Four types of AI-assisted invention, and the "evidence of human conception" companies need to preserve
Generative AI has moved well past being a mere search, translation, or calculation tool. It analyzes technical problems, searches a vast design space, and proposes candidate mechanical structures, compounds, algorithms, and control schemes. Human researchers then screen the AI's output, run experiments, diagnose why something failed, and revise the conditions until the technology is complete.
We have entered an era in which a substantial share of invention forms not in a single, linear train of human thought, but through repeated back-and-forth between humans and AI. So who should count as the inventor in this kind of AI-assisted invention? Answering that correctly means starting not from "how creative was the AI" or "did the human contribute more than the AI," but from this question:
Which natural person conceived the specific technical concept of the invention as ultimately claimed?
1. The Evolution of USPTO Guidance and Its Core Legal Principles
1.1 The February 2024 Guidance: Extending Joint-Inventorship Doctrine to AI-Assisted Invention
In February 2024, the USPTO issued its first inventorship guidance for AI-assisted invention. The guidance started from the premise that using AI does not, by itself, foreclose patent protection — but at least one natural person still needs to have made a sufficient contribution to the claimed invention. To make that determination, the USPTO borrowed the three factors courts have long used to assess joint inventorship among multiple natural persons under Pannu v. Iolab Corp. Under the 2024 guidance, examiners were to assess, on a claim-by-claim basis, whether a natural person made what the guidance called a significant contribution to the claimed invention.
1.2 The November 2025 Revised Guidance: A Full Rescission of the 2024 Guidance
The USPTO announced revised guidance on November 26, 2025, and published it in the Federal Register on November 28, 2025, rescinding the February 2024 guidance in its entirety.
Where a single natural person uses AI, the Pannu factors are not applied to compare the human's contribution against the AI's, because there is no joint-inventorship question between a human and an entity that is legally incapable of being an inventor in the first place. Joint inventorship among humans only becomes an issue where multiple natural persons used AI together.
1.3 The Analytical Framework as of 2025–2026
| Scenario | Governing Doctrine | Core Question |
|---|---|---|
| One human and an AI | Traditional conception doctrine | Did the human conceive the specific solution recited in the claim? |
| Multiple humans and an AI | Conception doctrine plus the Pannu factors | Was each human's contribution qualitatively significant? |
| AI generates the specific solution entirely on its own | Possible absence of any natural-person inventor | Does a natural person who conceived the claimed invention even exist? |
| Human conceives, AI verifies/computes | Traditional conception doctrine | Was AI used as an implementation/verification tool? |
2. What Is Conception?
2.1 Conception Is the Touchstone of Inventorship
Under U.S. patent law, the touchstone of inventorship is conception. The Federal Circuit, in Burroughs Wellcome Co. v. Barr Laboratories, Inc., described conception as the mental act of forming, in the inventor's mind, a definite and permanent idea of a complete and operative invention. More specifically, a particular solution to the problem being solved must have formed in the human's mind. A general goal or a future research plan is not enough.
A General Research Goal vs. a Concrete Technical Solution
Examples that are unlikely to qualify as conception (general goals):
- "Let's build a more efficient motor" / "Let's design a juicer with less vibration"
- "Let's develop a more accurate AI model" / "Let's build a control system that saves energy"
Examples closer to conception (concrete technical solutions):
- Arranging two rotating shafts at a specific angle and distance to avoid a resonance region
- A training architecture combining specific data preprocessing with a specific penalty term
- A closed-loop relationship that changes an actuation condition once a sensor signal crosses a threshold
- A specific process range that produces an unexpected effect
The key question is not whether the inventor merely wanted a particular result, but whether they mentally settled on the concrete technical means for achieving it.
