Bösherz Goebel Patent Attorneys
Protecting software and AI inventions
"Software cannot be patented" is a common answer, and it is still wrong. What cannot be patented is source code and software with no technical link. For everything else, what decides is what the software does for the customer and in the product.
What the sentence actually means
Patent law excludes "programs for computers" from patentability — but only in so far as protection is sought for them as such. Legal practice rests on that qualifier. It means an invention that uses software to solve a technical problem is not excluded. What is excluded is the program in itself, detached from any application.
Art. 52(2)(c) and (3) EPC · Sec. 1(3) no. 3 and (4) German Patents Act
The rule it comes down to: software and AI are patentable where they achieve a specific technical effect in a concrete application. For technical applications that is almost always the case. Anything that measures, controls, transmits or processes is on the right side of the line; anything that maps a commercial process is not.
Enlarged Board of Appeal of the European Patent Office, G 1/19 · Guidelines for Examination G-II 3.3, G-VII 5.4
Comprehensive protection for developments in software and AI matters more today than it ever has. It often fails because of misconceptions about what can be patented.
What the misconception costs
Anyone who believes software cannot be patented does not file. The real risk is then not the protection they miss out on, but other people's: a competitor who has filed on the same solution can later stop what has long been in routine use in your own house. Companies that never check for other people's rights may find out only through a warning letter or an enquiry asking on what basis the product is being used.
How to recognise it in your own product
Three examples from fields where the assessment is clear-cut:
- Error correction in a data transmission. The algorithm improves the transmission itself, a technical process. That it consists of mathematics changes nothing.
- Evaluating sensor data in a vehicle. Camera, radar, LIDAR and ultrasound signals become a statement about the surroundings. Input and output are both physical.
- Classifying tumours in image data. What looks like pure mathematics as a random forest is a technical contribution as medical image analysis. The model stays the same; the assessment changes with the application.
Open the example lists: what lands on which side in practice
Four lists from European Patent Office examination practice. They do not replace an assessment of your own case, but they show the pattern: what decides is not the technology inside, but the purpose outside.
Software
Patentable
- Controlling technical devices (X-ray, ABS)
- Data compression and encryption
- Error-correction coding
- Processor load balancing
- Efficient memory allocation
- Optimising compilers
- Security measures for boot integrity
- Protection against hardware attacks
- Algorithms for hardware optimisation
- Image restoration and enhancement
- Signal processing with a technical purpose
- Turning measurements into diagnoses
Not patentable
- Abstract algorithms with no technical link
- Business methods
- Accounting programs
- Word-processing programs
- General databases
- Merely general computer functions
- Algorithms without a technical effect
- Software with an overly general purpose
- Presentation of information
- Pure data processing without technical output
- Mathematical formulae and methods
- Software unrelated to physical quantities
Artificial intelligence
Patentable
- Heart monitoring for irregular heartbeats
- Image analysis for identifying people
- Speech recognition with technical output
- Determining material density
- Automated medical diagnosis
- Genotype estimation from DNA samples
- Detecting faults in industrial plant
- Optimising energy consumption
- Predicting raw-material availability
- Analysing low-level image features
- Controlling chemical processes
- Hardware-optimised computation
Not patentable
- Neural networks without a technical application
- Classifying text documents
- Abstract data classification
- Machine learning for purely statistical analysis
- Customer classification
- Stock-market analysis
- Learning algorithms without technical output
- Advertising targeting
- Support vector machines without a technical purpose
- Models with a purely linguistic purpose
- General business processes
- Algorithms for abstract data analysis
Illustration, without warranty of accuracy or completeness. The assessment depends on the individual claim.
Where the parts of a product sit
A patent is never granted on "the software" but on a particular part of a solution. So it pays to look at which part sits close to a practical application and which is closer to pure thinking.
Whether filing is worth it
Whether something is patentable and whether filing is worth it are two different questions. Plenty would be patentable without the filing paying for itself. A sound IP strategy therefore also includes a deliberate decision, in some cases, not to file. Three questions give a clear answer in most cases, and all three are asked of the product, not of the code.
What protects when a patent does not fit
A patent requires disclosure and pays off above all where an infringement can later be proved. A trade secret requires the opposite: it works only as long as it stays secret. Which route fits therefore depends on how much an outsider can make out of a given part.
Why disclosure is the harder part with AI
Whether an AI invention is patentable is often measured less by technical character as such than by whether the application discloses enough. Patent law requires that a skilled person can reproduce the invention from the description. For a model whose behaviour follows from data and training that is more demanding than for a machine — and it collides with the legitimate wish to keep training data and methods in-house. That trade-off belongs at the start of a project, not at the end.
Art. 83 EPC · T 161/18 · T 1669/21 · on plausibility G 2/21
Where applications fail in the drafting
The language of engineering has to be translated into the language of patent law. That is where the mistakes arise which are most expensive to correct later:
- Functional terms drawn too broadly. A claim that only says what is achieved, not how, covers a great deal on paper. It holds only as far as the description supports it. Without that support, less survives than a narrower claim would have delivered.
- No support in the description of the invention. What stands in the claim has to be found again in the description. Often nothing can be added afterwards: what was not disclosed on the filing date is then lost for that patent application.
- Different standards at different offices. The European Patent Office, the German office and offices outside Europe weigh structural and functional features differently. Anyone filing internationally takes that into account in the first draft, not at the subsequent filing.
Art. 84 and Art. 123(2) EPC
Can an AI be an inventor?
No. The inventor named has to be a natural person; an AI system cannot take that role. The question went through several instances under the name DABUS and has been answered consistently in Europe.
In practice this changes little: anyone using AI in development still files as usual and names the people who made the invention. A tool does not become a co-inventor.
European Patent Office, Legal Board of Appeal, J 8/20 and J 9/20
What this means for working together
Filing a software or AI invention needs an early conversation between the people who built it, the people who own the product, and the patent attorney. Improving the claims after filing is often no longer possible.