Patent Strategy for Software and AI: Technical Input for Stronger Patents

You're building groundbreaking AI technology. Your patent attorney asks for an invention disclosure. You draft something quickly, hand it over, and assume the legal team will handle the rest. Six months later, you receive office actions citing prior art you'd never seen. A year after that, your patent issues—but the claims are so narrow a competitor launches a nearly identical product without infringing.

This happens more often than it should. The problem isn't the patent attorney's legal expertise—it's the missing piece: technical expert strategic input during patent design.

Software and AI patents present unique challenges across all major jurisdictions. Whether you're filing at the USPTO, EPO, or UK IPO, the core issue remains the same: patent attorneys bring essential legal expertise, but these technologies require deep domain knowledge to navigate eligibility requirements, prior art landscapes, and enforcement realities.

In this article, we explain how technical experts strengthen software and AI patents through prior art analysis, claim scope strategy, and enforcement informed design. We've worked on both sides—helping to enforce patents and challenge them—and that perspective fundamentally changes how we approach patent design.

Why Technical Expertise Matters in Patent Design

Patent attorneys are experts in patent law, claim drafting, and prosecution strategy. They navigate USPTO, EPO, and UK IPO procedures, respond to examiner objections, and ensure applications meet legal requirements. This expertise is essential.

But software and AI patents present unique challenges that require deep technical domain knowledge:

• Subject matter eligibility hurdles: In the US, the Alice/Mayo framework under 35 U.S.C. § 101 requires distinguishing inventions from abstract ideas. In Europe, the EPO's "technical effect" doctrine under Article 52 EPC excludes computer programs "as such" but allows claims with technical character. Navigating these requires understanding what's technically novel and how to frame it appropriately for each jurisdiction.

• Prior art that exists outside traditional patent databases: AI innovations often build on academic research published on ArXiv, open source implementations on GitHub, or technical standards documents. Finding this prior art requires knowing where to look and how to evaluate technical similarity.

• Rapid technology evolution: Claiming specific implementations (particular neural network layers, specific code libraries) creates patents that become obsolete as technology evolves. Claiming the underlying logic and approach requires technical judgement about what's invariant across implementations.

• Provability in enforcement: A software patent that can only be proven with access to source code is significantly weaker than one provable from external system behaviour. Designing claims with enforcement in mind requires understanding both the technology and litigation realities.

We've seen patents drafted without sufficient technical input fail at multiple stages: rejected during prosecution due to poor prior art analysis, invalidated in litigation when better prior art surfaces, or designed around because claims were too narrow or didn't anticipate alternative implementations.

Technical expert input during patent design is a proactive investment that prevents these failures.

US vs Europe: Different Challenges, Same Need for Technical Expertise

United States: The Section 101 Challenge

The biggest hurdle for software patents in the US is 35 U.S.C. § 101 subject matter eligibility. Following Alice Corp. v. CLS Bank International, abstract ideas are not patentable unless the claims include "significantly more" than the abstract idea itself.

Courts apply a two step test:

1. Is the claim directed to an abstract idea?
2. If yes, does it include an inventive concept that transforms it into patent eligible subject matter?

Generic software claims like "system for processing data using machine learning" fail at step one—they're directed to the abstract idea of applying a mathematical algorithm.

The USPTO's 2024 guidance on AI inventions clarifies that claims directed to AI models or algorithms require additional elements demonstrating technical improvement or integration into a practical application. The guidance explicitly states that "claims that recite a judicial exception integrated into a practical application" are eligible, but determining what constitutes "practical application" requires technical understanding.

Europe: The Technical Effect Doctrine

The EPO takes a different approach under Article 52 of the European Patent Convention. Computer programs "as such" are excluded from patentability, but this exclusion is interpreted narrowly. If a software invention produces a "technical effect," it can be patentable.

The EPO's Guidelines for Examination state that "a computer program claimed by itself" is excluded, but "a claim directed to a technical process which process is carried out under the control of a program... is not excluded."

Key requirements for EPO software patents include:

• Technical character: The invention must have technical features that go beyond the excluded subject matter
• Further technical effect: Effects beyond the "normal" physical interaction between software and hardware—controlling industrial processes, processing signals, or improving computer functionality
• Inventive step assessed on technical features: Under the "Comvik" approach (T 641/00), only technical features contribute to inventive step

The landmark EPO Board of Appeal decision T 1358/09 (Cardinalcommerce) confirmed that "a computer implemented method that provides a further technical effect" may be patentable even if it involves excluded subject matter.