2.2 Conception Is Distinct From Reduction to Practice
Conception is the mental completion of the invention. Reduction to practice means actually building and testing the invention, or describing it in a patent application in a way that enables it to be practiced. Someone who builds a prototype from an already-completed design and confirms its performance may be an important research contributor. But if all they did was implement an already-completed technical concept in a routine way, that alone does not make them an inventor.
| Human Activity | Inventorship Assessment |
|---|---|
| Assembling a prototype exactly as the AI designed it | In principle, mere reduction to practice |
| Selecting a routine material | Likely a routine technical act |
| Merely confirming that the result works | Likely mere verification |
| Diagnosing the cause of failure and devising a new structure | Possible contribution to conception |
| Fixing new numerical ranges and relationships based on experiments | Possible contribution to conception |
| Combining features of multiple candidates in a new way | Possible contribution to conception |
3. The Standard for Joint Inventorship: The Pannu Factors
Pannu v. Iolab Corp. is the leading Federal Circuit decision on assessing joint inventorship among multiple natural persons. Since the 2025 revised guidance, these three factors are no longer used to compare a human against an AI. They are used to determine, among multiple natural persons, who made a qualitatively meaningful contribution to the claimed invention.
Contribute in Some Significant Manner
Each inventor must contribute in some significant manner to either the conception or the reduction to practice of the invention. Mere participation in reduction to practice is not enough — that contribution must rise above routine skill and connect to the formation of the claimed technical concept.
Not Insignificant in Quality When Measured Against the Full Invention
The contribution must be not insignificant in quality when measured against the dimension of the full invention. The standard is qualitative significance, not hours worked or volume of effort. Someone who proposed only a single element can be a joint inventor if that element is the key to solving the problem. Conversely, someone who participated in a long-running project but performed only routine testing, measurement, coding, or assembly may not be an inventor at all.
More Than Explaining Well-Known Concepts or the State of the Art
The contribution must go beyond simply explaining a well-known concept or the existing state of the art. The following, standing alone, generally will not make someone a joint inventor:
- Advising the use of a known material / introducing a textbook algorithm
- Explaining an industry standard / coding according to the inventor's specific instructions
- Running routine tests on a completed design / selecting a known manufacturing process
4. The Five Principles of the 2024 Guidance: How Should They Be Read Today?
The 2024 USPTO guidance set out five principles. Since the 2025 guidance rescinded that guidance in its entirety, those principles no longer stand as an independent legal test. They still have reference value, however, as an analytical checklist showing what facts to investigate when applying the ordinary conception doctrine to an AI-assisted scenario.
| Principle Under the 2024 Guidance | Its Current Legal Significance |
|---|---|
| Using AI does not negate inventorship | AI is a tool; whether a human conceived the invention is what matters |
| Raising a general problem is not enough | A particular solution is required |
| Mere reduction to practice or recognizing value is not enough | Conception and reduction to practice are distinguished |
| Contributing to designing or training the AI can count | It must connect to conception of the claimed invention |
| Mere ownership or oversight is not enough | Inventorship and ownership are distinct questions |
5. Four Types of AI-Assisted Invention
Entering a General Prompt and Claiming the AI's Output As-Is
A researcher enters a general-purpose generative AI prompt — "Design a compact transmission suitable for a radio-controlled car" — and carries the AI's design into the drawings and claims with little or no modification. Here, the human has done no more than pose a general problem or goal; it is hard to say they conceived the concrete technical means of the invention as finally claimed. Merely recognizing that the AI's output is useful, or choosing to file for it, is not enough.
Assessment: In principle, human inventorship is likely to be denied. The conclusion could differ, however, if the original prompt itself already contained concrete structural relationships, numerical conditions, or operating principles.
Building the AI's Design and Changing Only Routine Materials or Dimensions
The AI's design is actually built, and the exterior material is swapped from plastic to steel, or dimensions are adjusted within a routine range. A material substitution or dimensional change that a person of ordinary skill would make as a routine design choice is unlikely to be recognized as an independent act of conception.
Assessment: Inventorship is unlikely to be recognized in principle. But if the material change solved a previously unsolved problem by exploiting a relationship between a specific composite material's lay-up direction and its coefficient of thermal expansion, that could constitute a separate contribution to conception. What matters is not the magnitude of the change, but its qualitative significance to the technical core of the claimed invention.
