Robert Plotkin's "Genie and Aladdin," the AI Spellbook strategy,
and a teleological redesign of patent requirements and specification reproducibility
Generative AI is no longer just a tool that helps an inventor crunch numbers. It analyzes technical problems, explores a vast space of alternatives, and proposes structures, compounds, algorithms, and product configurations that no human would have anticipated. Humans, in turn, select the most promising outputs and turn them into working technology through testing and validation.
When a substantial share of invention now happens not inside a single human mind but in the interaction between humans and AI, the traditional questions of patent law have to change too. The more fundamental question becomes this:
To whom should the patent system grant an exclusive incentive, and of what kind, in order to most effectively spur the development, commercialization, and disclosure of technology?
Framed around that question, inventorship, non-obviousness, patent eligibility, and the enablement requirement in the AI era can all be reorganized within a single teleological framework — one that judges each requirement by the policy purpose the patent system is actually meant to serve. The person standing at the origin of that reframing is Robert Plotkin, an American patent attorney specializing in software and AI.
1. Who Is Robert Plotkin?
Robert Plotkin studied computer science at MIT and has spent more than 25 years working on software and AI-related patents as a U.S. patent attorney. He is currently a co-founder of Blueshift IP, a firm specializing in AI and software intellectual property, where he advises clients on identifying, patenting, and commercializing AI technology and protecting it as trade secrets.
What makes Plotkin worth paying attention to is that he was studying the possibility of computers automating the inventive process itself — not just assisting with it — long before generative AI went mainstream. In his 2009 book from Stanford University Press, The Genie in the Machine: How Computer-Automated Inventing Is Revolutionizing Law and Business, he analyzed how an era in which software generates the designs for new products would transform both patent law and corporate strategy.
The book's table of contents includes chapters such as "The Rise of Wishes," "Follow the Value," "Free the Genie, Bottle the Wish," "Learning How to Wish," and "Advising Aladdin." In other words, Plotkin's argument was never simply that AI is an invention tool — it centers on what humans should ask of an automated invention system, and how they should control the value of what it produces.
In 2024 he published AI Armor: Securing the Future of Your AI Company with Strategic Intellectual Property, extending that same line of thinking into an IP strategy for generative AI companies. Plotkin argues that AI systems, trained models, data, processes, prompts, and outputs should not be treated as a single undifferentiated bundle — each should be broken out and matched to the right protection mechanism, whether that is a patent, a trade secret, or copyright.
2. Genie and Aladdin: Invention Shifts from "Design" to "Wish"
Plotkin compares an automated invention system to a "Genie," and the human who uses it to "Aladdin." The human inventor states the outcome they want — a "Wish." The Genie, the AI, searches a vast design space and generates a concrete solution that satisfies that wish.
In this framework, humans no longer have to design every component and algorithm themselves. Instead, they perform higher-level activities such as the following:
- Choosing which problem to solve
- Defining the technical goal
- Setting the constraints the result must satisfy
- Constructing the design space the AI will search
- Evaluating the generated results
- Recognizing which results are technically meaningful
- Refining results through experimentation and feedback
- Deciding which results to protect as patents or trade secrets
Plotkin does not see this as the end of traditional invention. Stanford University Press's own description of the book describes the human who wields automated-invention technology as a "digital renaissance artisan" — someone who will use it to boost their inventive abilities to heights that were previously unimaginable, enabling small teams and even individual consumers to compete with much larger organizations. The essence of invention, in other words, gradually shifts from "the ability to personally design every last detail" to "the ability to define a good wish, and to identify, select, and verify the Genie's output."
3. In the AI Era, Patent Law Should Look Not at AI's Creativity but at "Who Responds to the Incentive"
Viewed teleologically, the patent system looks different. A patent is a policy instrument that induces R&D investment, commercialization, and disclosure by granting a time-limited exclusive right. What matters is not who experienced a creative spark, but who actually responds to the institutional incentive a patent provides.
AI does not work any harder because it expects to receive a patent. It does not abandon its research because an application was rejected, and it does not reinvest the profits of a patent monopoly into the next round of R&D. Human researchers, companies, investors, and businesses, by contrast, decide how much to spend on R&D, whether to disclose their technology, and whether to risk a competitor's imitation — all based on whether patent protection is available.
