The following content is for those who are not using Patenty.ai.
If you prosecute AI patents, you know the Examiner’s favorite template.
Whether it’s a Non-Final Office Action from the USPTO or an Examination Report from the EPO, the logic is often identical:
"The cited reference teaches the problem domain. A second reference teaches a Neural Network (CNN/RNN/Transformer). It would have been obvious to a Person Having Ordinary Skill in the Art (PHOSITA) to apply the known AI model to the known problem to automate the process."
(Cited as: 35 U.S.C. § 103 or Lack of Inventive Step)
It is frustrating. Your client spent months collecting data and fine-tuning the model to handle edge cases, but the Examiner dismisses it as a "predictable application of known tools" (citing KSR v. Teleflex in the US) or a "mathematical method with no further technical effect" (in Europe).
However, fighting on the "Novelty of the Model Architecture" is a losing battle. ResNet and BERT are prior art. To win, we must shift the battlefield from the Algorithm to the Data Pipeline.
Here is a strategic framework to overcome "Simple Application of AI" rejections by leveraging Data Preprocessing and Hyperparameter constraints.
When the Examiner argues that the Model is generic, agree with them. Then, pivot to the Input Data.
In both US and EP practice, transforming data to make it suitable for machine learning is often considered a technical feature that confers an inventive step.
Argue that applying a generic model to raw data would result in failure (overfitting or non-convergence). The invention lies in the specific preprocessing that enables the model to learn.
[Case Study: Medical Imaging]
The Rejection: "Applying a CNN to X-ray images to detect tumors is obvious."
The Pivot: We are not just claiming "CNN on X-rays." We are claiming:
"A method comprising: converting the X-ray image into the HSV color space; applying Contrast Limited Adaptive Histogram Equalization (CLAHE) to the Value (V) channel; and inputting the processed image into the CNN."
[Legal Argument]
US (§ 103): The specific combination of preprocessing steps yields "Unexpected Results" (drastic increase in accuracy in low-light conditions) that were not predictable by simply combining prior art references.
EP (Art. 56 - Comvik): The preprocessing is not a mere administrative or mathematical step; it is a technical process that improves the signal-to-noise ratio of the input signal, thereby contributing to the technical character of the invention.
Simply claiming "a learning rate of 0.001" will be rejected as an arbitrary design choice (result-effective variable). You must frame parameters as Structural Limitations or Critical Constraints.
Show that the specific parameter is not a routine choice but a solution to a specific technical hurdle.
Custom Loss Functions:
Instead of standard Cross-Entropy, claim a Weighted Loss Function designed to handle class imbalance (e.g., in fraud detection where 99.9% of data is normal). This is a technical solution to a data problem.
Training Strategy (Freezing & Fine-Tuning):
Explicitly claim the training regime: "Freezing the weights of the first N layers of the feature extractor and updating only the parameters of the classifier head."
This transforms the claim from a "Generic AI" to a "Resource-Optimized Computing Process," which is much harder to reject as "routine."
Examiners often assume that if Data A exists and Event B happens, feeding A into an AI to predict B is obvious.
Challenge this assumption. The discovery of the correlation itself can be the invention.
[Logic Construction]
Problem: PHOSITA would not expect that "Motor Vibration Data" (Input A) could predict "Machine Failure 3 days later" (Output B) due to high noise levels.
Solution: The Applicant transformed Input A into the Frequency Domain (FFT) and fed it into an LSTM.
Conclusion: The invention is not the use of LSTM. The invention is identifying that Frequency Domain Data contains the latent features necessary for prediction, establishing a non-obvious technical causal link.
To support these arguments, you need to prepare the "ammunition" during the drafting stage.
In US practice, Secondary Considerations (Objective Indicia of Non-Obviousness) are powerful. In EP practice, showing a Technical Effect is mandatory.
Include a graph in the specification comparing:
Baseline: Accuracy of Generic Model + Raw Data (e.g., 70%).
Invention: Accuracy of Generic Model + Claimed Preprocessing (e.g., 92%).
This creates an objective basis to argue "Unexpected Results" or "Technical Contribution."
Keep the independent claim broad, but ensure you have dependent claims specifically reciting:
Specific Preprocessing Steps (Normalization, Augmentation, FFT).
Loss Function Formulas.
Data Structure constraints.
Why? When the Examiner rejects the broad independent claim under § 101 or § 103, you can immediately pull up these dependent claims to negotiate an allowance.
When an Examiner says, "It’s just AI," our response should be:
"Two cars may have the same engine (Model), but the fuel (Data) and the tuning (Preprocessing) determine whether the car wins the race."
The "Inventive Step" in modern AI patents is rarely found in the neural network architecture itself. It is hidden in the data engineering and the optimization constraints.
Next time you face a "Routine Optimization" rejection, try shifting the focus to the Input Pipeline. You might find the Examiner has run out of ammunition.
[ ] Did you shift the argument from the "Model" to the "Data Preprocessing"?
[ ] Did you justify hyperparameters using "Criticality" (US) or "Purposive Selection" (EP)?
[ ] Are Custom Loss Functions or specific Training Strategies (e.g., freezing layers) included in dependent claims?
[ ] Do you have Comparative Data in the specification to prove "Unexpected Results" or "Technical Effect"?
An analysis of the operational shift from generative drafting to agentic workflow management. This report examines how AI-driven pacing and cognitive drag analysis are becoming critical infrastructure for risk mitigation in high-volume patent practices.
Prevent AI hallucinations and ensure full support (35 U.S.C. §112, EPO Art. 84) by generating the patent specification after finalizing claims. Discover 5 "Claims-First" prompt templates to lock the AI within the legal scope, covering system claims, method sequences, and means-plus-function limitations.
"What if I miss a killer X-reference simply because the applicant used 'Coupling Means' instead of 'Fastening Member'?" The nightmare of facing a rejection based on a document you missed—because of a single keyword—is real. Today, I’m sharing a method to drastically reduce this risk using Generative AI (ChatGPT, Claude, etc.). This isn’t just about "asking AI." It’s about strategic prompt engineering to filter out noise and extract the exact keywords needed for a robust search string.
