---
id: APP-MEDIA-001
title: "AI in Media and Creative Industries"
alternative_names:
  - "Generative AI in creative work"
  - "AI art and media production"
  - "Creative AI"
category: "Applications → Media and Creative Industries"
volume: VII
chapter: 42
entry_type: "Standard Article"
level: 3
status: "Draft"
last_reviewed: 2026-10-15
version: "0.1.0"
prerequisites:
  - "CAP-GENAI-001"
related_entries:
  - "TECH-DIFF-001"
  - "TECH-LLM-001"
  - "GOV-LAW-001"
  - "GOV-EU-001"
  - "ECON-LABOR-001"
  - "APP-EDU-001"
---

# AI in Media and Creative Industries

## 1. Concise Definition
**AI in media and creative industries** refers to the use of generative models for images, video, audio, music and text in publishing, advertising, film, games and design, and the legal, labour and cultural questions this raises. The main disputed issues are who owns and who may be compensated for outputs, whether training on copyrighted works is lawful, and how generative tools change the quality and diversity of creative work.

## 2. Overview & Significance
Generative models produce media that is cheap to make and difficult to distinguish from human-made content, which affects production costs, employment terms, the supply of training material and the information environment. This article does not duplicate the technical treatment of generative models (see CAP-GENAI-001 and TECH-DIFF-001). It separates three evidence classes:

| Class | Meaning | Examples below |
| :--- | :--- | :--- |
| **Deployed product** | Tools and policies in operation | Commercial generative tools; union contract terms; the EU transparency rule |
| **Validated result** | Experimental, legal-authority or regulator findings | A creativity experiment; the US Copyright Office report on copyrightability; appellate holding on human authorship |
| **Research demonstration** | Technical studies of risk or capability | Memorisation of training images by diffusion models |

Much material about this sector is advocacy from rights holders or developers. Legal questions in particular are unsettled, and facts below are dated.

## 3. Historical Development
Computational art and procedurally generated media predate deep learning. Modern generative image systems were enabled by generative adversarial networks and then diffusion models (TECH-DIFF-001), and by large language models for text (TECH-LLM-001). Public text-to-image and chat systems released from 2022 prompted a wave of litigation and labour negotiation. A 2023 Science policy forum, "Art and the science of generative AI", set out the research agenda on authorship, creativity and copyright (Epstein et al., 2023). The Writers Guild of America (WGA) reached a collective agreement in 2023 that addressed AI directly (WGA, 2023). In 2025 the US Copyright Office published its report on the copyrightability of AI-generated outputs (U.S. Copyright Office, 2025), the D.C. Circuit affirmed that copyright requires human authorship in Thaler v. Perlmutter (IPWatchdog, 2025), and two federal district courts in California ruled on fair use in training on books (Jones Walker LLP, 2025).

## 4. Technical / Conceptual Explanation
**Why memorisation matters legally.** A generative model learns a distribution $p_\theta(x)$ over training media. If it assigns very high probability to individual training examples it can reproduce them. Carlini et al. (2023) used a generate-and-filter pipeline to extract over a thousand training examples from state-of-the-art diffusion models, from photographs of individual people to trademarked logos, and concluded that diffusion models are less private than earlier generative models such as GANs. This is a research demonstration of a capability, not a measure of how often ordinary users obtain copies. It matters because litigation turns partly on whether outputs substitute for training works.

**Why output similarity matters for culture.** A tool that suggests ideas to many writers may push them to similar ideas. An online experiment found that access to generative AI ideas caused stories to be rated as more creative, better written and more enjoyable, especially among less creative writers, while AI-enabled stories were more similar to each other than stories written by humans alone (Doshi & Hauser, 2024). The authors describe this as a social dilemma between individual gain and collective novelty.

**Provenance and labelling.** Technical marking of synthetic content and legal disclosure duties are policy responses; the EU AI Act contains transparency obligations in Article 50 (European Union, 2024).

## 5. Types, Components & Classifications
- **Visual:** text-to-image, image editing, concept art, advertising imagery.
- **Moving image and audio:** video generation, voice cloning, music generation, dubbing.
- **Text:** drafting, translation, localisation, script and article assistance.
- **Interactive media:** non-player-character dialogue and asset generation in games.
- **Policy layers:** copyright (outputs and training), contract and collective bargaining, personality and voice rights, content labelling.