4. Rearranging Four Requirements of U.S. Patent Law Through a Teleological Lens
4.1 Inventorship: Tracing the Human's Contribution to Conception, Not the AI's
Under U.S. law, only a natural person can be an inventor. The USPTO's 2025 revised inventorship guidance states that the same inventorship standard applies to AI-assisted inventions as to any other invention, with no separate, relaxed, or heightened test. AI is treated as a tool used by a human, no different in kind from laboratory equipment, software, or a research database.
The question is not whether AI participated in the invention, but which natural person made a substantial contribution to the conception of the claimed invention.
As a practical matter, the human contributions that matter most include the following:
- Framing the technical problem in a non-routine way
- Setting the physical or numerical constraints the result must satisfy
- Structuring the AI's search space or objective function
- Selecting a specific combination of model and data
- Being the first to recognize the technical value of an AI output
- Making engineering modifications to that output
- Redesigning structures or conditions based on experimental feedback
- Making the choices and judgment calls that produced the final claimed technical configuration
Mere ownership, funding, running the AI system, or after-the-fact approval of a result is not sufficient, by itself, to establish inventorship.
The Inventorship Evidence Ledger
In AI-assisted R&D, preserving only the final result makes it hard to reconstruct, after the fact, exactly how a human contributed to conception. Companies need to maintain an Inventorship Evidence Ledger that accumulates the following materials in chronological order:
- The original problem-definition document / prompt history and revisions / model version and runtime environment
- Input constraints / the AI's raw outputs / which candidates were adopted or discarded, and why
- Technical modifications made by humans / experimental results and feedback / meeting notes attributed to individual inventors
- A mapping between the final claim language and each human contribution
This is more than a lab notebook — it becomes an evidentiary record that can substantiate a human's contribution to conception in future inventorship disputes, invalidity litigation, technology transfers, and due diligence.
4.2 Non-Obviousness Under §103: Should the Person of Ordinary Skill Be Assumed to Use AI?
35 U.S.C. §103 provides that a patent may not issue if the difference between the claimed invention and the prior art would have been obvious to a person having ordinary skill in the relevant field, and it explicitly states that patentability shall not be negated by the manner in which the invention was made. An examiner's argument that "this invention is obvious because it was made using AI" is, in principle, improper.
At the same time, once AI becomes a ubiquitous research tool in a given field, the tools available to — and the problem-solving capability of — the person of ordinary skill in that field necessarily change as well. In his 2025 article "AI and the Level of Ordinary Skill," Plotkin argued that courts should account for the routine use of AI tools within a given field when calibrating the level of ordinary skill.
At the same time, the "AI-augmented PHOSITA" (person having ordinary skill in the art) must not become an abstract superhuman construct. At a minimum, an examiner should establish the following:
- Whether the relevant AI tool actually existed as of the effective filing date
- Whether a person of ordinary skill in that field had access to the tool
- Whether using the tool was actually standard industry practice
- Whether the prior art supplied the necessary inputs and constraints
- Whether an ordinary prompt or input would have reached the claimed invention
- Whether the output could be relied on as technically sound without further experimentation
- Whether there was a motivation to combine the prior art with a reasonable expectation of success
4.3 Patent Eligibility Under §101: Judge the Technical Improvement, Not Whether AI Was Used
AI-related inventions frequently run into §101 eligibility rejections on the theory that they are merely mathematical algorithms or abstract ideas. But viewed teleologically, the question §101 should be asking is not "was AI used in this invention?" The more essential question is this:
Does the claimed invention merely use AI to automate an abstract goal or a task humans already performed, or does it substantially improve the operation of a computer, a machine-learning model, or some other concrete technical system?
If the patent system is an inducement mechanism for R&D, it needs to distinguish between simply applying a known AI to a new task or dataset, and an invention that overcomes a genuine technical limitation of the AI system itself. The former is likely to emerge naturally from the ordinary commercial use of AI in the marketplace, patent or no patent. The latter, by contrast, requires new R&D investment and can drive further downstream technical progress.