Testing the AI's Output and Redesigning the Structure and Conditions
The AI's design is used as a starting point, but the human uncovers defects through testing and substantially redesigns the structure. Suppose the AI's proposed transmission was installed in an actual device and, at a certain rpm, produced gear-mesh defects and vibration — and the engineer then changed the axial length of the housing, relocated the shafts, added support elements, designed a clip-fastening structure, and adjusted the gear ratio. In that case, the person who finalized the completed technical concept can fairly be said to be human.
Assessment: Human inventorship is likely to be recognized. But if the human's contribution appears only in dependent claims while the independent claim tracks the AI's output as-is, an inventorship problem can arise on a claim-by-claim basis.
Improving the Design Through Iterative Prompting and Feedback
This is the hardest category — where the human never directly modifies the physical design, but incrementally improves the result by repeatedly adjusting the prompt.
- An initial prompt generates a general design.
- The human identifies a stress-concentration problem in the generated design.
- Load conditions, allowable stress, and manufacturing tolerances are added to the prompt.
- Exclusion conditions are set to rule out a specific shape.
- Simulation results are fed back into the AI.
- The objective function is revised to avoid a resonant frequency.
- Candidates are discarded or adopted for specific technical reasons.
- The final design is fixed based on the iterative results.
Here, the number of prompts does not matter. A single concrete technical constraint can be a more important contribution to conception than a hundred abstract instructions. Assessment: Inventorship may be recognized conditionally, but requires a claim-by-claim factual analysis.
6. Where Copyright and Patent Law Diverge on Iterative Prompting
The same act of prompting raises a different question depending on the body of law. Copyright law asks whether the human sufficiently determined the expressive elements of the final output. Patent law, by contrast, asks whether the human conceived the particular technical solution claimed in the invention.
6.1 The Beijing Internet Court's Generative-AI Image Case
On November 27, 2023, China's Beijing Internet Court recognized copyright authorship for a human user in connection with an image generated using Stable Diffusion. What the court emphasized was not the raw number of inputs, but the selection and ordering of prompts, the setting of an expressive approach, the composition and arrangement of the image, parameter adjustments, the selection among generated results, and the aesthetic judgment reflected in the final output.
6.2 The U.S. Copyright Office's Approach: Supplying a Prompt Alone Is Generally Not Enough
In Part 2 of its AI report, released January 29, 2025, the U.S. Copyright Office concluded that merely supplying a prompt is generally not sufficient. Copyright law's core question is this:
Did the human supply only the idea, mood, or style they wanted, or did they actually determine the concrete expressive elements of the final output — the specific lines, colors, sentences, notes, or arrangement?
6.3 The Critique of "Sweat of the Brow"
The U.S. Supreme Court, in Feist Publications, Inc. v. Rural Telephone Service Co., held that copyright cannot be granted merely because substantial effort and expense went into compiling data. The number of prompt iterations and the amount of trial and error does not, by itself, establish expressive originality. The real question is this:
Did the human's repeated inputs and selections substantially determine the final expression, or did the human supply only preferences and ideas while the AI actually determined the expression?
6.4 Patent Law Looks at Technical Conception, Not Expression
Suppose a human instructed an AI as follows:
"Arrange the first and second shafts asymmetrically, limit the shaft spacing to 18–22 mm, and adjust the gear ratio to avoid a resonance band of 7,500–8,200 rpm. Exclude any structure whose maximum deflection exceeds 0.3 mm."
Suppose the AI generated a concrete CAD shape in response. Because the human did not personally determine every line and curved surface in the final drawing, the authorship of that CAD image can be debated separately. But the asymmetric arrangement of the two shafts, the numerical range for shaft spacing, the resonance band to be avoided, the gear-ratio-based solution principle, and the maximum-deflection exclusion condition — all contained in the prompt — can connect directly to conception under patent law.
| Copyright Law | Patent Law | |
|---|---|---|
| What Is Protected | Human-created expression | The invention as defined by the claims |
| Core Question | Did the human determine the expressive elements? | Did the human conceive the technical solution? |
| A General Prompt | Closer to an idea or instruction | Closer to a general research goal |
| A Technically Constrained Prompt | May still fall short without expressive control | Can support conception if reflected in the claims |
| Number of Iterations | Not originality in itself | Not inventorship in itself |
| Unit of Analysis | The final expression and the human-authored portion | The claims and their technical limitations |
7. Prompt Length or Repetition Count Is Not the Standard
What matters under patent law is not whether a prompt is long or literarily polished. A long sentence can still be nothing more than a general goal, and a short instruction can still contain a concrete technical solution.
