Generative AI in game development: from input rights to release evidence

Research cut-off: 22 July 2026
General information only. This research note is not legal advice and does not determine the position for any particular product, transaction or jurisdiction.

Abstract

Generative artificial intelligence has entered online games at three connected points: as a production tool, as part of the technology supply chain, and as a player-facing runtime system. It can generate concept art, textures, music, voice, code and narrative. It can also generate non-player character (NPC) dialogue, quests, speech, images and player-created content while a game is running. The legal question is therefore not a single question about whether an AI output has copyright. It is a chain of questions that must be kept separate: may the source material and player data be supplied to a model; what do the model and API contracts say about inputs, outputs, training, suspension and termination; does the result contain a human author’s original expression; can a protectable result still infringe an earlier work, character, portrait, voice or trade secret; how must a product label, review, distribute and preserve evidence of AI-generated content?

This note takes mainland China as its primary frame. It compares eight cases or regulatory samples, six distribution-channel policy groups and, on a limited basis, Article 50 of the European Union Artificial Intelligence Act. It makes three working conclusions. First, material that is publicly accessible, material held by a company and material covered by a complete training licence are three different states. The ability to see a file, a purchase receipt or possession of a file on a company drive does not by itself establish permission for training, fine-tuning, retrieval-augmented generation (RAG) or re-publication. Second, Chinese cases do not support a categorical rule that AI images are either protected or unprotected. The difference between outcomes turns largely on the generation process, the human contribution and the original evidence. Third, runtime generation moves clearance from a one-time pre-launch exercise to an operating control system. Model versions, safety layers, player prompts, visible and embedded labels, reports and rollback records jointly shape risk.

For a game company, the most practical governance objects are three auditable tables: an input-source and licence table, a model/API obligations table, and a generated-asset and release-evidence table. Connecting those tables to the asset library, version control, build pipeline and store-submission process is what turns an opaque production shortcut into an asset that can be released, financed and defended.

Keywords: generative AI; online games; copyright; voice rights; player UGC; AI-content labelling; platform rules; evidence chain

1. One AI feature can contain six different legal questions

A model-generated character illustration can raise at least six issues. Was the reference image permitted as an input? Does the model licence allow commercial use? Will the provider retain the input or use it to improve a service? Does the image meet the threshold for a work and, if so, which human contribution matters? Does it reproduce an earlier character or other protected expression? Must the game disclose or mark the content at release? An affirmative answer to one question does not answer the next.

This paper separates the issues as follows:

1. Input legality: whether material, data and confidential information may be used for training, fine-tuning, RAG or prompting.
2. The model supply chain: what the foundation-model, API, plug-in and platform contracts permit and exclude.
3. Protectability of the output: whether the result is an original intellectual creation attributable to a human author or to an organisation under the applicable rules.
4. Output infringement: whether the result uses or reproduces an earlier expression, character, voice, portrait or secret.
5. Distribution governance: labelling, safety controls, UGC mechanisms and store-submission requirements.
6. Evidence and allocation of responsibility: who can prove source, process, review and response, and who controls, benefits from or can stop the risk.

The mainland Chinese Interim Measures for the Administration of Generative Artificial Intelligence Services require providers serving the public to respect intellectual-property rights, process personal information lawfully and use data with lawful sources. [1] An employee using a public tool internally, a game operator providing generation directly to players, and a publisher uploading an AI-made asset to a store are different actors and scenarios. The same obligation cannot be mechanically applied to all three.

This note does not attempt to restate the general protection of games, gameplay, characters or audiovisual works. Nor does it provide a full treatment of game analytics, player profiling or cross-border transfer. An AI character enters the separate anthropomorphic-interaction analysis only when it combines a stable persona, continuing interaction and emotional care, companionship or support; a task NPC, combat assistant or one-off dialogue does not become such a service merely because a large model is used.

The method combines statutory interpretation, case comparison, public contract analysis, content-lifecycle analysis and limited comparative law. The public platform policies cited here are snapshots of public material. PlayStation and Nintendo require partner access for some certification materials, so their private submission requirements cannot be inferred from this paper.

2. Classify the use before classifying the risk

The position of a model in the game determines the intensity and timing of the controls. Internal tools and player-facing systems should not be put in one bucket merely because they use the same model.

These uses sit on a spectrum from company-controlled material that can be fixed before launch to player-triggered output that changes while the game runs. The closer a feature is to runtime generation or player UGC, the less a one-time clearance exercise can do. Review, labels and evidence must operate continuously.

