When a model is "open" and when that word is just marketing — an overview.
Fundamentals 12 min Beginner April 13, 2026
Every time you use an AI chatbot, you trust a system whose inner workings you cannot see. You cannot check what data it was trained on, whether it has hidden biases, or what it does with your input.
What if you could look under the hood — not just at the results, but at the engine itself? That is the promise of open source. By the end of this article, you will understand three things: what source code is, how licenses determine what you can do with it, and why all of this becomes critical when the software is an AI system.
What You See — and What You Don't
Source Code — The Recipe Behind the Software
AnalogyDefinition
Open source is like a published recipe: you can read the ingredients, follow the steps, adjust the seasoning, and share your improved version. Closed source is like a factory-sealed ready meal: you can eat it, but you have no idea what is inside or how it was made.
Analogy:
Open source is like a published recipe: you can read the ingredients, follow the steps, adjust the seasoning, and share your improved version. Closed source is like a factory-sealed ready meal: you can eat it, but you have no idea what is inside or how it was made.
Definition:
Source code is the human-readable text written in a programming language (Python, C, Java) that defines what a program does. A compiler or interpreter translates this into machine code (a binary) that the processor can execute. Source code is readable, modifiable, and auditable — compiled binaries are opaque.
The recipe analogy has limits: software can be copied infinitely at zero cost (cooking requires ingredients). A "recipe" with millions of lines of code is only readable by specialists. Some licenses allow reading but restrict commercial use (source-available).
Linux — Open Source
Over 40 million lines of code, publicly visible on GitHub. Anyone can read the code, report bugs, or submit patches. Started in 1991 with ~10,000 lines.
Windows — Closed Source
Proprietary source code that only Microsoft employees can see. Users receive installation packages, not source files. Independent inspection is not possible.
Important: code on GitHub is not automatically open source. Without a LICENSE file, full copyright protection applies — no one may copy, modify, or distribute the code, even if it is publicly visible. Public visibility does not grant usage rights.
Common Source Code Misconceptions
"Open Source = free" — Not necessarily. Open source means access to source code and certain freedoms, but companies like Red Hat earn billions with open-source software through support and services.
"Closed Source = more professional" — Most web servers (Apache, Nginx), programming languages (Python), and cloud infrastructures are based on open source. Professionalism does not depend on the licensing model.
"Code on GitHub = Open Source" — Without a LICENSE file, all rights belong to the author. Visibility is not the same as usability.
The Rules of the Game — Software Licenses
License — The Legal Rules of Open Source
AnalogyDefinition
Licenses are like the rules of a community garden. A permissive garden (MIT) says: "Pick what you want, use it or sell it — just mention where it came from." A copyleft garden (GPL) says: "You may harvest and breed new plants, but your new varieties must also be available to everyone." A source-available garden lets you look and taste, but prohibits selling the harvest beyond certain limits.
Analogy:
Licenses are like the rules of a community garden. A permissive garden (MIT) says: "Pick what you want, use it or sell it — just mention where it came from." A copyleft garden (GPL) says: "You may harvest and breed new plants, but your new varieties must also be available to everyone." A source-available garden lets you look and taste, but prohibits selling the harvest beyond certain limits.
Definition:
A software license is a legal agreement that defines who may use, modify, and distribute the code — and under what conditions. Without an explicit license, software is fully protected by copyright. The Open Source Initiative (OSI) certifies which licenses qualify as truly "open source."
The garden analogy has limits: licenses are legally enforceable contracts with real consequences. Code copies perfectly while plants do not. Subtle legal differences like patent clauses have no garden equivalent.
Permissive (MIT, Apache 2.0)
Maximum freedom. Code visible, modifications allowed, commercial use allowed. Only obligation: maintain attribution. Changes do not need to be shared.
Copyleft (GPL)
Same freedoms, but with one condition: derivative works must also be under GPL. If you modify the code and distribute it, you must release your source code.
The third family is called source-available: the code is visible, but commercial use or redistribution is restricted. Important: source-available is not the same as open source — the OSI does not recognize such licenses as open source.
Case Study: LLaMA vs. Mistral
Meta called its LLaMA models "open source," but the Llama Community License restricts companies with over 700 million monthly active users and requires separate licensing. This violates OSI criteria, which do not allow discrimination against specific users.
In contrast: Mistral publishes models under Apache 2.0 — genuine open source with unrestricted commercial use. The LLaMA case shows why you must read the actual license, not just the marketing.
Common License Misconceptions
"GPL prohibits commercial use" — No. GPL allows commercial use but requires derivative works to also release their source code.
"No license = public domain" — No. Without a license, full copyright protection applies. No one may use the code.
"Open Weights = Open Source" — No. Model weights without open training data and an OSI-approved license are a different category. LLaMA is an example.
