Unity: an assistant that reads the project
Unity's public beta of AI tools built into the Editor is open to developers on Unity 6.0 or later. It includes the in-editor AI Assistant with Ask, Plan and Agent modes, the Unity MCP Server, AI Gateway, and asset Generators. Unity's own pitch is that this is not a general-purpose chatbot bolted onto the side of the editor, but an agent that understands the project's scene graph, GameObjects, components, packages and target platform.
Unity's documentation describes how the three modes differ in autonomy:
- Ask: answers questions and explains console errors using the project's context.
- Plan: breaks a multi-step task into a reviewable plan before anything changes.
- Agent: can write C# scripts, modify scene components and create prefabs. Unity says every change is reversible and that permissions are configurable. That range runs from read-only queries to full autonomy.
MCP Server and AI Gateway
The MCP Server matters most to developers who live in an IDE. Unity says that without MCP an external assistant sees only code files. With it, the assistant can inspect scene state, read GameObjects and components, check console logs, and trigger Editor actions. Unity's examples of compatible clients include Claude in VS Code and Cursor. Third-party coverage also lists VS Code Copilot among supported clients, "depending" on configuration.
AI Gateway lets studios route model subscriptions they already pay for through the Editor. Unity says this does not consume Unity AI credits.
Plans, pricing and caveats
These are vendor-stated details, and they are beta details:
- Unity's feature page says Pro, Enterprise and Industry customers get the in-editor assistant as part of their subscription. Personal users can take a free trial, which converts to a $10/month subscription.
- Unity's guide describes the Personal trial as 1,000 credits free for 14 days.
- The beta needs a Unity Cloud-connected project and a package installation. It is not on by default.
- Unity warns that features and availability may change or be discontinued without notice.
- Unity says the tools are not Unity Muse, the earlier product built on Unity's own models. The beta tools use third-party models.
So "Unity's current suite" is accurate only with the word beta attached. Nothing here establishes that the assistant's output is correct. Unity's claims about faster iteration are workflow benefits, not measured productivity results.
Roblox: generation inside Studio
Roblox's Creator Hub documentation says Assistant answers questions, generates content and performs actions inside Studio. It pairs a large language model with a run-command system that acts on the place's data model.
For scripting, it can create scripts and modify existing ones, including across several objects. The docs give an example of a remote event that produces a local script, a server script and the event together. Roblox is blunt that generated scripts might not function flawlessly. Its answer is to edit them yourself or ask Assistant to revise them. That is the review step in practice.
Meshes and procedural models
/generate_meshtakes either a text prompt or an image, but not both in one request. Developers can set a bounding Part and a maximum triangle count, which defaults to 10,000./generate_procedural_modelbuilds parametric models. Roblox says they behave like standard objects until their parameters change, and are kept in a dedicated GeneratedFolder.- Roblox documents a limit of 50 procedural models per 24-hour rolling window.
- Segmentation splits a model into up to eight parts, so creators can change materials, attach scripts or swap components without regenerating the whole thing.
- Mesh segmentation for imported assets is labeled early preview.
Roblox's prompting guidance is realistic about the process. It says AI tools are non-deterministic, so the same prompt can produce different results. It tells creators to name editable properties explicitly, and that generation can take longer than a text reply. It also states that AI in its current state requires work. Treat generation as a drafting step, not delivery of a shippable asset.
What the adoption numbers do and don't show
Unity's 2025 Gaming Report says 96% of the studios it spoke to are integrating AI tools into their workflows. It adds that less than half of developers use AI tools in any single workflow. The report says machine learning has not delivered the full-scale revolution many expected. Its sentiment finding is that 79% of respondents felt positive about AI in game development.
The methodology matters. The report draws on a Cint telephone survey of 300 developers, with a margin of error of ±5.66% at 95% confidence. It also draws on in-person surveys of 144 studios at GDC 2024 and 231 at gamescom 2024, plus Unity Engine data. That is a vendor-run snapshot from 2024 data, not a census. It does not show that 96% of all studios worldwide use AI. It also should not be set against more recent, differently designed surveys without noting the difference. A third-party summary of a later GDC survey, which I have not verified, reported a much lower share of developers using generative AI. Different questions and samples give different answers.
The consistent finding is selective adoption: AI used for specific, often mundane problems rather than everywhere.
Rights, provenance and disclosure
Unity says assets from its in-editor assistant carry embedded metadata flagging them as AI-generated. It also says developers are fully responsible for app-store declarations and for verifying usage rights. It lets admins disable Generators in the dashboard while keeping the assistant active.
On the legal side, the U.S. Copyright Office's AI report is being issued in parts. Part 2, on copyrightability of generative-AI outputs, came out January 29, 2025. A pre-publication Part 3 on training was released May 9, 2025, with a final version to follow. The Office's analysis centers on human authorship. Using AI as an assistive tool does not by itself bar protection for a larger work. Wholly AI-generated output is a different matter. "AI content can never be copyrighted" is therefore too blunt, and so is "metadata settles ownership."
The practical takeaway, which is general industry knowledge rather than legal advice, is to keep provenance and rights review in the pipeline:
- Track which assets were AI-generated.
- Check the tool's terms and your publishing platform's disclosure rules.
- Confirm the legal position in each market you ship to. The Copyright Office's work covers U.S. law only.
NVIDIA ACE: AI in the shipped game
The other side of the story is characters players meet directly. NVIDIA describes ACE as a suite for building conversational, actionable in-game characters, with cloud and on-device models for speech, intelligence and animation. Its NVIGI plugins schedule inference alongside graphics workloads.
For Unreal Engine 5 developers, NVIDIA lists plugins for facial animation (Audio2Face-3D), speech recognition, a local LLM with function calling, and text-to-speech. Supported engine versions differ by plugin. The facial animation plugin is listed for UE 5.4 to 5.6, while the ASR, LLM and TTS plugins are listed for 5.5 to 5.7. Check the specific plugin before planning around it. NVIDIA also offers a beta Game Agent SDK in C/C++ for custom engines.
The design questions are the ones the games.gg piece raised: what the character may say or do, whether it fits the game's tone, and what it costs in performance. Local and cloud deployment each carry trade-offs. NVIDIA's page does not quantify the performance cost, and nothing in the evidence shows AI characters replacing scripted dialogue.
The takeaway
The evidence supports a measured conclusion. Generative AI is becoming a set of workflow-specific tools: context-aware coding and scene help, placeholder and parametric asset generation, and experimental runtime characters. The vendors' own documentation repeats the same caveat. Output must be integrated, tested, judged creatively and checked for rights. Faster first drafts are plausible, but better games, cheaper production and smaller teams shipping more are not shown here.
For Windows developers, the immediate points are practical. Unity's tools need Unity 6.0+, a cloud-connected project and a package install. External agents can connect through MCP, and Roblox's tools live inside Studio. Whether it helps depends on whether your team reviews the output as carefully as it would a junior colleague's pull request.