This is a series of four articles on AI and IP protection:
I. Setting the stage
Working alongside early-stage companies, venture funds, and corporations, I often wonder what the word “defensibility” means nowadays for an AI startup. And even if I find myself often thinking and advising people on why a patent on a machine learning product/algorithm makes sense or not, I recently realized that this problem is actually related to how I see an AI company be protected in the long-run.
I will blog later about the broader concept of defensibility and rather focus here only on IP protection. I am not a patent lawyer, so there might likely be many details I am either missing or misunderstanding (and you should definitely check with one if you are thinking about patenting something) but from a business perspective, I think I have now a pretty good picture on what is going on.
So let’s start with the fundamentals.
II. Why startups patent inventions (and why is different for AI)
Patenting an invention is one of the four forms of IP protection (i.e., patents, trademarks, copyrights, and trade secrets) and it basically gives you the chance to use and exploit the economic benefits of a certain invention for a quite long period (usually 20 years) demanding in turn that you make your invention public for the sake of the scientific and technological progress.
For a startup, having a patent has historically been a great advantage over the competition, but is also a burdensome cost to cover especially at a very early stage (around 20k both in Europe and US over a 3 to 5 years period). And this cost is not even getting into account the fact that obtaining a patent is the same of tossing a coin (only 55% of the applications result into a granted patent, eventually — see Carley et al., 2014 for more details).
So why companies get themselves into this tricky and expensive process in the first place? Well, there are multiple reasons for doing it. You can indeed create a strong competitive advantage or shape a new stream of revenues by licensing your technology, or simply increase your confidence that what you are building is being recognized by the rest of the world as useful.
Indeed one of the companies I am working with called Meeshkan ML recently applied for a patent for their distributed machine learning algorithm because in the founder’s words “we feel that we can share it with our local community of engineers and build new algorithms on top of it without worrying about losing our business advantage. We therefore made a calculated risk in investing time into the R&D necessary for the patent instead of pushing a product out.”
But most of all, a startup often heavily depends on external financing — and investors love(d) patents. It is a simple way for many of them to achieve three things at once: they get sure the tech is legit (i.e., they outsource the technical due diligence to patent lawyers and offices); they are more confident the technology is feasible (reducing product risks); they get more confident the team can actually build what they claimed and that is committed to it (reducing execution and team risks).
This is especially true for deep science ventures as well as emerging technologies where understanding those three aspects is incredibly cumbersome. AI is clearly one of those fields, but I am also claiming it stands apart from other technological inventions for a bunch of different reasons:
- A strong open-source community: many AI algorithms/libraries/packages are completely open-sourced, and you usually tend to build on top of those (which makes patenting extremely hard to be managed);
- Confusion around the (real) invention: it is still hard to identify which part of an AI algorithm is the real source of invention. Is it the source code? Or rather the data used? Or maybe the process or the human-machine blending interaction?
- Continuous evolution: feedback loops in machine learning push the code to keep evolving, which complicates the understanding of whether a new invention breaks the initial patent.
- It stems from an academic community: traditionally many inventions are industry-driven while AI is historically intrinsically related to academia, where the approach to scientific research and development is culturally very different;
- It is still half science half art: even though a patent requires a fairly decent degree of details around the invention in question, when it comes to AI the devil is into the smallest details (e.g., tuning). In other words, even with a full access to the process or to the algorithm, you might not be able to use it and implement it correctly.
I am always interested in speaking to, learning from or simply connecting with interesting founders working in highly impactful fields like life sciences, energy, and others. If you are one of them, feel free to reach out here!
Carley, M., Hegde, D., Marco, A. (2014). “What is the Probability of Receiving a U.S. Patent?”. Yale Journal of Law and Technology 17 (1): 204–224.