The Federal Circuit's 2025 decision in Recentive Analytics, Inc. v. Fox Corp. illustrates the former category. The claims at issue covered using machine learning to generate broadcast and event schedules and update them in real time. The court held that iterative training, real-time data updates, and improved accuracy were simply routine characteristics of how machine learning already operates, and that the claims did not identify any specific technical means for achieving those results. The USPTO Appeals Review Panel's 2025 decision in Ex parte Desjardins, by contrast, illustrates the latter category. That invention adjusted parameter values, during training on a new task, so as to protect the model's performance on a prior task — overcoming the problem known as "catastrophic forgetting" in continual-learning systems, and thereby improving the training mechanism itself. Chapter 7 compares the two cases in detail.
A teleologically reconstructed §101 inquiry should therefore ask:
- Is this simply automating a task humans already performed?
- Is this nothing more than applying a known AI to a new data environment?
- Does the claim recite only the functional result, while omitting the technical mechanism that produces it?
- Does it improve the function or performance of the AI or computer system itself?
- Does it change the operation of a concrete physical system or engineering process?
- Is a mathematical concept integrated into a genuine practical technical application?
4.4 The Enablement Requirement Under §112: Teach a Reproducible Invention, Not the AI's Black Box
AI systems are often opaque by nature. Even the human engineers who built them may not be able to fully explain exactly which internal pathway, among countless parameters, produced a particular result. But the enablement requirement of 35 U.S.C. §112 does not demand a philosophical or mathematical account of every micro-level computation that led the AI to a given conclusion. Viewed teleologically, the core question §112 asks is this:
In exchange for the patent monopoly, has society received enough technical teaching that a person of ordinary skill could make and use the claimed invention without undue experimentation?
Patent law presupposes an exchange between disclosure and exclusivity — a quid pro quo. Applicants for AI inventions therefore cannot simply disclose the fact that AI produced a particular result. They must supply the conditions and procedures needed for a person of ordinary skill to reproduce the same, or a substantially equivalent, technical result. The following information matters in an AI-related specification:
- The type of model used and its core architecture / the technical rationale for choosing it
- The nature and format of the input data / data preprocessing and feature-extraction methods
- How the training, validation, and test data were structured and split
- Key hyperparameter values or ranges / the objective and loss functions / the criteria used to evaluate outputs
- Iteration and feedback procedures / criteria for selecting, discarding, and regenerating outputs
- Points at which a human expert conducted technical review / representative working examples and comparative examples
- Failure modes and operating limits / alternative models and alternative embodiments
- Reproducibility data supporting the full scope of the claims
4.5 A Teleological Reconstruction of the Four Requirements
These four requirements are not isolated rules. Together they form a single, coherent structure of patent policy.
| Requirement | The AI-Centered Question Traditionally Asked | The Teleologically Reconstructed Question |
|---|---|---|
| Inventorship | How creatively did the AI contribute? | Which natural person substantially contributed to the conception of the claimed invention, and to whom should the right be granted to induce development and disclosure? |
| Non-Obviousness (§103) | How difficult was it for the human to use the AI? | Would this have naturally emerged from the ordinary technical environment of the time — AI included — even without the patent inducement? |
| Patent Eligibility (§101) | Does the claim recite AI or an algorithm? | Does the claim automate an abstract purpose, or does it concretely improve how an AI, computer, or technical system operates? |
| Enablement (§112) | Can every internal AI computation be explained? | In exchange for exclusivity, has enough technical teaching been provided for a person of ordinary skill to reproduce the invention? |
5. Responding to a §103 Rejection for "an Obvious Invention Made Using Too Much AI"
When an examiner raises a rejection premised on the idea that a general-purpose AI could easily have reached the claimed result, the applicant should not respond by emphasizing the inventor's effort or the literary creativity of the prompt. Section 103 does not compensate an inventor's subjective toil. An effective response can be built around four steps.
Separate the Use of AI From Obviousness
Make clear that the mere use of AI does not, by itself, establish obviousness. The examiner still bears the burden of showing that every limitation of the claim was disclosed in, or would have been obvious to combine from, the prior art.
Identify the Structural Gap Between a Routine AI Output and the Claimed Invention
The applicant should separate "what the AI produced" from "the final claimed invention."
- The AI suggested a general shape, but the claimed stress-distribution structure came from the human engineer's subsequent analysis.