A General Goal
"Make it faster," "reduce vibration," "increase accuracy," "save energy"
→ Generally not recognized as conception.
Search Conditions or Evaluation Criteria
A maximum-stress limit, excluding a specific material, minimizing energy consumption, avoiding a specific resonant frequency
→ May contribute to conception, but requires examining how much of the specific solution ultimately claimed it actually determined.
A Concrete Technical Means
- A structural relationship between specific components / a specific numerical range
- A combination of data preprocessing and a loss function
- A closed-loop relationship between a sensor signal and an actuation condition
- A specific combination of substituents and reaction conditions
- A relationship governing validation, rejection, and regeneration of AI outputs
→ If a human established this and it made its way into the final claims, it becomes strong evidence of conception.
8. Flash of Genius and the Paradox of Iterative Prompting
8.1 Cuno Engineering and the "Flash of Creative Genius"
In 1941, the U.S. Supreme Court, in Cuno Engineering Corp. v. Automatic Devices Corp., held that a new device being useful was not enough — it needed to reveal, beyond ordinary skill, a "flash of creative genius." That standard drew criticism, and when the 1952 Patent Act codified the non-obviousness standard in §103, it included the following principle:
Patentability shall not be negated by the manner in which the invention was made.
Patent law therefore does not favor an invention reached through long experimentation and iterative prompting over one reached through a single, elegant prompt, or vice versa. Non-obviousness is judged objectively, against the prior art and the person of ordinary skill in the art.
8.2 If Iterative Effort Isn't Credited on Its Own, Does That Protect Only a Single Prompt?
That concern has some validity, but it needs a doctrinal correction. Not crediting repeated effort as such is not the same thing as favoring a single prompt. The law does not choose between "a single flash of insight" and "sustained effort." Whichever path was taken, it asks whether a clear and lasting technical solution to the claimed invention actually formed in a human mind.
8.3 §103's "Manner of Invention" and Inventorship Must Be Kept Separate
Inventorship: It is still necessary to investigate which natural person, in the course of that iteration, established the concrete technical means recited in the claims — not how much effort was expended, but who recognized the technical problem, who formed the working hypothesis, who set the technical constraints, which results were discarded for which technical reasons, and whose judgment the final claim limitations trace back to.
9. An Integrated Test for Applying to Type 4
AI-assisted invention arising from iterative prompting can be assessed along four axes.
The Specificity of the Prompt Content
Did it state only a general goal, or did it include concrete structures, numbers, relationships, or process conditions?
The Human's Technical Diagnosis
Did the human technically identify a defect or the cause of a failure in the AI's output? There is a difference between simply disliking a result and diagnosing stress concentration, resonance, overfitting, toxicity, or degraded durability.
The Correspondence Between Human Input and the Claims
Did the condition the human added end up reflected in some limitation of the final claims? A sophisticated prompt that has nothing to do with the claimed invention is weak evidence of inventorship.