3. Source material: public does not mean licensed

“Input” is not one legal act. Foundation training can involve large-scale copying and technical processing. Fine-tuning or LoRA often uses a smaller and more targeted dataset, increasing the probability of reproducing a particular character or style. RAG may leave model parameters unchanged while still copying and indexing protected material and returning it to users. A prompt may be processed once, retained by a provider, seen by a human reviewer or fed into later service improvement. A project record should therefore state separately whether a copy is made, whether parameters change, whether the result is made public, whether the source text is retained and whether it contains personal information or secrets.

The same Chinese source rules can intersect with several interests:

• copyright in text, images, music, code, character settings and database arrangement;
• performers’ rights, sound-recording rights and the limits in the production contract;
• name, portrait, voice, privacy and personal-information interests;
• trade secrets in source code, unpublished narrative, numerical design, model weights and operating data;
• competitive interests arising from a data collection or platform investment; and
• website terms, asset-library licences, API policies and confidentiality agreements.

Chinese copyright law does not create a general commercial text-and-data-mining licence for generative AI. A fair-use analysis must fit the statutory circumstances and must not unreasonably prejudice the normal exploitation of the work or the legitimate interests of the right holder. “Technical innovation” is not a substitute for that analysis. [2]

3.1 Four common sources

Public web data. Public accessibility shows that there is no access barrier; it does not show consent to train. A web image may be protected, a public player post may contain personal information, and the website’s terms may restrict scraping or automated use. Robots.txt, a login wall and technical measures may affect authorisation, fault and unfair-competition analysis, but compliance with robots.txt is not itself a licence. If a vendor sells a “public web dataset”, the game company should ask for collection sources, exclusions, rights-request channels, personal-information handling, deduplication and deletion capability. “Collected from the internet” is not a rights warranty.

Licensed libraries. A normal commercial asset licence often permits use in the final product but not model training, a derivative training library, a style capability or upload to a third-party API. The contract should address machine learning, training, fine-tuning, synthesis, generated derivatives, sublicensing and model sharing. “Commercial use” and “perpetual buyout” do not automatically answer those questions.

Company-held assets. A file on a company drive does not prove that the company owns every right in it. A founder’s pre-incorporation demo, an outsourced illustration, an employee’s personal music library, a co-branded IP or a voice licence limited to one title or territory may not have been completely transferred. Training amplifies the gap: permission to use a character in one game does not necessarily include permission to train a general model or serve other projects.

User content. A standard user term saying that content may be used “as necessary to provide the service” cannot automatically expand into permission to improve a general model or conduct commercial training. Player material can include a portrait, voice, chat history, minor’s data or a third-party work. Sensitive personal information used for training requires the stricter controls of Chinese personal-information law; anthropomorphic-interaction rules separately require specific consent for sensitive interaction data used in training. [3]

3.2 Build a purpose-rights matrix

The same asset can have different answers for different purposes. A workable ledger should include at least these rows:

4. Model and API contracts: “the output belongs to you” is not a clearance opinion

Public commercial terms commonly say that the customer retains input and receives whatever rights in an output the supplier can transfer. Public OpenAI and Anthropic commercial terms are examples of that structure. [4][5] It has at least three limits.

First, a contract cannot create a copyright that the law does not recognise. If a purely machine-generated output lacks the required human contribution, “output ownership” does not turn it into an exclusive copyright. Second, a supplier-customer contract cannot eliminate a third party’s rights in a similar character, voice or work. Third, non-unique-output language means another customer may receive a similar result; a transfer clause is not proof of independent creation or market exclusivity.

The contract review should cover six variables:

1. Input use. Is customer content used for service improvement? Are there exceptions for abuse monitoring, human review, backups, feedback or beta features?
2. Input warranties. The customer normally warrants that it has the right to submit the input. An employee uploading an external artist’s work or unreleased code may create both a rights-chain problem and a customer breach.
3. Output review. The customer is usually responsible for suitability, accuracy and lawful use. Game review must also ask about character similarity, open-source fragments, voice recognisability and content safety.
4. IP indemnity. Supplier indemnities often exclude customer input, customer applications, combinations, modifications, known infringement, certain patent uses and commercial trademark use. A game’s combination of the output with an engine, character and brand may fall exactly within those exclusions. [4][5]
5. Suspension and change. Policy violations, legal requirements, security events, third-party dependencies and territory limits can cause a service to be suspended. Model and safety-layer updates can change a runtime NPC after launch.
6. Termination and data. Deletion deadlines, legal retention, abuse-content exceptions, export formats and log access determine whether the company can reconstruct a generation after a dispute.