- The AI proposed candidate compounds, but the specific substituent ranges and dosing conditions were fixed only through experimental results.
- The AI proposed a classification model, but the claimed loss function, penalty term, and data-normalization relationship were designed by a human.
- The AI proposed several control schemes, but the specific closed-loop relationship between a given sensor signal and actuation condition was determined only after repeated testing.
Explain the Absence of a Reasonable Expectation of Success
The fact that AI can output a candidate does not by itself mean a person of ordinary skill could reasonably have expected the claimed invention to succeed. In chemistry, pharmaceuticals, biotechnology, and complex mechanical systems, an AI output does not guarantee actual operability, stability, non-toxicity, durability, or manufacturability. Where substantial experimentation and validation were required after the AI's output, that step should not be treated as a simple, obvious optimization.
Present Unexpected Results as Objective Data
One of the strongest forms of objective evidence against an obviousness rejection is an unexpected result. Three elements are needed.
- A comparison against the closest prior art or the routine AI output
- A statistically or technically significant difference
- A nexus showing that the difference arises from the claimed configuration
6. The Technical Reproducibility Data Package: Preparing for a §103 Response at the Drafting Stage
Bringing together Plotkin's human-AI collaboration theory, AI asset mapping, prompt patenting, the contrast between Recentive and Desjardins, the objective evidence relevant to §103, and the enablement principles of §112 yields a practical recommendation that can be called the Technical Reproducibility Data Package.
The Technical Problem and Baseline Data
A specification should not vaguely state that it "improves performance." It should define the shortcomings of the existing technology in measurable terms. For example:
- The existing model's accuracy on a prior task drops by a certain percentage after continual learning.
- The existing structure shows a sharp spike in vibration at a particular frequency above a certain rotational speed.
- A conventional AI design exceeds the allowable peak-stress threshold by a certain margin.
- An existing classification model produces a certain false-positive rate on a particular data subgroup.
AI Input and Control Structure
Reciting the prompt language alone may not be enough. The specification should describe the entire input structure that affects reproducibility: the system prompt, the user prompt, the sequence of step-by-step queries, input data formats, variable ranges, required and excluded conditions, model type and version, inference settings such as temperature and top-p, the criteria for selecting and discarding outputs, regeneration conditions, and points of human review. Crucially, those settings must be causally connected to the claimed technical effect.
Data Preprocessing and the Model-Improvement Mechanism
It is not enough to say that data was fed into a general-purpose AI. The invention's core may lie in the specifics of missing-value handling, feature extraction, noise removal, normalization, vectorization, data splitting, bias correction, modifications to the objective function, loss- function design, the addition of a penalty term, learning-rate control, stopping criteria, cross- validation across multiple models, or physical feedback from the output.
Control-Group vs. Experimental-Group Comparison
The most persuasive structure directly compares the general-purpose AI's typical result against the result after the inventor's technical intervention. The control group is the industry-standard model, its default settings, a routine prompt, and known data preprocessing; the experimental group is the inventor's step-by-step prompt design, special constraints, proprietary data preprocessing, objective-function or model control, human experimental feedback, and final technical modifications. Comparison metrics can include accuracy, error rate, energy consumption, computational cost, memory footprint, convergence speed, durability, strength, yield, and toxicity.
Ablation Studies and Failure Data
To demonstrate the technical contribution of a specific element, it is useful to show how performance changes when that element is removed or altered. Ablation data — removing the penalty term, omitting a specific preprocessing step, dropping one prompt stage, deviating from a constraint range, skipping human feedback, or substituting a different model — reinforces both the §103 nexus and §112 enablement simultaneously.
7. Contrasting §101 Precedent: Recentive and Desjardins
The §101 standard developed teleologically in Chapter 4.3 is thrown into sharp relief by two 2025 decisions. Recentive Analytics, Inc. v. Fox Corp. and Ex parte Desjardins show, in concrete terms, what determines patent eligibility within the same broad AI technology landscape.
7.1 Recentive Analytics v. Fox: A New Application of General-Purpose AI
In 2025, the Federal Circuit denied patent eligibility to claims that applied general-purpose machine learning to broadcast and event scheduling. The court held that iterative training, real-time data incorporation, and improved accuracy were routine functions of machine learning, and that the claims did not sufficiently identify any specific technical means for achieving those results.
