ElevenLabs, meanwhile, have secured an IBM partnership integrating its TTS and STT into watsonx Orchestrate with PCI compliance and zero-retention mode. That enterprise signal is not insignificant: 70 languages, 10,000+ voices, and the compliance certifications that procurement teams actually require. On the ASR side, Cohere shipped Transcribe under Apache 2.0, a 2B parameter model it claims outperforms ElevenLabs Scribe on the Hugging Face Open ASR Leaderboard. tech.yahoo.com
Video & Animation
OpenAI has scrapped Sora, with no plans to integrate video generation into ChatGPT. The Verge reports the company is also winding down a $1 billion Disney deal, citing compute costs that weren’t generating returns. TechCrunch called it a “reality check moment” for AI video, a verdict reinforced by ByteDance delaying its Seedance 2.0 launch over engineering and IP challenges. For studios evaluating Sora for pre-vis or marketing asset workflows, the vendor stability question just answered itself. theverge.com
SimilarWeb traffic data for February 2026 puts the surviving market at xAI’s Grok (~40% share), Runway (~14%), Google Flow (~10%), and Kling AI (~7%). Runway continues targeting filmmakers with Gen-4.5 and 4K exports. Luma AI expanded its stack with Uni-1, a multimodal reasoning model for image creation and precision editing, alongside its Ray 3 video tools. officechai.com
EA’s Battlefield 6 offered a concrete production case: Voice2Face generated all basic facial animation from audio, with 30% requiring no manual edits, while FaceRig integration with Maya reduced face sculpting from approximately two weeks to a few hours. Tencent showed genAI animation tools at GDC; Game Developer’s assessment was blunt, calling the demo “technically impressive” but questioning whether the tools “offer anything for developers.” gamedeveloper.com
Tool Updates
Cursor: Cursor launched Composer 2 and an X user quickly identified the model was built on Kimi 2.5, an open-source model from Chinese company Moonshot AI. VP Lee Robinson confirmed the open-source base but said roughly three-quarters of training compute was broadly Cursor’s own work. Co-founder Aman Sanger acknowledged the missing attribution as a “miss” and said they would correct it. Moonshot AI confirmed an authorized commercial partnership. (TechCrunch)
Gemini: Google added “Import Memory” and “Import Chat History” features to Gemini on desktop. Users can transfer personalized data from a competing AI by copying a Gemini-provided prompt into their current assistant and pasting the output back, or by uploading an exported chat history file. Both features target switching friction: anyone locked into another assistant’s learned preferences can now migrate that context to Gemini without starting over. (The Verge)