The Technical Character of the Selection
Was the reason for choosing the final candidate mere preference, or a technical judgment about operating principle, effect, manufacturability, or safety?
| Form of Iterative Prompting | Copyright Assessment | Patent Inventorship Assessment |
|---|---|---|
| Repeated adjustment of mood/style | Depends on the degree of expressive control | Low if unrelated to technical conception |
| Repeated requests for general performance improvement | Depends on the outcome | Likely just a general goal |
| Setting concrete structures/numbers | May fall short without expressive control | High if reflected in the claims |
| Revising conditions based on experimental results | May be unrelated to copyright | Strong evidence of conception |
| Selecting a candidate on aesthetic grounds | May be recognized depending on jurisdiction | Low absent a technical reason |
| Recognizing an unconventional effect and redesigning | Separate from copyright | High likelihood of recognized inventorship |
10. Inventorship and Non-Obviousness Are Different Questions
A common error in AI-assisted invention is conflating inventorship with non-obviousness. Even where a human is a legitimate inventor, the invention can still be obvious. Conversely, even a technically highly non-obvious AI-generated result can raise an inventorship problem if no natural person actually conceived it.
Once AI becomes a routine research tool in a given field, the capabilities of the PHOSITA (person having ordinary skill in the art) can change too. Asserting an AI-augmented PHOSITA requires examining, at minimum, the following:
- Did the relevant AI tool exist as of the filing date, and could a person of ordinary skill access it?
- Was it actually in routine use in that field?
- Did the prior art supply the necessary inputs and constraints?
- Could a routine input have reached the claimed invention?
- Was the AI's output technically reliable, or was there a reasonable expectation of success without experimentation?
11. Is the Person Who Discovered and Selected an AI-Generated Candidate the Inventor?
AI can generate thousands, or millions, of candidates. Where one of them showed unexpected performance and a human discovered it and filed for patent protection, the basic rule is that mere after-the-fact recognition is not enough.
Mere Hindsight Recognition
- Adopting the AI's top-ranked recommendation as-is
- Choosing a visually appealing shape / an arbitrary selection
- Merely confirming that the result is useful
Genuine Technical Reconstruction
- A human discovers a physical principle the AI failed to capture
- Recognizing an unconventional effect in a candidate that would ordinarily have been discarded
- Fixing a new use and specific operating conditions
- Combining features of multiple candidates in a new way
- Deriving an independent structure or process from the AI's result
12. Software and AI Inventions Also Raise §101 Issues
Even where a legitimate human inventor exists, a §101 problem arises if the claim is directed to nothing more than an abstract idea. In 2025, the Federal Circuit in Recentive Analytics, Inc. v. Fox Corp. denied patent eligibility to claims that applied general-purpose machine learning to broadcast and event scheduling. In Ex parte Desjardins, by contrast, a concrete computational and control mechanism that adjusted parameters during training on a new task — so as to protect the model's performance on a prior task — was recognized as 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 |
| Claiming Style | Result/function oriented | Computation/control relationship oriented |
| Technical Problem | Making a task more efficient | Knowledge loss during continual learning (catastrophic forgetting) |
| Technical Effect | Faster results | Reduced storage/complexity and knowledge preservation |
| §101 | Ineligible | Eligible |
Specifications and claims should not stop at saying that "the AI analyzes," "optimizes," or "determines." They need to show which data structures, parameters, loss functions, iteration conditions, or signal relationships actually change how a computer or physical system operates.
13. Practical Recommendation for Companies #1: An Inventorship Evidence Ledger
In AI-assisted R&D, preserving only the final result makes it hard to reconstruct, after the fact, the relationship between a human's conception and the AI's output. Companies therefore need to maintain a structured Inventorship Evidence Ledger.
Materials to Accumulate in Chronological Order
- The original problem-definition document
- The solution hypothesis the human proposed
- The AI model and version used
- The system prompt and the user prompt
- Input data and constraints
- The AI's raw output
- Which candidates were adopted or discarded, and the reason for each
- Technical problems the human discovered
- Technical elements the human added or changed
- Test and simulation results / failure conditions and their causes
- Meeting notes attributed to individual inventors
- A mapping between the final claims and each human's contribution
- The allocation between patent and trade-secret protection
A prompt log alone is not enough. The record should also capture why a prompt was changed, which technical defect the human identified, and which claim limitation that judgment fed into.