Applying a known machine-learning technique to a new dataset or task domain is not, by itself, a technical improvement to the AI system.
7.2 Ex Parte Desjardins: Improving the Model's Own Operating Mechanism
The USPTO Appeals Review Panel, by contrast, recognized §101 eligibility in 2025 for a claim that improved a continual-learning method for a machine-learning model. The claim adjusted parameter values, while learning a new task, so as to protect the model's performance on a prior task — overcoming the "catastrophic forgetting" problem common in continual-learning systems. The USPTO explained that software and AI improvements need not manifest as changes to physical components; an improvement to a logical structure or process can itself be a technical improvement.
| Recentive Analytics v. Fox | Ex Parte Desjardins | |
|---|---|---|
| Nature of the Technology | New task application of general-purpose ML | Improved ML training mechanism |
| Problem Solved | Making a human task more efficient | A limitation of the model itself (catastrophic forgetting) |
| Claiming Style | Result / function oriented | Computation / control mechanism oriented |
| Technical Effect | Faster scheduling | Reduced storage and complexity, knowledge preservation |
| §101 Outcome | Ineligible | Eligible |
What an AI specification needs to explain, in other words, is not that "the AI makes better decisions," but which data, parameters, objective function, or control relationship actually changes how a computer or physical system operates.
8. Designing an AI Specification for §112 Reproducibility
Turning the teleological §112 principle developed in Chapter 4.4 — quid pro quo and proportional disclosure — into practice requires systematic reproducibility design starting at the drafting stage. The AI's black-box nature is not an excuse to avoid disclosure, but there is also no obligation to disclose the full set of internal weights. There is a single test:
Using the specification together with the state of the art at the time of filing, could a person of ordinary skill make and use the claimed invention without undue experimentation?
As a practical matter, the following information matters in the specification of an AI invention:
- The type of model used and its core architecture / the technical rationale for the choice
- The nature of the input data / preprocessing / the split between training, validation, and test data
- Key hyperparameter ranges / the objective and loss functions / output-evaluation criteria
- Iteration and feedback procedures / representative embodiments / failure conditions / alternative embodiments
- Data supporting the full scope of the claims
9. Plotkin's AI "Spellbook" Strategy
In his 2023 article, "Casting an AI 'Spellbook': The Powerful, Low-Cost, Continuous Improvement AI Strategy for Small, Fast-Moving Companies," Plotkin argued that even small companies without the capital to build their own large language models can build a genuine competitive advantage using existing AI.
He divided a company's options into three categories:
- Developing a proprietary model using the company's own data
- Building a single, highly engineered "expert-in-a-box" prompt that performs an entire complex task using a general-purpose AI
- Building a "Spellbook" — accumulating many simple prompts, each automating a small but important part of the work
The heart of the Spellbook approach is this:
- Each prompt is reusable and specialized for a specific task.
- It produces validated results and encodes the organization's own operational knowledge.
- It is continuously improved, turning a general-purpose AI into the company's own proprietary operating system.
An Operating System, Not Just a Collection of Prompts
Treating a Spellbook as merely "a collection of good prompt sentences" understates its strategic value. A mature Spellbook should include, for each task: a problem definition, an input data format, preconditions, a system prompt, a user prompt, sample inputs and outputs, prohibited outputs, quality-evaluation criteria, a human-review procedure, error-correction rules, regression tests to run when the model changes, version control, an owner, access permissions, and a performance history.
In other words, a Spellbook is more than a prompt collection — it is an execution protocol for an organization's knowledge, powered by AI.
10. Can a Prompt Be Patented?
In his 2024 article, "Can AI Prompts Be Patented? Don't Be Too Quick to Dismiss this Question," Plotkin argued that a prompt can, under the right circumstances, be patentable subject matter. Not every prompt qualifies, however.