United Kingdom: Following EPO with Local Nuances

UK patent law under the Patents Act 1977 largely follows the EPO approach, but UK courts have developed their own jurisprudence. The key test from Aerotel/Macrossan 2006 EWCA Civ 1371 involves:

1. Properly construe the claim
2. Identify the actual contribution
3. Ask whether it falls solely within excluded matter
4. Check whether the contribution is technical in nature

The UK IPO's guidance on patenting AI inventions notes that "an AI invention may be patentable if it solves a technical problem using technical means or produces a technical effect."

Why Technical Expertise Bridges Both Systems

Whether facing a Section 101 rejection in the US or a technical character objection at the EPO, the solution is the same: demonstrating genuine technical innovation. This requires:

• Understanding what's technically novel versus routine implementation
• Framing the invention as solving a technical problem
• Identifying technical improvements over prior art
• Articulating the technical mechanisms, not just business outcomes

A patent attorney can draft claims that recite technical language, but understanding whether the underlying innovation actually delivers technical improvements requires domain expertise.

The Patent Design Process: Where Technical Experts Add Value

Here's how we approach patent strategy when engaged early in the process:

Phase 1: Invention Disclosure and Analysis

When we receive an invention disclosure, our first task is identifying the core inventive concept. What's truly novel here, and what's just implementation detail?

For a machine learning system, the inventive concept might be:

• A novel training data preparation method that improves model accuracy
• An architecture modification that reduces inference latency
• A technique for adapting a pre trained model to new domains with minimal additional training

It's rarely "we use machine learning to solve problem X." That's application of known technology, which faces eligibility issues in both the US (Section 101) and Europe (technical character).

We also identify multiple embodiments—different ways to implement the core invention. This serves two purposes: it supports broader claim scope (showing the invention isn't limited to one specific implementation), and it helps patent attorneys draft dependent claims that cover design around attempts.

Phase 2: Prior Art Landscape Analysis

This is where technical expertise becomes critical. We conduct comprehensive prior art searches across:

• Patent databases: USPTO, EPO, WIPO, UK IPO, national patent offices worldwide
• Academic publications: ArXiv, IEEE Xplore, ACM Digital Library, Google Scholar, Semantic Scholar
• Open source: GitHub repositories, GitLab projects, technical documentation
• Standards bodies: W3C, IEEE, IETF, ETSI technical specifications
• Industry publications: Conference proceedings (NeurIPS, ICML, CVPR), technical blogs from major technology companies

For AI and software inventions, much of the relevant prior art exists outside patent databases. We've found invalidating prior art in academic papers, open source projects, and technical standards that examiner searches missed because they're not indexed in patent search tools.

Research from the USPTO shows that AI related patent applications cite non patent literature at significantly higher rates than other technology areas. A 2023 USPTO study found that AI patents averaged 15.3 non patent literature citations compared to 6.8 for all technologies—highlighting the importance of searching beyond patent databases.

The goal isn't just to find prior art—it's to understand the technical landscape well enough to design claims that avoid what's already known whilst maintaining maximum defensible breadth.

Phase 3: Claim Scope Strategy

With prior art mapped, we work with patent attorneys to determine optimal claim scope.

Independent claims should be as broad as defensibly possible without reading on prior art. Each element must be essential. Can you remove an element and still have the invention? If yes, that element doesn't belong in the independent claim.

Dependent claims serve multiple strategic purposes:

• Fallback positions if independent claims are narrowed during prosecution or challenged in litigation
• Coverage of specific commercial implementations
• Defence against design around attempts

We think like competitors: if someone wanted to copy this invention whilst avoiding the claims, what would they change? We then draft dependent claims that cover those variations.

For example, if the independent claim recites "using a neural network," dependent claims might specify:

• "wherein the neural network is a convolutional neural network"
• "wherein the neural network is a recurrent neural network"
• "wherein the neural network comprises a transformer architecture"

This anticipates that a competitor might use a different neural network type and still achieve the same result.