14. Practical Recommendation for Companies #2: A Claim-to-Human-Contribution Matrix
Inventorship should not be assessed abstractly against the invention as a whole — it should be reviewed claim by claim.
| Claim Feature | Who First Proposed It | Subsequent Human Contribution | Supporting Evidence | Candidate Inventor |
|---|---|---|---|---|
| Basic gear arrangement | AI output | None | Output log | Unclear |
| Asymmetric support structure | Researcher A | Designed after stress analysis | CAD / meeting notes | A |
| A specific gear-ratio range | Researcher B | Fixed through resonance-avoidance testing | Test data | B |
| Clip-fastening structure | Researchers A & C | Jointly designed | Sketches / meeting notes | A, C |
| Control algorithm | Researcher D | Designed the loss function and stopping condition | Code history | D |
15. Practical Recommendation for Companies #3: A Technical Reproducibility Data Package
A strong AI patent comes not from the claim that "our researchers used AI creatively," but from objectively demonstrating the technical gap between a routine AI output and the final claimed invention.
The Technical Problem and a Baseline
Define the shortcomings of the existing technology using measurable data, along these lines (the figures below are illustrative drafting examples — actual numbers must be measured case by case):
- The existing structure shows a sharp rise in vibration above a certain rotational speed
- A conventional AI design exceeds the allowable stress by a certain margin
- The existing model's accuracy on a prior task drops by a certain amount after continual learning
- The conventional classification model reaches a certain false-positive rate on a particular data subgroup
AI Input and Control Structure
- The model and version / input data format / variable ranges
- Required and excluded conditions / the sequence of step-by-step queries
- Output-selection criteria / regeneration conditions / points of human review / experimental and simulation feedback
Human Technical Intervention
- Data preprocessing / feature extraction / changes to the objective or loss function
- Setting physical constraints / modifying shape and layout
- Designing the sensor-actuator relationship / diagnosing failure causes / combining and reconstructing candidates
Control Group and Experimental Group
Control group: a general-purpose AI, default settings, a routine prompt, results with no further human modification
Experimental group: unconventional conditions, proprietary data handling, experimental feedback, the human's final technical modifications
Ablation Studies and Failure Data
- Results with the penalty term removed / results with a specific preprocessing step omitted
- Results without human feedback / results outside the constraint range
- Results using a different model / results with one prompt stage removed
This data can reinforce not just inventorship, but also non-obviousness and the nexus between the technical effect and the §112 enablement requirement.
16. Specification and Claim Drafting Strategy
16.1 Write the Mechanism, Not Just the Result
The following phrasing, alone, is not enough:
- "Analyzes the data" / "determines the optimal result" / "improves accuracy"
Wherever possible, make it concrete:
- Computing a feature vector from sensor values
- Comparing the computed value against a stored threshold to generate a control signal
- Iteratively adjusting parameters to minimize a defined loss function
- Regenerating with revised constraints when the error falls outside the allowable range
- Changing the operating state of an actual actuator based on the final output
16.2 Draft the Original Specification Broadly and Deeply
- The broad system concept / data flow / data preprocessing
- Model architecture / the objective and loss functions / parameter ranges
- Integration with a physical system / concrete working examples / comparative experiments
- Failure conditions / alternative configurations / points of human intervention
This layered structure is what allows amendment without adding new matter during prosecution, and lets scope be adjusted through continuations.
17. A Corporate Checklist for Reporting AI-Assisted Inventions
Identifying the Original Technical Problem
- Who identified the original technical problem?
- Was it a general goal, or an unconventional problem formulation?
- What was the failure mode of the existing technology?
Documenting How AI Was Used
- Which model and version were used?
- What data and conditions were entered?
- How does the initial output differ from the final invention?
- Were prompt changes wording tweaks, or changes to technical conditions?
Attributing Conception
- Who set the key constraints?
- Who discovered the defect in the AI's output?
- Who modified the structure, values, or algorithm?
- Who fixed the solution principle through experimental results?
- Whose contribution does each claim feature trace back to?
Confirming the Scope of Each Participant's Contribution
- Which claims reflect which participant's contribution?
- Who performed only implementation, testing, or management?
- Was there a contribution beyond merely explaining well-known technology?