For a core production or runtime feature, procurement should fix permitted games and territories, input classes, the default training position, human access, data location and subprocessors, model-version notices, service levels, safety responsibilities, IP cooperation, logs, transition support and alternative-model migration. If a runtime AI controls core play, the contract must answer who retests after an update, who freezes an old version when a policy changes, who preserves prompts after a complaint and how player memories and paid rights migrate after termination. “API replaceable” is a technical hypothesis until data formats, prompts, safety rules and regression tests prove it.

5. Output protection, authorship and ownership

5.1 Human expression, not polish, is the common threshold

Under the Copyright Law of the People’s Republic of China, a work is an original intellectual achievement in literature, art or science that can be expressed in a certain form. [2] The model is not a natural-person author. The relevant question is whether the final result can be attributed to a human’s original intellectual activity or, under the applicable employment or organisational rules, to an organisation.

A polished result, a high production cost or permission for commercial use does not satisfy the work requirement. Conversely, extensive AI assistance does not automatically defeat protection. A person may make identifiable expressive choices in composition, character design, source selection, rhythm, camera, local repainting, code structure or overall arrangement. The output should be split into protectable and unprotectable elements rather than assigned a binary “AI” label.

As a limited foreign comparison, the US Copyright Office asks for identifiable human creative contribution, disclosure and exclusion of AI-generated material that is not protected. Zarya of the Dawn separated protected text, selection and arrangement from Midjourney images that were not registered as the author’s own expression. Thaler v. Perlmutter addressed the narrower question of an application naming AI as the sole author. Neither decision supports the proposition that every AI-assisted work is unprotected. [23][24][25][26]

5.2 Evidence differs by medium

Prompts can be ideas, instructions or expression. Their evidentiary value usually lies in showing how the user controlled the final expression, not in their length. Parameters prove the operating environment but may not themselves be original. Repeated generation proves a process but not purposeful choices. Selection can be an arrangement contribution, but selecting the “prettier” of two random results is weaker evidence than iterative selection against a defined expressive goal. Local repainting, layer editing, cutting, composition and code changes are easier to separate from the model’s contribution. A complete compilation or arrangement may be protected even if an individual AI element is not.

If a human artist or designer contributes enough expression, company ownership still returns to ordinary employee-work, commissioned-work, organisational-work and assignment rules. Outsourcing contracts should address AI permission, input warranties, delivery of process records, transfer of protectable rights, use of unprotectable material, third-party claims and model-file ownership. Authorship, AI disclosure and platform disclosure answer different questions and can coexist.

6. Process evidence and contrasting cases

The Beijing Internet Court’s 2023 image case (2023) Jing 0491 Min Chu No. 11279 involved Stable Diffusion, positive and negative prompts, parameter settings, repeated adjustment and candidate selection. The court found, on the evidence of that process, that the image reflected the claimant’s individual expression and constituted an artistic work. [6] The limiting phrase is “on the specific generation process”, not “all Stable Diffusion images are protected”.

The publicly reported 2025 “crystal diamond cat pendant” case reached the opposite result. The claimant did not provide the original generation record. During the lawsuit, a describe function was used to describe the finished image and simulate a reconstruction. The court found that the original prompt, environment and selection could not be restored and that the reconstruction differed. The evidence did not prove sufficient human creative labour, and both the first and second instances dismissed the claim. [7]

The difference can be reduced to one chain:

Original input → environment and parameters → candidates and iterations → human edits → final output.

The first case could be traced along the chain; the second began with a finished file and a later simulation. The issue was not two incompatible slogans about AI. It was whether the claimant could prove which person made which choices, when and how those choices formed expression.

Traditional asset management keeps the final PNG, WAV, script or source file. With AI, that file only proves that a result exists. A company should preserve tool and model name, exact version, account and workspace, timestamp, system and user prompts, negative prompt, parameters, seed, plug-ins and LoRA; source IDs and hashes; representative candidates and selection reasons; PSD or engineering files, local repainting, multitrack edits, code diffs and human commits; review, approval, build number and first release. That evidence supports a copyright claim, a response to a copying allegation, proof of input permission and proof that the platform took reasonable care.