Phase 4: Technical Specification Development

The specification must enable a person having ordinary skill in the art to make and use the invention. For software and AI patents, this requires:

• Sufficient algorithmic detail without being so specific that simple modifications avoid infringement
• Multiple working examples showing different embodiments
• Technical explanation of why the invention works (not just what it does)
• Figures and diagrams that clearly illustrate the technical concepts

We ensure the specification includes enough technical depth to survive enablement challenges (35 U.S.C. § 112 in the US, Article 83 EPC in Europe) whilst supporting the breadth of the claims.

Phase 5: Enforcement Perspective Review

Before finalising claims, we ask: if this patent is asserted, can infringement be proven?

For software patents, this question is critical. If the only way to prove infringement requires access to source code, enforcement becomes significantly harder and more expensive. Claims provable from external system behaviour (API calls, network traffic, user visible functionality) are stronger.

We also evaluate whether claims are clear enough to avoid indefiniteness challenges and specific enough to survive validity challenges, but broad enough to be commercially valuable.

Claiming Technical Solutions: Practical Guidance

The key to patentable software and AI claims—in both the US and Europe—is framing your invention as a technical solution to a technical problem, not just applying known technology in a new domain.

Good vs Poor Claim Framing

Poor Approach	Better Approach
"System for detecting fraud using machine learning"	"System for real time fraud detection that reduces false positives by analysing transaction patterns using ensemble of specialised classifiers, each trained on specific fraud signatures, with dynamic weighting based on transaction context"
"Method for personalising content using AI"	"Method for content personalisation comprising: generating user embeddings from behavioural sequences using attention mechanisms weighted by temporal decay; computing content relevance scores through dot product similarity in learned embedding space; and re ranking results based on diversity constraints optimised through reinforcement learning"
"System for processing data using neural network"	"Data processing system comprising distributed preprocessing modules that normalise heterogeneous data streams using format specific parsers; inference engine that applies ensemble of specialised models selected based on data characteristics detected during preprocessing; and caching layer that stores intermediate results keyed by semantic hashes of input data"

The better approaches claim specific technical mechanisms, show technical improvement (reduced false positives, improved relevance, reduced computational cost), and explain how improvements are achieved.

Function Over Code

A common mistake is claiming specific implementations that become obsolete:

Too specific: "wherein the neural network uses TensorFlow framework with ResNet 50 architecture"

Better: "wherein the neural network comprises convolutional layers for feature extraction and attention mechanisms for identifying relevant image regions"

The first ties the patent to particular software and architecture choices that may become outdated. The second claims the functional approach that remains valid regardless of implementation framework.

AI and Machine Learning: What's Actually Patentable

Not all AI innovations are patentable. Here's what typically qualifies:

Generally patentable:

• Novel training methods (new loss functions, training data preparation techniques, curriculum learning approaches)
• Model architecture innovations (new layer types, attention mechanisms, skip connections with specific technical benefits)
• Inference optimisation techniques (model compression, quantisation methods, pruning strategies with measurable performance improvements)
• AI applied to solve specific technical problems (image denoising, speech recognition in noisy environments, anomaly detection in network traffic)

Generally not patentable:

• Generic "apply machine learning to X" claims
• Using existing ML models in new business contexts without technical innovation
• Obvious combinations of known techniques
• Claims that amount to "train a model to do Y" without specifying how

The EPO Guidelines explicitly state that "AI and ML are based on computational models and algorithms" which "per se are of an abstract mathematical nature." However, they also confirm that "when an AI or ML algorithm is trained to perform a technical task... the training method may have a technical character."

Common AI Patent Pitfalls

We regularly see these mistakes:

Overly broad claims: "System for using artificial intelligence to analyse data" fails because it doesn't specify what technical problem is solved or how.

Missing technical detail: Claims that recite "machine learning model" without any architectural specifics are vulnerable to indefiniteness and enablement challenges.

Ignoring prior art: The academic AI community publishes extensively. A novel commercial application may not be novel from a patent perspective if the underlying technique was published in a research paper. ArXiv alone hosts over 50,000 machine learning papers annually.

Claiming the training rather than the trained: The EPO has noted that claims to trained models face different considerations than claims to training methods. A trained neural network may be difficult to distinguish from other trained networks, whilst the training method that produces superior results may be clearly novel and inventive.

