5.1 AI as Inventor: The Patent Question
Can AI be named as inventor on a patent? This question has been litigated globally with consistent result: current patent law requires human inventors. Understanding this is crucial for AI patent strategy.
Facts: Dr. Stephen Thaler sought patents for inventions created by his AI system "DABUS" (Device for the Autonomous Bootstrapping of Unified Sentience). He listed DABUS as inventor.
Held: An "inventor" under the Patents Act must be a natural person. DABUS is not a person and cannot be named as inventor. Patent applications rejected.
Reasoning: The Act assumes inventors have rights that can be transferred - AI cannot hold or transfer rights. Parliament did not intend machines to be inventors.
Global Consensus on AI Inventorship
| Jurisdiction | Decision | Reasoning |
|---|---|---|
| UK | Rejected (Thaler 2023) | Inventor must be natural person |
| USA | Rejected (Thaler 2023) | 35 USC refers to "individuals" |
| EPO | Rejected (2022) | EPC requires human inventor |
| Australia | Initially allowed, reversed (2022) | High Court: inventor must be human |
| India | No ruling; follows UK position | Patent Act mirrors UK terminology |
Indian Patent Act Analysis
While the Indian Patent Act does not explicitly define "inventor" as human, several provisions imply human agency:
- Section 6: Application by "any person claiming to be true and first inventor"
- Section 10(4)(a): Declaration of inventorship required
- Section 28: Rights of employees - assumes human employment
India is likely to follow the UK/Thaler approach given the similar legislative language. The Indian Patent Office has not issued formal guidance, but practitioners should assume AI cannot be named as inventor.
Practical Patent Strategy for AI Inventions
- Name Human Inventor: Person who conceived the problem, designed the AI, or selected the output
- Document Human Contribution: Maintain records of human creative input
- Claim AI as Tool: Describe AI as instrument used by human inventor
- Protect AI System: Separately patent the AI methodology, not just outputs
When advising AI companies, recommend: (1) Establish invention disclosure processes identifying human contributors, (2) Document human oversight and selection of AI outputs, (3) Consider trade secret protection for AI-generated innovations where human contribution is minimal.
5.2 Copyright in AI-Generated Works
Can AI-generated content receive copyright protection? Who owns copyright when AI creates art, music, or text? Indian copyright law presents unique challenges.
Indian Copyright Act, 1957
Unlike many jurisdictions, India explicitly recognizes "computer-generated works" in Section 2(d)(vi). This potentially covers AI-generated content, with authorship attributed to "the person who causes the work to be created."
Who "Causes" AI Work to be Created?
The phrase "causes the work to be created" is ambiguous. Potential claimants:
- AI Developer: Created the system enabling generation
- Prompt Engineer: Provided input causing specific output
- Platform Operator: Deployed and maintained the AI system
- End User: Initiated the generation request
Originality Requirement
"Copyright subsists in original literary, dramatic, musical and artistic works." Section 13(1)(a), Copyright Act 1957
Does AI output meet originality standards?
- Eastern Book Company v. D.B. Modak (2008): "Skill and judgment" standard, not mere labor
- Question: Can AI exercise "judgment"? Is prompt engineering sufficient creativity?
- Trend: Courts may require human creative choices for originality
US Copyright Office Position (Comparative)
The US Copyright Office has rejected AI-generated works lacking human authorship:
- Zarya of the Dawn (2023): AI-generated images denied copyright; human text retained
- Theater D'Opera Spatial (2023): AI-generated artwork denied registration
- Guidance (2023): AI output only copyrightable if human creativity involved
AI-generated content may receive weaker or no copyright protection. Advise clients: (1) Document human creative contributions, (2) Consider contractual protections, (3) Use trade secrets for commercial AI outputs where copyright uncertain.
5.3 Training Data & Copyright
AI systems are trained on vast datasets, often including copyrighted material. This raises critical questions about infringement, fair dealing, and licensing requirements.
The Training Data Problem
AI training typically involves:
- Data Collection: Scraping web content, books, images (often without permission)
- Reproduction: Copying works into training datasets
- Model Training: Using works to create model parameters
- Output Generation: Creating content potentially similar to training data
Copyright Infringement Analysis
1. Reproduction in Training
Section 14(a) grants exclusive right to reproduce the work. Training AI on copyrighted works involves reproduction. Is this infringement?