7. Infringement, personality rights, secrets and responsibility

An image can be a protectable new work and still reproduce another character. Code can contain human choices and still include a governed open-source fragment. A synthetic voice can be a recording and still violate the cloned person’s voice interests. The order of analysis should be: identify earlier elements; assess access, substantial similarity, exceptions and permission; then allocate conduct and fault. Lawful training does not automatically make every output lawful, and disputed training does not make every output substantially similar to a particular work.

Style is generally more abstract than a specific character design, scene, melody, code or arrangement. Saying that a result is “in the style of” a studio does not itself prove copyright infringement. A stable reproduction of a character’s clothes, face, marks, composition or a sequence of plot events raises a much stronger risk, potentially alongside trade mark, merchandising and unfair-competition claims.

The Hangzhou Internet Court’s “Ultraman LoRA/model” dispute shows that the model file and sharing function can be the infringement node, not only the final image. The user uploaded Ultraman images, trained a model, and made the model, cover and similar outputs available for browsing, download, publication or sharing. The court held the platform liable for helping infringe the right of communication through information networks and maintained the result on appeal. [8] A game AI workshop should therefore govern downloadable models and templates, with notice, takedown or link-disconnection measures coordinated with the network-information-right framework. [27]

7.1 Voice and likeness

The Civil Code provides that natural-person voice is protected by reference to portrait-right rules. [9] In a Beijing Internet Court voice case (2023) Jing 0491 Min Chu No. 12142, a performer had recorded an audiobook, but the data authorisation was not signed by the performer and the existing commission, name and portrait terms did not cover AI processing and synthetic voice. A court-generated test found the output recognisable by timbre, intonation and pronunciation style. [10]

For a game, a recording licence is not automatically a licence to clone the voice. The authorisation should specify samples, model training, synthesis, character, language, term, territory, sublicensing, API calls, withdrawal and model deletion. Identifiability is the centre of the question: the listener need not know the performer’s name if relevant listeners can associate the timbre, intonation and pronunciation with a particular person. A real name, face or life story adds name, portrait, privacy, endorsement and false-association risks. Player systems should offer complaints for real people’s names, voices and portraits, quick freezing and repeat-infringer controls.

7.2 Secrets and model parameters

Uploading unreleased code, numerical tables, anti-cheat logic, launch plans or player complaints to a public model risks more than provider training. It can breach confidentiality and weaken the company’s proof that reasonable secrecy measures were maintained. The Supreme People’s Court’s 2025 unfair-competition case on a “cartoon transformation” effect treated model structure and parameters, developed through data training and tuning and capable of generating a commercial advantage, as a protected competitive interest. [11] A company must protect both external inputs and its own weights, system prompts, evaluation sets, safety rules, experiment records and deployment parameters through classification, access control and audit logs.

Responsibility should be analysed through control, knowledge, benefit and ability to respond:

The “neutral platform” label cannot hide training, model sharing, recommendation, trading or technical controls. But the existence of filtering does not make a platform strictly liable for every unforeseeable result. A functional reasonable-care standard, followed by notice-specific action, is a more useful operating model.

8. Labelling and runtime generation

China’s AI-generated-content regime is a metadata chain, not one corner badge. The 2025 Measures for Labelling AI-Generated and Synthetic Content distinguish explicit labels that users can see in text, sound or graphics from implicit labels written into file data. [12] The generation service, propagation service, user and app store each have a role. A generator may need to add labels and metadata; a propagation service may verify metadata and add a generated or suspected-generated notice; the user must declare generated content and must not maliciously remove or falsify labels; an app store may ask whether an app provides generation and verify the material. In April 2026, the Cyberspace Administration of China reported action against products including Jianying, Maoxiang and Jimeng for failure to implement labelling, showing that the requirement has moved into enforcement. [13]

The product must be assessed by function:

Human editing may matter to copyright, but it does not automatically remove a labelling obligation. A rule about generated content and service context has a different purpose from a rule about human authorship.