Prior Art Informed Design: Building Stronger Patents

Filing a patent without prior art analysis is expensive gambling. Here's what we've seen happen:

A company files a broad AI patent claiming "system for content recommendation using neural networks." During prosecution, the examiner cites five prior art references—academic papers and earlier patents—showing that content recommendation using neural networks is well known. After multiple office actions and claim amendments, the patent issues with claims so narrow they provide minimal competitive protection.

If we'd conducted prior art analysis before filing, we would have:

• Identified those references immediately
• Determined what was actually novel (perhaps a specific technique for handling cold start problems or a particular attention mechanism for capturing user intent)
• Designed claims around the true innovation
• Avoided months of prosecution and narrowing amendments

What Prior Art Analysis Delivers

Higher allowance rates: Claims designed to avoid known prior art face fewer examiner rejections. USPTO data shows that applications with prior art submitted via IDS have higher ultimate allowance rates than those without.

Stronger patents: Patents that survive detailed prior art analysis are more likely to survive invalidity challenges in litigation. A study of USPTO inter partes review outcomes found that patents with comprehensive prior art consideration during prosecution faced lower invalidation rates.

Faster prosecution: Fewer office actions, fewer continuations, faster issuance. The average USPTO AI patent takes 35 months from filing to grant—thorough preparation can reduce this significantly.

Strategic portfolio building: Understanding where prior art exists shows where you can and cannot claim protection, informing both patent strategy and product development strategy.

Practical Examples: Impact of Technical Expert Input

Example 1: AI Training Method

Without technical input: "Method for training neural network to classify images"

This claim is abstract, obvious over prior art (supervised learning for image classification is well known), and provides no competitive advantage.

With technical input and prior art analysis: "Method for training image classification neural network comprising: generating augmented training samples by applying transformations selected based on predicted model uncertainty for each image region; dynamically adjusting augmentation intensity during training based on validation loss trajectory; and updating model parameters using loss function weighted by sample difficulty scores derived from attention layer outputs"

This claim specifies:

• Novel augmentation strategy (uncertainty guided)
• Dynamic adjustment mechanism (adaptive augmentation)
• Specific technical implementation details (attention based difficulty scoring)

It avoids prior art (standard data augmentation techniques), solves a technical problem (improving model generalisation), and is specific enough to be defensible yet broad enough to cover commercial implementations.

Example 2: Software System Patent

Without technical input: "System for processing user data using artificial intelligence"

This fails Section 101 in the US as abstract and lacks technical character in Europe.

With technical input: "Real time data processing system comprising: distributed preprocessing modules that normalise heterogeneous data streams using format specific parsers; inference engine that applies ensemble of specialised models selected based on data characteristics detected during preprocessing; and caching layer that stores intermediate results keyed by semantic hashes of input data to reduce redundant computations"

This claims specific technical architecture, shows technical improvements (reduced computational cost through semantic caching), and is provable from system behaviour (latency measurements, cache hit rates).

Example 3: Prior Art Informed Claim Revision

Initial disclosure: "Method for image segmentation using convolutional neural networks"

Prior art search findings: Extensive existing patents and academic publications on CNN based image segmentation. Key references include U Net (Ronneberger et al., 2015), Mask R CNN (He et al., 2017), and DeepLab (Chen et al., 2017).

Actual innovation identified: Novel boundary refinement technique using attention mechanisms that improves segmentation accuracy at object edges—a known problem in standard approaches.

Revised claims: Focus entirely on the boundary refinement method, with CNNs as background technology. Claims specify:

• How attention weights are computed from multi scale features
• How boundary pixels are identified and processed differently
• Measurable improvements in edge accuracy

Result: Patent allowed on first office action. Claims cover commercially valuable innovation and avoid extensive prior art.

Costs and Practical Realities

Patent strategy requires understanding not just what's possible, but what makes commercial sense.

Patent Filing and Prosecution Costs

Stage	UK (via UK IPO)	Europe (via EPO)	US (via USPTO)
Filing fees	£30 - £310	€1,425 - €1,630	$1,820 - $3,760
Search fees	£150	Included in filing	Included in filing
Examination fees	£100	€1,930	$800 - $1,600
Grant/issue fees	£0	€1,040	$1,200 - $2,400
Typical attorney costs	£5,000 - £15,000	€8,000 - €20,000	$15,000 - $35,000

Costs vary significantly based on complexity. AI and software patents often fall at the higher end due to subject matter eligibility considerations and extended prosecution.

