A published paper and the algorithm it describes are protected differently — and conflating the two is one of the most common student errors in this area. Understanding the three forms of intellectual property, the specific limit Section 3(k) of the Patents Act places on algorithm patents, and what the MIT, GPL, and Apache licences each require turns IP knowledge into something practically useful for a researcher.
Sign in to mark this article as read and track your progress.
| Right | What it protects | In India |
|---|---|---|
| Copyright | The specific expression of an idea — source code as written, a paper's exact text, a diagram, a dataset structure | Copyright Act, 1957. Arises automatically on creation; no application needed |
| Patent | A novel, non-obvious, industrially useful invention — the underlying technical solution, not its expression | Patents Act, 1970. Must be applied for and examined; does not arise automatically |
| Trade secret | Confidential information — a proprietary detection algorithm, a training dataset not publicly released — protected only for as long as it is kept secret | No registration; no fixed term |
The three forms protect different things. A tool's source code is copyrighted the moment it is written. The algorithm that source code implements may or may not be patentable. If neither is filed nor disclosed, the algorithm may be a trade secret — but only while it remains confidential.
A published research paper and the algorithm it describes are protected differently. The paper's text, figures, and equations are copyrighted: that is why the rules on quoting and paraphrasing apply to them. The algorithm itself is not covered by that copyright.
In India, Section 3(k) of the Patents Act, 1970 states that "a mathematical or business method or a computer programme per se or algorithms" cannot be patented. The phrase per se creates a narrow opening: a claim that demonstrates a concrete technical effect beyond the bare algorithm — for example, a novel hardware pipeline that implements the algorithm with a specific physical efficiency gain — may qualify. A pure algorithmic improvement described in a paper, by itself, does not.
Publishing a novel ML architecture does not grant IP protection over the architecture. It grants copyright over the specific paper text. Anyone can re-implement the architecture from the published description.
Ronald Rivest, Adi Shamir, and Leonard Adleman published the RSA public-key cryptosystem in 1977. MIT, their employer, was granted US Patent 4,405,829 in September 1983. Because the algorithm had already been publicly described before the application was filed, most other countries would not grant an equivalent patent — only the US patent applied. For seventeen years, anyone wanting to use RSA commercially within the United States needed a licence. The patent expired in September 2000, after which RSA became free for anyone to implement.
The case illustrates the patent bargain precisely: a public, legally binding disclosure of exactly how the invention works, in exchange for a time-limited monopoly, followed by the invention entering the public domain when that term ends. The disclosure is not optional — it is what the monopoly is exchanged for.
When a researcher builds a tool using open-source libraries, the licence of each library determines what the researcher must do when distributing the tool:
MIT licence. Permits almost any reuse, including in closed-source commercial products, provided the original copyright notice is kept. Minimal obligation.
GPL (GNU General Public Licence). Requires that any modified version distributed to others be released under the GPL too. If you incorporate a GPL-licensed library into your tool and distribute the tool, the tool's source code must become publicly available under GPL. A tool built on a GPL library cannot be incorporated into a closed-source commercial product without the commercial product's source code also becoming publicly available.
Apache 2.0. Permits broad reuse including in closed-source products, and explicitly grants users a patent licence from contributors — relevant if the contributor holds patents covering the code.
A researcher releasing a tool alongside a paper should choose the licence deliberately. Using a GPL-licensed component without understanding the obligation is not a defence; it exposes anyone who builds on the released tool to downstream licence violations.
A PhD student publishes a paper describing a novel graph-neural-network architecture for malware classification. Her university's technology transfer office asks whether the architecture is patentable in India. Apply Section 3(k) of the Patents Act, 1970 to give an answer. Then: she has released her code on GitHub under MIT, but built it on a logging library released under GPL v3. A company wants to incorporate her tool into a closed-source endpoint product. Which licence creates an obstacle, and what would the company need to do to resolve it?