- Does an amended claim require a change in named inventors?
Confirming Patentability and Specification Content
- Could the same result have been reached with routine AI use?
- Is there an unexpected effect / a comparison and objective data?
- Can a person of ordinary skill reproduce the invention?
- Are there working examples supporting the full scope of the claims?
Deciding How to Protect the IP
- Which parts could a third party detect as infringement from outside?
- Which know-how should stay confidential?
- Should the claim be drafted around the system process rather than the prompt wording?
- Are access permissions and version control in place for prompts and AI workflow protocols?
18. Conclusion: The AI-Era Inventor Is the Human Who Created the Technical Gap
Inventorship in AI-assisted invention is not a question of comparing who was more creative, the AI or the human. Under current U.S. law, AI cannot be an inventor. But the mere fact that AI was used neither negates nor establishes a human's inventorship. What matters is which natural person contributed to the conception of the specific technical concept ultimately claimed.
Someone who enters a general goal and simply adopts the AI's output is unlikely to be recognized as an inventor. The same is true of someone who merely builds the AI's output or swaps in routine materials. By contrast, someone who tests the AI's proposed candidates, diagnoses their defects, and substantially redesigns the structure, conditions, or control principle is likely to qualify as an inventor.
Copyright law asks whether the human controlled the final expression; patent law asks whether the human conceived a concrete technical solution. So even if a prompt amounts to nothing more than supplying an idea under copyright law, a contribution to conception can still be recognized under patent law if it contained the specific structure, values, and relationships of the final claimed invention. Conversely, even hundreds of rounds of prompting will not establish inventorship if all they did was repeat a general goal and preference.
What matters is not the claim that "our researchers used AI very creatively," but the demonstration that
"a routine AI output combined with the prior art could not have reached this technical configuration and effect, and a human researcher specifically designed and validated this gap."
19. 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?
The four types and the conception-attribution analysis in this article all rest on the premise that a human is using AI as a tool. 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 Inventorship Evidence Ledger in Chapter 13 and the patent/trade-secret layering strategy in Chapter 17, Check 6, 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
- USPTO, Inventorship Guidance for AI-Assisted Inventions, 89 Fed. Reg. 10043 (February 13, 2024) Federal Register The original guidance applying the Pannu factors to AI-assisted invention; fully rescinded in November 2025.
- USPTO, Revised Inventorship Guidance for AI-Assisted Inventions (announced November 26, 2025; published in the Federal Register November 28, 2025) Federal Register Rescinds the 2024 guidance in its entirety and returns to the traditional conception doctrine (the natural-person inventor standard).
- Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998) The leading decision establishing the three-factor test for joint inventorship among multiple natural persons.
- Burroughs Wellcome Co. v. Barr Laboratories, Inc., 40 F.3d 1223 (Fed. Cir. 1994) Defines conception as the touchstone of inventorship and distinguishes conception from reduction to practice.
- Cuno Engineering Corp. v. Automatic Devices Corp., 314 U.S. 84 (1941) The source of the "flash of creative genius" language; the 1952 Patent Act's §103 codified the principle that patentability does not turn on the manner of invention.
- Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991) Confirms that the standard for copyright is independent creation and minimal creativity, not labor; rejects the "sweat of the brow" theory.
- Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025) Held that applying general-purpose ML 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 a continual-learning claim that adjusts parameters to protect performance on a prior task while learning a new one. Decided September 26, 2025; designated precedential on November 4, 2025.
- U.S. Copyright Office, Copyright and Artificial Intelligence — Part 2: Copyrightability (January 29, 2025) Copyright Office AI Policy Page An official report addressing the copyrightability of generative-AI output and why supplying a prompt alone generally falls short.
- Beijing Internet Court, Li v. Liu — Stable Diffusion Image Copyright Case, (2023) Jing 0491 Min Chu No. 11279 (decided November 27, 2023) A Chinese decision recognizing copyrightability in an AI-generated image where aesthetic judgment was reflected in the selection, ordering, and parameter adjustment of prompts and in the composition of the image.