Technical Expert Engagement Costs

Technical expert involvement in patent design typically adds 15 to 30 percent to overall patent costs, but the return on investment comes from:

• Reduced prosecution costs: Fewer office actions means fewer response fees and attorney hours
• Stronger patents: Higher likelihood of surviving validity challenges
• Broader claims: Better competitive protection
• Faster grant: Reduced time to enforceable rights

A study of EPO opposition proceedings found that patents drafted with comprehensive prior art consideration faced opposition at lower rates and survived opposition at higher rates than those without.

Multi Jurisdictional Filing Strategy

For technology companies, a single jurisdiction is rarely sufficient. Typical strategies include:

PCT route: File a single PCT application within 12 months of priority date, then enter national/regional phases by month 30/31. Costs approximately £2,500 - £5,000 for the international phase, plus national phase costs.

Direct filing: File directly in priority jurisdictions (typically US, Europe, and one or more Asian markets). Higher initial cost but faster prosecution in key markets.

Portfolio approach: Not every invention needs protection everywhere. Technical expert analysis helps identify which inventions warrant broad protection and which need only targeted filing.

Timeline Expectations

Milestone	US (USPTO)	Europe (EPO)	UK (UK IPO)
Filing to first office action	16 - 24 months	18 - 26 months	12 - 18 months
Average prosecution length	24 - 42 months	30 - 48 months	24 - 36 months
Accelerated examination available	Yes (Track One, prioritised)	Yes (PACE)	Yes (accelerated)
Average AI patent pendency	~35 months	~42 months	~30 months

Collaboration with Patent Attorneys: The Optimal Workflow

Technical experts don't replace patent attorneys—we complement them. The strongest patents result from true collaboration.

Our typical workflow:

1. Invention disclosure → We review with inventors to understand core innovation
2. Technical analysis → We conduct prior art search, identify novel aspects, assess eligibility issues for target jurisdictions
3. Strategy session → We meet with patent attorney to discuss claim scope, embodiments, prosecution strategy
4. Patent attorney drafts application → Leveraging legal expertise and our technical input
5. Technical review → We review draft for technical accuracy, completeness, claim support
6. Joint refinement → We collaborate on revisions before filing

This combines legal expertise (claim language, prosecution strategy, procedural requirements) with technical expertise (prior art knowledge, domain understanding, enforcement considerations).

When to Engage Technical Experts

Ideal timing: Before the patent attorney drafts the application. Prior art analysis and technical strategy inform drafting, not retrospective fixes.

Minimum engagement: Technical review of draft application before filing. We can still identify technical issues, missing embodiments, or claim scope problems.

Also valuable: During prosecution when facing difficult office actions (particularly Section 101 rejections in the US or technical character objections at the EPO), or when developing patent portfolio strategy across multiple related inventions.

What Not to Do: Common Strategic Mistakes

Based on our experience reviewing hundreds of software and AI patents, here are mistakes to avoid:

Mistake	Consequence	Better Approach
Filing without prior art search	Claims narrowed during prosecution, weak final patent	Comprehensive search before drafting
Claiming specific frameworks or libraries	Patent obsolete when technology evolves	Claim functional approaches, not implementations
Generic AI/ML terminology	Section 101 rejection (US), technical character objection (EPO)	Specific technical mechanisms and improvements
Single independent claim	One successful challenge invalidates key protection	Multiple independent claims covering different aspects
Specification limited to single embodiment	Narrow claim scope, enablement challenges	Multiple embodiments showing breadth of invention
Ignoring enforcement practicalities	Patent valid but unenforceable without source code access	Claims provable from external behaviour

Conclusion: Invest in Strong Patent Design

Software and AI patents are too valuable and too expensive to leave to chance. A well designed patent—informed by both legal expertise and deep technical knowledge—protects your innovation, withstands challenges, and provides real competitive advantage.

We've seen the difference technical expert input makes. Patents designed with prior art awareness, strategic claim scoping, and enforcement considerations are simply stronger. They're more likely to be allowed, more likely to survive validity challenges, and more valuable when asserted or licensed.

