Most AI ideas are not patentable. A few are. Here is how to tell the difference — from someone who has sat on both sides of that line.
A few years ago, I found myself in a room with a group of like-minded SMEs, a whiteboard full of diagrams that only the people who drew them could understand, coffee levels dangerously low, and confidence levels dangerously high. We all looked at each other with a question we had never seriously considered before:
“Is what we built actually patentable?”
Surely the answer was yes. After all, it involved AI.
I had filed patents before — across various domains and technologies. The process was familiar. But AI, I was about to discover, is a different animal entirely. The rules that applied to industry innovations, IoT systems, cryptographic solutions, and conventional software did not translate cleanly to AI. What counted as novel, what counted as abstract, what the examiner would accept and what they would reject — all of it required relearning from scratch.
We had built something we were proud of. An AI system that did something genuinely useful, in a way nobody had done before. Naturally, we assumed a patent was waiting for us.
Then, I learned that brilliance and patentability are not the same thing. This article is what I wish someone had told me before that meeting.
Why Most Engineers Never Think About Patents
There is a cultural reason engineers do not think about patents. In the software world, the dominant wisdom for decades has been: ship fast, iterate, build moats through execution, not legal protection. Patents are for large corporations with legal departments. They are slow, expensive, and by the time your patent is granted the technology has moved on.
There is truth in this. But it is incomplete — and it is increasingly incomplete as AI moves from software into systems, algorithms, computations, integrations and applications.
The more your AI invention touches something fundamental — either at the core of how AI works, or at the edge where AI meets the real world — the more patents matter.
At the core, this means inventions that advance the foundations of AI itself: a novel algorithm to invert a neural network without compromising its weights, a more efficient method for token search computation, a training technique that changes how models generalise. These are contributions to the science of AI — and when they are genuinely novel and non-obvious, they are patentable.
At the edge, this means applied engineering — AI systems embedded in telecommunications, manufacturing lines, healthcare diagnostics, smart farming, industrial automation. Here the invention is not the model itself but what the model makes possible in a specific domain.
If you are building AI purely as a web application — a chatbot, a recommendation engine, a productivity tool — patents may genuinely not matter to you. But if your work touches the core foundations of how AI thinks, or the applied engineering of how AI operates in a specific field — read on.
What We Built
I will be deliberately general here — not because the specifics are secret, but because the specifics are not the point. What matters is the pattern.
We built systems that proved more efficient than existing ones, tweaked algorithms to use fewer computations, and applied system intelligence to areas where nobody had thought to apply it before. Each of these felt like a genuine breakthrough in the moment.
Naturally, we assumed a patent was waiting for us with open arms.
The question was: which parts of this system were patentable? The answer, it turned out, was more nuanced than either “yes, the whole thing” or “no, it is just software.”
The First Thing I Learned — Abstract Ideas Are Not Patentable
Abstract ideas — including mathematical concepts and mental processes — are not patentable on their own, even when implemented on a computer. An AI invention must do more than apply an abstract idea using generic computer functions. It must provide a specific technical improvement to the functioning of a computer or another technology.
This is where most AI patent applications fail at the first hurdle. The applicant describes a machine learning model — the architecture, the training procedure, the loss function — and the examiner correctly observes that this is an abstract mathematical method. Not patentable.
Do not claim the model. Claim what the model makes possible.
The Three Tests Every AI Invention Must Pass
Test 1: Is it novel?
Novelty means the invention has not been publicly disclosed anywhere before your filing date. Academic paper? Counts. Conference presentation? Counts. GitHub repository? Counts. Blog post? Also counts. The article you’re excited to publish today might accidentally destroy the patent you wanted tomorrow. File before you publish. Always.
Test 2: Is it non-obvious?
Even if an invention is novel, it must not be obvious to a person skilled in the relevant field. This is my favourite test because it is essentially asking: “Would a reasonably smart engineer have thought of this anyway?” If the answer is yes, you are probably not getting a patent.
Test 3: Does it have utility?
The invention must have a specific, substantial, and credible use. AI embedded in a manufacturing system, a medical device, or an agricultural monitoring system easily meets this test. Pure abstract AI — a recommendation algorithm, a text classifier with no specific application — can struggle.
Practical Advice for AI Builders
Keep a dated record of everything — notebooks, commit logs, email threads, even ugly papers with coffee stains. Future you will thank present you.
File before you publish — the excitement of sharing your invention lasts a few days. The regret of accidentally killing your patent lasts much longer.
Think about the system, not just the model — start from the output (what does your AI make happen in the world?) and work backward.
Do the prior art search yourself first — Google Patents, Espacenet, USPTO database. If you find exact prior art, better to know before you invest further.
What the Patenting Process Taught Me
The biggest lesson wasn’t about patents. It was about thinking. When you are writing patent claims, you are forced to articulate — with precision that a trained examiner will scrutinise — exactly what is novel about your invention.
I now think about novelty differently when I design AI systems. Not primarily in terms of “is this patentable” — but in terms of “what is genuinely new here, and can I articulate it clearly?” That clarity of thinking makes better engineers. It makes better systems. And occasionally, it makes something worth protecting.
The engineer who understands both the technology and the intellectual property landscape around it is increasingly rare — and increasingly valuable. AI is moving faster than our brains can keep up. The frameworks are still being written, the precedents still being set, the boundaries still being tested. That ambiguity is not a reason to ignore patents. It is precisely the reason to pay attention now — before the rules harden and the window closes.
The model was never the invention. The thinking behind it was. Protect it.
This article focuses specifically on the nuances of patenting AI inventions — a domain that sits at an uncomfortable intersection of mathematics, software, and physical application. I have a separate story to tell about my broader journey through the patent process — the filings, the rejections, the negotiations, the lessons learned across domains far beyond AI. That article is coming.