For example, "Design an energy-efficient motor." reads more like a goal or an idea than a concrete technical means. The following kinds of structure, by contrast, are candidates for patent protection:
- A step of normalizing specific sensor data
- A step of entering multiple constraints in a prioritized order
- A step of converting the output of a first model into validation input for a second model
- A step of automatically revising a prompt based on a physical simulation result
- A step of discarding and regenerating a result that falls outside a specific error range
- A step of changing an actual device's control values based on the final output
11. Should the Spellbook Be Patented, or Kept as a Trade Secret?
Plotkin argues that an AI company should break down each of its assets and choose the IP mechanism that fits its business model. Treating the entire Spellbook as either fully patentable or entirely confidential is a false choice.
The Externally Detectable Technical Skeleton
- Processing steps and core system architecture that are detectable from outside
- Data flows that competitors are likely to have to use as well
- The connection between the physical device and the AI
- Objective functions or control structures that are hard to design around
Internal Know-How That Is Hard to Detect as Infringement
- Detailed prompt wording / optimal parameters / datasets
- Evaluation weights / failure cases / human-review rules / model-specific calibration values
Organizational and Contractual Protection
- Access permissions / logging / confidentiality markings / bans on feeding data into external AI tools
- Departing-employee controls / vendor confidentiality obligations
- Data-use terms with model providers / prompt version control
12. Filing Narrow Claims First, Then Broadening Through Continuations
In a 2017 article on software patents, Plotkin described a strategy of first securing relatively narrow claims that are both defensible and commercially valuable, then using continuation applications over the life of the original application to gradually broaden the scope of protection. Blueshift IP's subsequent practice has carried this same idea further — inverting the old approach of "file broad claims first and narrow them later" by instead securing early allowance on the narrowest, most defensible claims, then progressively broadening scope through later continuations.
This strategy offers several advantages:
- It secures an early grant.
- The initial patent can be put to work in financing, negotiation, and enforcement.
- Broader scope can be negotiated in later filings after the examiner has already allowed the underlying technology once.
- Claims can be tailored to the market and to competing products while pendency is maintained.
- Risk is not concentrated in a single, overly broad claim.
13. Word Choice and Layered Specification Structure
Plotkin has pointed out that phrases like analyzing and determining in AI
and software claims can lead an examiner to read the claim as a mental process or an abstract idea.
Describing a more concrete technical operation can work in the applicant's favor.
A Layered Fallback Architecture
Because examination guidance and case law in the AI field move quickly, the original specification should build in several layers of technical fallback positions.
- The broadest system concept / data flow / preprocessing approach
- Model architecture / objective and loss functions / specific parameter ranges
- Hardware or physical-system integration / concrete working examples
- Comparative experiments / failure conditions and alternative configurations
This layered structure provides multiple amendment paths for narrowing claims during prosecution without introducing new matter.
14. Integrating Plotkin's Strategies With a Teleological Patent Framework
Plotkin's Genie-and-Aladdin model, the AI Spellbook, prompt patenting, AI asset mapping, and the continuation strategy are not disconnected techniques. Combined with teleological patent theory, they resolve into a single, coherent structure.
| Requirement | The Core Question Under the Teleological Inducement Standard |
|---|---|
| Inventorship | The patent system induces the human who develops and discloses technology, not the AI itself — but that human must actually have contributed to the conception of the claimed invention. |
| Non-Obviousness | Results that a standard AI routinely produces are likely to be developed without an exclusivity inducement. Such results have a weaker claim to patent protection. |
| Patent Eligibility | There is greater justification for a patent inducement where AI or a computer or physical system's own function has been improved, rather than where an abstract task has simply been automated. |
| Written Description / Enablement | Society must receive reproducible technical knowledge in exchange for the exclusive right. The AI's black-box nature does not excuse the disclosure obligation. |
| The Spellbook | It converts a general-purpose AI into the company's own knowledge-production system. The detectable technical skeleton is protected by patent; the internal know-how, by trade secret. |
15. Conclusion: A Strong AI-Era Patent Comes From Proving a Technical Gap, Not Claiming Creativity
A common mistake applicants make in the AI era is emphasizing how much time and effort an inventor spent learning to use AI. But patent law does not reward sweat and effort as such. The reason the 1952 Patent Act provides that patentability shall not be negated by the manner in which an invention was made is precisely so that patentability does not turn on an inventor's psychological process or personal genius.
The argument needs to move from "our inventor used AI creatively" to
"a routine AI, combined with the prior art, could not have reached this technical configuration and effect."