Interactive: Open Source, Free Software & Proprietary
Click on the regions of the Venn diagram to explore the overlaps between Open Source, Free Software, and Proprietary software. The intersections reveal hybrid models like dual licensing or open core.
👆Click on a region in the Venn diagram to see its meaning.
Click on a region
Click on one of the circles or intersections to find out which software belongs there and what the category means.
Why This Matters for AI — Transparency
A closed AI system is like a car with a sealed hood: you can drive it, but you cannot inspect the engine, verify the safety ratings independently, or repair it yourself. You must trust the manufacturer's claims entirely.
An open AI system lets you open the hood, check every bolt, and even build a better engine. However: most users cannot read source code even when available. And open-source projects sometimes lack resources for thorough review.
Security: open code allows community-driven vulnerability discovery. But it is not a guarantee — the Heartbleed bug hid for over 2 years in OpenSSL, even though the code was open.
Trust: independent audits can verify claims about safety and fairness. With closed systems, you must blindly trust the provider.
Innovation: open models let anyone build on existing work. The history of GNU/Linux proves that open-source collaboration produces systems that power the world.
C
Closed
GPT-4, Claude: API-only access. No source code, no training data. You can use the model but cannot inspect, run locally, or modify it.
W
Open Weights
LLaMA: Model weights available for download, but with a restrictive license (700M user limit). You can run the model locally but cannot freely redistribute or use it for certain enterprises.
O
Open Source
Mistral (Apache 2.0): Full freedom. Source code and weights available, no usage restrictions. You can use, modify, sell, and redistribute the model.
Deep Dive: The Open-Source AI Ecosystem
Hugging Face Central platform for AI models, datasets, and demos
Ollama Run language models locally on your computer
LM Studio Graphical interface for local language models
Stable Diffusion Open-source image generation on your own hardware
n8n Open-source automation with AI integration
Deep Dive: Linus's Law and Its Limits
"Given enough eyeballs, all bugs are shallow" — this is how Eric S. Raymond formulated the principle named after Linus Torvalds. The idea: the more people read the code, the more likely bugs are found.
Reality is more complicated. The Heartbleed bug hid for over 2 years in OpenSSL, one of the most widely used security libraries in the world — despite the code being open. The "many eyes" were apparently looking elsewhere.
In 2024, a sophisticated backdoor was discovered in the xz compression tool — a targeted supply chain attack where an attacker gained the maintainers' trust over years. The open code became the attack vector.
The takeaway: transparency is necessary but not sufficient. It must be accompanied by active, funded security review. Open source provides the ability to inspect — but someone must actually do it.
Key Takeaways
Source code is the human-readable recipe; a compiled binary is the sealed product. Open source gives you the recipe; closed source gives you only the product.
A license turns visible code into legally usable code. Without a LICENSE file, code on GitHub is all-rights-reserved by default.
Permissive licenses (MIT, Apache) let you do almost anything. Copyleft (GPL) requires sharing changes. Source-available lets you look but restricts what you can do.
"Open weights" (like LLaMA) are not the same as "open source" — without open training data and an OSI-approved license, key transparency is missing.
For AI systems, transparency is not a luxury — it enables independent security audits, bias detection, and reproducible research.
Quiz: Open Source vs. Closed Source
Question 1 / 4
Not completed
What does a LICENSE file in a GitHub repository do?
1. What does a LICENSE file in a GitHub repository do?
☐ A) It makes the code private
☐ B) It defines who can use, modify, and distribute the code and under what conditions
☐ C) It encrypts the source code
☐ D) It shows who wrote the code
2. You find useful code on GitHub but there is no LICENSE file. What are you allowed to do?
☐ A) Anything — public code is free to use
☐ B) Nothing — without a license, full copyright applies and you may not copy, modify, or distribute
☐ C) Only read the code
☐ D) You can use it if you name the author
3. A company wants to use MIT-licensed open-source code in a commercial product. What must they do?
☐ A) Pay a license fee
☐ B) Release their product as open source
☐ C) Keep the original copyright notice
☐ D) Get written permission from the author
4. Meta calls LLaMA "open source." Why is this technically not accurate according to OSI?
☐ A) Because the code is written in Python
☐ B) Because the license restricts large companies (700M+ users) — OSI forbids discrimination against specific users
☐ C) Because the model is too large
☐ D) Because Meta is a corporation
Answer Key: 1) B · 2) B · 3) C · 4) B
Checkpoint: Test Your Knowledge
Why is "the code is on GitHub" alone not enough for something to be open source?
What is the key difference between a permissive license (like MIT) and a copyleft license (like GPL)?
Name one advantage and one limitation of open-source AI models compared to closed models.