The investment in robust patent design is far smaller than the cost of weak patents that fail when you need them most. Whether facing Section 101 challenges in the US or technical character requirements in Europe, the solution is the same: demonstrating genuine technical innovation through claims that reflect deep understanding of both the technology and the prior art landscape.

This is general information about patent strategy, not legal or technical advice for specific situations. Always work with qualified patent attorneys for legal guidance, and consider engaging technical experts who understand both your technology domain and patent enforcement realities.

---

Sources

Primary Legal Sources: US

1. 35 U.S.C. § 101 — Inventions patentable (subject matter eligibility statute). https://www.law.cornell.edu/uscode/text/35/101

2. 35 U.S.C. § 102 — Conditions for patentability; novelty. https://www.law.cornell.edu/uscode/text/35/102

3. 35 U.S.C. § 103 — Conditions for patentability; non obvious subject matter. https://www.law.cornell.edu/uscode/text/35/103

4. 35 U.S.C. § 112 — Specification requirements (enablement, written description, best mode). https://www.law.cornell.edu/uscode/text/35/112

5. Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014) — Supreme Court decision establishing framework for software patent eligibility. https://supreme.justia.com/cases/federal/us/573/208/

6. Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012) — Supreme Court decision on patent eligibility for laws of nature. https://supreme.justia.com/cases/federal/us/566/66/

7. Diamond v. Diehr, 450 U.S. 175 (1981) — Supreme Court decision holding software controlling industrial process is patentable. https://supreme.justia.com/cases/federal/us/450/175/

Primary Legal Sources: Europe

8. European Patent Convention (EPC), Article 52 — Patentable inventions and excluded subject matter. https://www.epo.org/en/legal/epc/2020/a52.html

9. European Patent Convention (EPC), Article 83 — Disclosure of the invention (enablement). https://www.epo.org/en/legal/epc/2020/a83.html

10. EPO Guidelines for Examination, Part G Chapter II — Inventions (including AI and software). https://www.epo.org/en/legal/guidelines-epc/2024/g_ii.html

11. EPO Board of Appeal Decision T 641/00 (Comvik) — Approach to assessing inventive step for mixed technical/non technical inventions. https://www.epo.org/en/boards-of-appeal/decisions/t000641eu1

12. EPO Board of Appeal Decision T 1358/09 (Cardinalcommerce) — Computer implemented inventions and technical effect. https://www.epo.org/en/boards-of-appeal/decisions/t091358eu1

13. EPO Board of Appeal Decision T 0154/04 (Duns Licensing) — Technical character of computer programs. https://www.epo.org/en/boards-of-appeal/decisions/t040154eu1

Primary Legal Sources: UK

14. Patents Act 1977 (UK), Section 1 — Patentable inventions. https://www.legislation.gov.uk/ukpga/1977/37/section/1

15. Aerotel Ltd v Telco Holdings Ltd & Macrossan's Patent Application 2006 EWCA Civ 1371 — Court of Appeal test for excluded subject matter. https://www.bailii.org/ew/cases/EWCA/Civ/2006/1371.html

16. HTC Europe Co Ltd v Apple Inc 2013 EWCA Civ 451 — Court of Appeal on technical contribution for software patents. https://www.bailii.org/ew/cases/EWCA/Civ/2013/451.html

17. UK IPO, "Examining patent applications relating to artificial intelligence (AI) inventions" (2022) — Official examination guidance. https://www.gov.uk/government/publications/examining-patent-applications-relating-to-artificial-intelligence-ai-inventions

USPTO Guidance and Statistics

18. USPTO, "2019 Revised Patent Subject Matter Eligibility Guidance" — Framework for examining Section 101 issues. 84 Fed. Reg. 50 (January 7, 2019). https://www.federalregister.gov/documents/2019/01/07/2018-28282/2019-revised-patent-subject-matter-eligibility-guidance

19. USPTO, "Guidance Update on AI and Other Emerging Technologies" (2024) — Current examination guidance for AI inventions. https://www.uspto.gov/patents/laws/examination-policy/subject-matter-eligibility

20. USPTO, "Inventing AI: Tracing the diffusion of artificial intelligence with U.S. patents" (2020) — Statistics on AI patent trends and citations. https://www.uspto.gov/sites/default/files/documents/OCE-DH-AI.pdf