## 6. Applications & Concrete Examples

### 6.1 Deployed products and instruments in force
- **Commercial generative tools** are available for images, audio, video and text. Studio-by-studio adoption claims and revenue figures are largely company-reported and were not verified for this draft.
- **WGA 2023 Minimum Basic Agreement.** According to the guild's summary, AI cannot write or rewrite literary material, and AI-generated material is not considered source material under the agreement. A writer may use AI with the company's consent and under company policies, but the company cannot require it, and the company must disclose AI-generated material given to a writer. The guild reserved the right to assert that training AI on writers' material is prohibited by the agreement or other law (WGA, 2023). The summary gives a term running to 2026-05-01; subsequent agreements were not verified for this draft [VERIFICATION NEEDED: terms in force after 2026-05-01].
- **EU transparency rule.** Article 50 of the AI Act sets transparency duties, including machine-readable marking of AI-generated outputs. Under Regulation (EU) 2026/1744 (the 2026 Digital Omnibus, Official Journal text read on EUR-Lex), Article 50 was not deferred and applies from 2026-08-02, with providers of generative systems already on the market before that date required to comply with Article 50(2) marking by 2026-12-02.

### 6.2 Validated results and legal authority
- **Copyrightability in the United States.** The Copyright Office's Part 2 report (2025-01-29) concludes that outputs of generative AI can be protected only where a human author has contributed sufficient expressive elements, and that prompts alone do not suffice (U.S. Copyright Office, 2025, as summarised in its release; the full report was not read).
- **Human authorship.** The D.C. Circuit affirmed denial of registration for an image generated by an AI system with no human author, holding that the Copyright Act requires human authorship (IPWatchdog, 2025; decision of March 2025). It did not decide that all AI-assisted works are unprotectable.
- **Training on copyrighted books.** In June 2025 the Northern District of California ruled in Bartz v. Anthropic and Kadrey v. Meta that training language models on books was fair use on the records before the courts, with differing reasoning: one judge treated training as transformative and found the downloading of unauthorised copies separately problematic, while the other stressed that the outcome depended on the plaintiffs' weak showing of market harm (Jones Walker LLP, 2025, secondary summary). These are district-court rulings; the Anthropic case ended in a reported class settlement and the Meta ruling was partial. Appellate status after mid-2026 was not verified [VERIFICATION NEEDED: current appellate and settlement status]. Other pending suits (for example on news and image training data) are not assessed here.
- **Creativity experiment.** See section 4 (Doshi & Hauser, 2024).

### 6.3 Research demonstrations
- Extraction of training images from diffusion models (Carlini et al., 2023).

## 7. Advantages, Limitations & Unresolved Questions
**Advantages.** Lower cost of drafts, localisation and variants; support for less experienced creators (Doshi & Hauser, 2024); new forms of interactive media.

**Limitations and risks.**
- **Legal uncertainty.** Courts have reached fact-specific fair-use outcomes, and US copyright protection for AI-generated material is limited to human-authored contributions. Rules differ by jurisdiction.
- **Homogenisation.** Individual quality gains can reduce collective variety (Doshi & Hauser, 2024).
- **Memorisation and privacy.** Models can emit training examples, including images of identifiable people and logos (Carlini et al., 2023).
- **Labour displacement and bargaining.** Contract terms address AI (WGA, 2023), but aggregate employment effects in creative occupations were not established by the sources used here (see ECON-LABOR-001).
- **Authenticity.** Synthetic media complicates trust in news and evidence; labelling obligations are new and their effectiveness is untested.

**Unresolved questions.** Whether training is fair use or requires licensing across jurisdictions; how compensation could work; how much creative value is lost if outputs converge; whether labelling and provenance technologies are robust.

## 8. Current State of Knowledge
As of 2026-10-15: US law currently requires human authorship for copyright (Well Supported, D.C. Circuit holding and agency report). Two US district courts have found AI training on books to be fair use on their records, while the extent of lawful training remains unsettled across other media and jurisdictions (Contested). Experimental evidence shows individual creative gains with reduced collective diversity in one task setting (Preliminary). Memorisation of training data by diffusion models is demonstrated (Well Supported as a capability; prevalence in ordinary use is uncertain). Legal facts in this field change frequently; consult a current legal tracker for the status of specific cases.

## 9. Future Trajectory & Strategic Outlook
Near-term (0-3 years): appellate decisions, licensing deals and statutory labelling rules shape the market; more field evidence on creative output. Medium-term (3-10 years): potential legislative settlements on training and compensation. Long-term claims about the fate of creative professions are speculative and not rated.

| Field | Content |
| :--- | :--- |
| **Forecast ID** | `FCST-MEDIA-001` |
| **Forecast Statement** | A US federal court of appeals will issue a published merits decision on whether training a generative AI model on copyrighted works is fair use. |
| **Operational Definition** | A precedential opinion of a US court of appeals resolving the fair-use question for generative-model training (not merely a procedural or interlocutory ruling). |
| **Time Horizon** | By 2028-12-31 |
| **Assumptions** | District-court cases proceed to final judgment and are appealed; settlements do not moot all of them. |
| **Confidence Level** | Medium-High (about 65%); editorial judgement. |
| **Alternative Scenarios** | Settlements and licensing agreements moot the principal cases; legislation intervenes first. |
| **Review Schedule** | Review each October and on any appellate filing. |
| **Status** | `Active` |