Runtime NPCs move review into every session. Pre-launch tests should include celebrity, artist, real-person, game-competitor and famous-character prompts; sexual, hateful, self-harm, fraud, cheating, exploit and minor-safety prompts; multilingual, obfuscated, encoded, image-prompt and multi-turn bypasses; RAG prompt injection, document overreach, secret leakage and cross-player data mixing; model and knowledge-base updates; provider outage, filtering timeout and human-review backlogs. A stable persona with cross-session memory and emotional interaction may enter the separate anthropomorphic-interaction regime; ordinary task NPCs should not be classified that way without the facts. [14]

9. Player AI UGC and distribution channels

An in-game UGC agreement should answer five questions:

1. Does the player have the right to upload the image, music, voice, code or real-person material, and what personal information or secrets are prohibited?
2. What is the platform promising about output ownership? It should not promise that pure AI output must be copyrighted. Player expression, platform templates and pre-existing game assets should be separated.
3. What licence does the platform need for operation, display, promotion and technical processing? Commercial release or sale should require a further permission and revenue arrangement rather than an unbounded transfer.
4. How are the dataset, model, finished UGC and account reviewed? The mechanism should include reports, counter-notices, appeals, repeat infringer controls and emergency freezing.
5. Is player content used to train a model, and how is sensitive personal information excluded or separately consented to?

Apple’s UGC rules require filtering, in-app reporting and timely response, blocking abusive users and public contact information. Google Play requires ongoing review of UGC and AI-generated content, in-app reporting or flagging and use of reports to improve filtering. [15][16] These channel rules do not replace copyright or privacy law, but they can determine whether a game launches and remains available.

Steam distinguishes pre-generated AI content from live-generated content. Developers disclose the former when AI was used for art, code or sound during development; for the latter, they must disclose runtime generation and explain safeguards against illegal content. Steam may display part of the disclosure and provides a reporting path. [17] Apple’s public review guidelines do not use the same universal production-stage field, but they impose UGC, dynamic-software and third-party-AI data obligations. Google’s public rules focus on content safety and in-app reporting. Microsoft Store policy 11.16 requires disclosure of live generative AI in store metadata and in Partner Center. [18]

Public PlayStation and Nintendo material is incomplete without partner access. That absence is itself a due-diligence conclusion: the release team should export the current questionnaire, content policy, age-rating and online-safety requirements from the actual account and preserve the answers.

10. China and the EU: a limited comparison

Article 50 of the EU AI Act distributes transparency duties between providers and deployers. An AI system that directly interacts with a person should make the AI nature known; providers of systems generating synthetic text, audio, images or video should make outputs machine-readable; deployers have additional duties for emotion recognition, biometric categorisation, deepfakes and certain public-interest text. [19] The European Commission’s July 2026 guidance states that Article 50 applies from 2 August 2026 and gives a limited extension for certain marking obligations of systems already on the market. [20][21]

China and the EU can share a technical base—generation ID, model version, content hash, metadata and a visible interaction notice—but the scope, exceptions, responsible party and transition period differ. China’s framework links generation, propagation, user declaration and app-store checks. The EU uses provider/deployer roles and horizontal transparency. A game company deploying a third-party model and offering a branded AI system can occupy both roles.

11. A seven-gate operating model

The following gates connect legal analysis to production:

1. Source admission: record rights holder, source, use, term, territory and restrictions for each dataset, asset, knowledge base and voice corpus.
2. Tool admission: assess model/API, privacy, security, business continuity and data-processing terms; keep the company account and DPA under control.
3. Generation trace: archive model and version, input, parameters, reference assets, candidates and timestamp.
4. Human creation: preserve repainting, editing, arranging, code changes and story revisions in professional tools and version control.
5. Asset approval: review similarity, open-source licences, real-person elements, secrets, content safety and labels together.
6. Release submission: keep metadata through the build; generate jurisdiction and channel disclosures; preserve receipts and answers.
7. Runtime response: make output auditable, reportable, freezeable and reversible; model changes trigger regression testing.

A minimum internal record can include:

asset_id / project / creator / employment_or_vendor_basis
source_asset_ids / source_licence / allowed_ai_uses
provider / model / version / workspace / generated_at
system_prompt / prompt / negative_prompt / parameters / seed / plugins
candidate_ids / selection_reason / human_edit_files / commit_hash
similarity_review / voice_portrait_review / licence_review / approver
explicit_label / metadata_label / build_id / store_submission
complaint_id / action / rollback_version / retention_and_deletion

Not every field should be public. The purpose is internal traceability, with access and retention limited by necessity, disputes, regulation and personal-information controls. Contract and policy versions should be archived at important releases: a current web link cannot prove what a model or store permitted six months earlier.

When a complaint arrives, locate four objects at once: the content instance, the model/LoRA/template, the uploading or publishing account, and the recommendation, trading or export paths that can reproduce or spread it. The response may be a single-item takedown, model suspension, keyword restriction, trade freeze, evidence preservation, supplier notice or version rollback. A counter-notice or appeal matters too; moderation should not become a tool for competitive suppression.