2. Fair Dealing Defence (Section 52)
| Fair Dealing Ground | AI Training Application | Viability |
|---|---|---|
| Research (S.52(1)(a)) | AI development as research | Weak - commercial research limited |
| Private Use (S.52(1)(a)) | Internal training use | Weak - commercial use problematic |
| Criticism/Review | Not applicable | No |
| Transient Copy (S.52(1)(b)) | Temporary processing | Possible - if truly transient |
India's fair dealing provisions are narrower than US fair use. There is no "transformative use" doctrine. AI training on copyrighted works without license is a significant legal risk in India.
Global Litigation Landscape
Claim: Stability AI trained Stable Diffusion on 12 million Getty images without license. Getty alleges copyright and trademark infringement.
Status: Pending. Key test case for AI training practices.
Claim: OpenAI trained ChatGPT on NYT articles. AI can reproduce substantial portions, undermining NYT subscription model.
Status: Pending. May determine scope of AI training fair use.
Licensing Solutions
To mitigate infringement risk, AI developers should consider:
- Licensed Datasets: Use properly licensed training data
- Synthetic Data: Train on AI-generated or synthetic datasets
- Open Source Content: Use Creative Commons, public domain works
- Data Licensing Agreements: Negotiate training rights from content owners
- Content Filtering: Exclude copyrighted content from training
5.4 Patenting AI Innovations
Beyond the inventorship question, AI-related patents face additional challenges under Indian patent law, particularly Section 3(k) exclusions.
Section 3(k) - Computer Program Exclusion
AI algorithms are essentially computer programs. How to navigate Section 3(k)?
Indian Patent Office Approach
CRI Guidelines (2017) and subsequent clarifications indicate:
- "Per se" interpretation: Software with technical effect may be patentable
- Technical Contribution: AI solving technical problem may qualify
- Hardware Integration: AI embedded in physical device more likely patentable
Patentable AI Claim Strategies
- Technical Effect Claims: Focus on technical improvement achieved by AI
- Hardware-Software Integration: Claim AI as part of physical system
- Method Claims: Process claims with technical steps beyond mere algorithm
- Training Data Processing: Novel data processing methods
- Architecture Claims: Novel neural network architectures
When drafting AI patent claims: (1) Emphasize technical problem and solution, (2) Include hardware elements where possible, (3) Describe technical effects beyond mere computation, (4) Avoid claiming "algorithm" or "method" without technical context.
Section 3(d) - Incremental Innovation
AI improvements may face Section 3(d) challenges if considered mere "enhancement of known efficacy." Ensure claims demonstrate genuinely new technical effect, not just improved performance.
5.5 Trade Secret Protection for AI
Given challenges with patents and copyright, trade secrets may offer the most robust protection for AI innovations in India.
Trade Secret Protection for AI
Elements protectable as trade secrets:
- Training Data: Curated, proprietary datasets
- Model Architecture: Specific neural network designs
- Hyperparameters: Training configurations, optimization settings
- Weights: Trained model parameters
- Prompt Engineering: Effective prompt formulations
Maintaining Trade Secret Status
- Confidentiality Measures: NDAs with all personnel accessing AI systems
- Access Controls: Limited access to model weights, training data
- Technical Protection: Encryption, secure deployment practices
- Documentation: Record confidentiality measures taken
- Exit Procedures: Ensure departing employees don't take AI knowledge
AI models can be reverse-engineered through "model extraction attacks" - querying the model to recreate its functionality. Consider: (1) Rate limiting API access, (2) Detecting extraction attempts, (3) Watermarking outputs, (4) Contractual prohibitions on extraction.
Key Takeaways
- AI Inventorship: AI cannot be named inventor (Thaler); name human contributors
- Copyright: India's "computer-generated work" provision may cover AI outputs; authorship goes to person causing creation
- Training Data: Using copyrighted works for training is legally risky; fair dealing defence weak in India
- Patents: Section 3(k) challenges; focus on technical effect, hardware integration
- Trade Secrets: May offer strongest protection; implement robust confidentiality measures
- Document human creative contributions for all IPR claims
- Consider multi-pronged protection: patents for architecture, trade secrets for weights, contracts for commercial terms