21. USPTO, "Patent Examination Research Dataset" (PERD) — Source data on prosecution timelines and outcomes. https://www.uspto.gov/ip-policy/economic-research/research-datasets

22. USPTO, "AI/ML Patent Landscape Study" (2023) — Analysis of AI patent applications and allowance rates. https://www.uspto.gov/about-us/news-updates/ai-patent-landscape

EPO Guidance and Statistics

23. EPO, "Guidelines for Examination: Artificial Intelligence and Machine Learning" (G II 3.3.1) — Official examination approach for AI inventions. https://www.epo.org/en/legal/guidelines-epc/2024/g_ii_3_3_1.html

24. EPO, "Patents and the Fourth Industrial Revolution" (2020) — Statistics on AI patent filings and grant rates. https://www.epo.org/en/news-events/press-centre/news/2020/20201214

25. EPO, "Patent Index 2023" — Annual statistics on patent filings by technology area. https://www.epo.org/en/about-us/statistics/patent-index-2023

26. EPO, "Opposition statistics" — Data on opposition rates and outcomes by technology area. https://www.epo.org/en/about-us/statistics/opposition

Academic and Research Sources

27. Ronneberger, O., Fischer, P., & Brox, T. (2015). "U Net: Convolutional Networks for Biomedical Image Segmentation." MICCAI 2015. https://arxiv.org/abs/1505.04597 — Foundational image segmentation architecture frequently cited as prior art.

28. He, K., Gkioxari, G., Dollar, P., & Girshick, R. (2017). "Mask R CNN." IEEE ICCV 2017. https://arxiv.org/abs/1703.06870 — Instance segmentation architecture commonly appearing in AI patent prior art searches.

29. Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2017). "DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs." IEEE TPAMI. https://arxiv.org/abs/1606.00915 — Semantic segmentation approach frequently cited.

30. Vaswani, A., et al. (2017). "Attention Is All You Need." NeurIPS 2017. https://arxiv.org/abs/1706.03762 — Transformer architecture paper cited in most modern AI patent applications.

31. Samuelson, P. (2016). "Freedom to Tinker." Theoretical Inquiries in Law, Vol. 17. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2837244 — Analysis of software patent scope and enforcement.

32. Lemley, M. A., & Feldman, R. (2016). "Patent Licensing, Technology Transfer, and Innovation." American Economic Review. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2747213 — Empirical study of patent value and licensing.

Industry and Practice Sources

33. WIPO, "Patent Landscape Report: Artificial Intelligence" (2019) — Comprehensive analysis of global AI patent filing trends. https://www.wipo.int/publications/en/details.jsp?id=4386

34. WIPO, "Technology Trends 2019: Artificial Intelligence" — Analysis of AI patent portfolios by company and technology area. https://www.wipo.int/tech_trends/en/artificial_intelligence/

35. UKIPO, "AI and IP: copyright and patents" Call for Views Government Response (2023) — UK policy position on AI patentability. https://www.gov.uk/government/consultations/artificial-intelligence-and-ip-copyright-and-patents

36. Lewis, T. & McCracken, S. "Patenting AI in Europe: EPO Examination Approach" (2024). IAM Magazine. — Practical analysis of EPO AI patent prosecution.

37. Intellectual Property Office, "Facts and figures: patent, trade mark, design and hearing data" (2024) — Official UK patent statistics. https://www.gov.uk/government/collections/intellectual-property-office-facts-and-figures-patent-trade-mark-design-and-hearing-data

38. ArXiv.org machine learning category statistics — Data on annual ML paper submissions demonstrating prior art volume. https://arxiv.org/list/cs.LG/recent

Fee and Cost Sources

39. USPTO, "USPTO Fee Schedule" (2024) — Official patent fee information. https://www.uspto.gov/learning-and-resources/fees-and-payment/uspto-fee-schedule

40. EPO, "Fees, payments and refunds" (2024) — Official European patent fee schedule. https://www.epo.org/en/applying/fees

41. UK IPO, "Patent fees" (2024) — Official UK patent fee information. https://www.gov.uk/government/publications/patent-forms-and-fees

42. AIPLA, "Report of the Economic Survey" (2023) — Industry data on patent attorney costs in the US.

43. CIPA, "IP Cost Survey" (2023) — Industry data on patent costs in the UK.