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[CAP-GENAI-001] Generative AI](../vol04_capabilities/CAP-GENAI-001_generative_ai.md): Capabilities of generative models.
  - [[TECH-DIFF-001] Diffusion Models](../vol03_technical/TECH-DIFF-001_diffusion_models.md): The image and video generation technique studied for memorisation.
- **Related Methods & Technologies**:
  - [[TECH-LLM-001] Large Language Models](../vol03_technical/TECH-LLM-001_large_language_models.md): Text generation.
  - [[GOV-LAW-001] AI Governance](../vol08_safety_governance/GOV-LAW-001_ai_governance.md): General regulatory framework.
  - [[GOV-EU-001] The EU AI Act](../vol08_safety_governance/GOV-EU-001_the_eu_ai_act.md): Transparency obligations.
- **Advanced Implications & Future Trajectories**:
  - [[ECON-LABOR-001] AI, Productivity and Labor Markets](../vol09_economy/ECON-LABOR-001_ai_productivity_and_labor_markets.md): Employment effects.
  - [[APP-EDU-001] AI in Education and Knowledge Work](../vol07_applications/APP-EDU-001_ai_in_education_and_knowledge_work.md): Related questions on writing and skill.

## 11. References & Foundational Literature
1. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramèr, F., Balle, B., Ippolito, D., & Wallace, E. (2023). *Extracting Training Data from Diffusion Models*. USENIX Security Symposium 2023; arXiv:2301.13188. https://arxiv.org/abs/2301.13188. [Demonstrates memorisation and extraction of training images.]
2. Doshi, A. R., & Hauser, O. P. (2024). *Generative AI enhances individual creativity but reduces the collective diversity of novel content*. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290. [Experiment on story creativity and similarity.]
3. Epstein, Z., Hertzmann, A., & the Investigators of Human Creativity, et al. (2023). *Art and the science of generative AI*. Science, 380(6650), 1110-1111. https://doi.org/10.1126/science.adh4451. [Policy forum framing questions on authorship, creativity and copyright; text not read beyond metadata.]
4. European Union. (2024). *Regulation (EU) 2024/1689 (Artificial Intelligence Act)*. Official Journal of the European Union. http://data.europa.eu/eli/reg/2024/1689/oj. [Source of the Article 50 transparency obligations; Article content recalled and not re-read in this session. VERIFICATION NEEDED: confirm Article 50 text and application dates.]
5. IPWatchdog. (2025, March 24). *D.C. Circuit Upholds Human Authorship Requirement in Thaler v. Perlmutter*. https://ipwatchdog.com/2025/03/24/d-c-circuit-upholds-human-authorship-requirement-thaler-v-perlmutter/. [Tier 4 summary of the D.C. Circuit holding.]
6. Jones Walker LLP. (2025). *Federal Courts Find Fair Use in AI Training: Key Takeaways from Kadrey v. Meta and Bartz v. Anthropic*. https://www.jw.com/news/insights-kadrey-meta-bartz-anthropic-ai-copyright/. [Tier 4 law-firm summary of the June 2025 district-court rulings; not a substitute for the opinions.]
7. U.S. Copyright Office. (2025, January 29). *Copyright and Artificial Intelligence, Part 2: Copyrightability*. https://copyright.gov/newsnet/2025/1060.html. [Release announcing the report's conclusion on human authorship and prompts; the report itself was not read.]
8. Writers Guild of America. (2023). *Summary of the 2023 WGA MBA*. https://wgacontract2023.org/the-campaign/summary-of-the-2023-wga-mba. [Guild summary of AI provisions in the 2023 agreement.]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Diffusion models memorise and emit training images | Well Supported | Carlini et al. (2023) | High as capability; prevalence unknown |
| AI ideas raise individual creativity ratings, lower collective diversity | Preliminary | Doshi & Hauser (2024) | Moderate; one task |
| Human authorship required for US copyright | Well Supported | IPWatchdog (2025); U.S. Copyright Office (2025) | Moderate; secondary summaries |
| Training on books held fair use in two district rulings | Contested | Jones Walker LLP (2025) | Moderate; not appellate, status needs checking |
| WGA 2023 agreement AI terms | Well Supported | WGA (2023) | High for the 2023 term |
| EU Article 50 dates | Well Supported pending primary text | European Union (2024); secondary summaries | Moderate; verify |
| Appellate fair-use decision by 2028 | Forecast | `FCST-MEDIA-001` | Medium-High |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.