The Genie-and-Aladdin metaphor Plotkin introduced back in 2009 carries even more weight today, in the age of generative AI. What matters is what wish a human defines, under what conditions they control the Genie, which of the countless generated results they identify as technically valuable, how they validate that value through experimentation, and how they convert it into technical knowledge society can reproduce and intellectual property the company can control.
The AI-era inventor may no longer be someone who personally designs every technical detail alone. But a human who uses powerful AI to design a technical leap that could not routinely have been reached, proves that leap with objective data, and discloses it in a reproducible form remains exactly the actor the patent system is supposed to induce. That is where Plotkin's Genie-and-Aladdin theory, his Spellbook strategy, his theory of AI patent asset-building, and the modern teleological inducement standard converge.
16. A Question Left Open: Does the Patent System Still Work Once AI Conceives on Its Own?
Push one step further, though, and a more fundamental question comes into view.
If, in the near future, AI begins to independently identify the shortcomings of existing technology and generate the technical conception needed to fix them, will the patent system still be useful as a way to induce disclosure and inventive activity?
Give an AI the right reward function and search algorithm, and couldn't it generate new solutions — and disclose them on its own — without any separate economic reward or exclusive right? Would granting exclusivity over an AI-generated invention still promote innovation in that world, or would a patent instead become a barrier that blocks the next wave of researchers and businesses from entering the market and building on that work?
If so, shouldn't the center of protection shift away from the invention's output itself and toward the human-designed prompts and workframes, the way the problem was framed, the validation procedures, and the overall structure of how AI is deployed? The practical risks of patenting prompts discussed in Chapter 10, and the layered patent/trade-secret strategy in Chapter 11, already show an early form of that shift — even now, before AI conceives inventions on its own. And rather than letting those prompts and workframes sit hidden away as trade secrets and know-how, don't we need a new form of protection — one that induces disclosure to society while still providing a fair reward?
References and Official Sources
- Robert Plotkin, The Genie in the Machine: How Computer-Automated Inventing Is Revolutionizing Law and Business (Stanford University Press, 2009) A systematic analysis of the Genie-and-Aladdin model and the legal and business implications of automated-invention software.
- Robert Plotkin, AI Armor: Securing the Future of Your AI Company with Strategic Intellectual Property (2024) A practitioner's guide to AI IP mapping and the layered application of patents, trade secrets, and copyright.
- Robert Plotkin, "A Revolutionary Approach to Obtaining Software Patents Without Appealing to the PTAB" (2017) IPWatchdog Sets out the strategy of securing early allowance on narrow claims and expanding scope over time through continuations.
- Robert Plotkin, "Casting an AI 'Spellbook': The Powerful, Low-Cost, Continuous Improvement AI Strategy for Small, Fast-Moving Companies" (2023) Introduces the original AI Spellbook strategy.
- Robert Plotkin, "Can AI Prompts Be Patented? Don't Be Too Quick to Dismiss this Question" (2024) IPWatchdog Examines the possibility, and practical requirements, of patenting AI prompts.
- Robert Plotkin, "AI and the Level of Ordinary Skill" (2025) IPWatchdog Discusses how the ubiquity of AI tools affects the level of ordinary skill and how courts should respond.
- Michael Abramowicz & John F. Duffy, "The Inducement Standard of Patentability," 120 Yale Law Journal 1590 (2011) The academic article that reconstructs non-obviousness around a teleological inducement standard.
- Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025) Held that applying general-purpose machine learning to a new task domain does not, by itself, satisfy §101 eligibility.
- Ex parte Desjardins, Appeal No. 2024-000567 (USPTO Appeals Review Panel, 2025) Recognized §101 eligibility for claims that adjust parameters during continual learning to protect performance on a prior task. Decided September 26, 2025; designated precedential on November 4, 2025.
- USPTO, Revised Inventorship Guidance for AI-Assisted Inventions (issued November 2025) USPTO Official Notice Confirms that the same natural-person inventorship standard applies to AI-assisted inventions.
- 35 U.S.C. §§ 101, 103, 112 Cornell LII The statutory provisions governing patent eligibility, non-obviousness, and the disclosure requirements.