12. Conclusions, limits and pre-publication checks

Generative AI has not removed a game’s rights chain. It has extended it from “who delivered the file?” to “who supplied the data, controlled the model, made the expressive choices, reviewed the result and released it?” The business should be able to answer why an input was permitted for that use, what the supplier terms allowed at the time, which human choices formed the expression, what review covered earlier rights and personality interests, what Chinese, EU and store labels applied, and how the product can locate and roll back a disputed output.

For a small team, the minimum viable system is not a long policy. It is the three linked tables described at the start: input sources and licences, model/API obligations, and generated assets and release evidence. Once connected to the asset and build pipeline, legal review can rely on records rather than memory.

This study has limits. Chinese copyright treatment of AI outputs is developing through individual cases; the “crystal diamond cat pendant” public report does not disclose a case number and should be checked against the judgments before formal litigation citation. EU Article 50 guidance and enforcement were still new at the source cut-off. Private PlayStation and Nintendo certification materials were unavailable. Enterprise orders can differ from public model terms, especially on indemnity, data retention, versions and termination. This note also flags, rather than fully analyses, cross-border data, minor data and anthropomorphic emotional interaction.

Before release, the team should re-run the channel and jurisdiction matrix for the selected countries; map the real data flow of NPC, AI UGC, export and sharing; collect model orders, DPAs and console materials; sample ten AI assets from final file back to source, generation, human editing and release; and separately review real-person voice, portrait, co-branded IP, outsourced datasets and downloadable LoRA. EU deployments around 2 August 2026 should receive an additional Article 50 review.

References

[1] Cyberspace Administration of China et al., Interim Measures for the Administration of Generative Artificial Intelligence Services, 2023-07-13, source

[2] Copyright Law of the People’s Republic of China (2020 revision), source

[3] Personal Information Protection Law of the People’s Republic of China, source; Interim Measures for the Administration of Anthropomorphic Interaction Services, source

[4] OpenAI, OpenAI Services Agreement, source

[5] Anthropic, Commercial Terms of Service, source

[6] Beijing Internet Court, (2023) Jing 0491 Min Chu No. 11279 and related case materials, source

[7] Beijing Court Network, “AI-generated image copyright claim rejected where the generation record did not prove creative input,” 2025-09-17, source

[8] China Intellectual Property News, “Hangzhou Internet Court issues judicial recommendations on AI platforms,” 2025-04-02, source

[9] Civil Code of the People’s Republic of China, Articles 1018, 1019 and 1023, source

[10] Beijing Internet Court, (2023) Jing 0491 Min Chu No. 12142 and related case materials, source

[11] Supreme People’s Court, 2025 Typical Anti-Unfair-Competition Cases, “Cartoon transformation effect,” source

[12] Cyberspace Administration of China et al., Measures for Labelling AI-Generated and Synthetic Content, source

[13] Cyberspace Administration of China, “Authorities investigate products that failed to implement AI-generated-content labelling,” 2026-04-28, source

[14] Cyberspace Administration of China et al., Interim Measures for the Administration of Anthropomorphic Interaction Services, source

[15] Apple, App Review Guidelines, source

[16] Google Play, AI-Generated Content and User Generated Content, source; source

[17] Steamworks Development, “AI Content on Steam,” 2024-01-10, source

[18] Microsoft, Microsoft Store Policies, v7.19, §11.16, source

[19] Regulation (EU) 2024/1689, Article 50, source

[20] European Commission, Guidelines on Transparency of AI-Generated Content, source

[21] European Commission, Transparency obligations under Article 50 of the AI Act — Questions and Answers, source

[22] Regulation (EU) 2024/1689, Article 50 implementation material, source

[23] US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, 2025, source

[24] US Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 2023, source

[25] US Copyright Office, Zarya of the Dawn registration reconsideration letter, 2023-02-21, source

[26] Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), source

[27] Regulations on the Protection of the Right of Communication through Information Networks, source

[28] Cyberspace Administration of China et al., Provisions on the Administration of Deep Synthesis in Internet-based Information Services, source

[29] OpenAI, Data Processing Addendum, source; Anthropic, Data Processing Addendum, source

[30] National Information Security Standardization Technical Committee, Practice Guide on the Security of Agent Deployment and Use, 2026-07-01, source