How Will AI Affect Open Source Development? (linuxfoundation.org) 39
This week the Linux Foundation's Open Source Security Foundation (OpenSSF) announced it would broaden its AI security resources, including a new guide to finding and fixing vulnerabilities using AI, "from preparing and sandboxing AI to validating, fixing, and preventing issues." The guide "stresses that AI-generated patches should never be trusted without automated tests and human review. An online course (LFD126) is planned for the coming months."
OpenSSF also released the State of AI Impact on Software Security and Cybersecurity 2026 report (supported by OpenSSF, ActiveState and the Linux Foundation AI & Data Foundation). And it turns out the 2026 report "found a growing gap between the pace of AI-driven software development and organizations' ability to secure the code being produced." 58% of surveyed teams ship code faster than security reviews can keep up, while 60% say AI coding assistants have increased the number of external packages in their codebases. This data underscores the importance of trusted, reliable guidance as AI deployment continues to grow.
Meanwhile the Linux Foundation announced at this week's Web Summit that it was releasing a new study on "how commercial open source companies generate revenue, convert users into customers and build their businesses," emphasizing the question's new relevance in the age of AI. "AI now writes code almost for free, and that raises a hard question for the whole software industry: if code costs next to nothing, where does the value of software lie?" Commercial open source [COSS] companies have been answering that question for more than 30 years. They have built large businesses on code they give away, which makes them a good guide to how software will make money in the age of AI... Seventy-two percent of respondents report a net positive impact from AI, against 8% negative. So far, the gains show up mostly inside companies. Internal productivity is the most cited benefit (73%), ahead of market expansion (46%) and new monetization opportunities (39%). AI roles are the only hiring category growing on net.
COSS strategy is mostly unchanged by AI. But when companies do adapt, they are nearly twice as likely to open more of their core product rather than become more closed. When asked what sets them apart, 57% of companies name brand recognition and trust, well ahead of enterprise features, integrations and community size. They turn that trust into revenue through support and services (63%), managed cloud (49%) and open core (43%). Most combine these: 65% run two or more revenue models.
Companies stack models as they grow, rather than switching between them: managed cloud rises from 35% of companies below $5 million ARR to 66% above $20 million. Venture-backed companies lean toward open core (53% vs. 38% for others).
The report draws on a survey of 225 commercial open source founders and enterprise leaders (published by Linux Foundation Research in partnership with Serena, and with sponsorship from AWS, Red Hat, Futurewei and the Cloud Native Computing Foundation. And when it comes to open source, "over half of founders point to brand recognition and community trust as their greatest asset," said Chris Aniszczyk, CTO of the Cloud Native Computing Foundation. So "The findings from this year's report prove that AI isn't a threat to commercial open source software — it's an accelerating tailwind. With 72% of companies viewing AI as a net positive primarily driving internal productivity, the ecosystem is leaning further into collaboration."
The Linux Foundation's announcement includes an interesting perspective from Matthieu Lavergne, partner at the venture capital company Serena. "Open source won the developer. Next, it will win the agent. The 225 founders we surveyed mostly see AI as a productivity gain today, but that's only the start: agents are set to become open source's largest users, and they choose what is open, proven and trusted. The founders who build and price for these new users will define the infrastructure of the AI era."
OpenSSF also released the State of AI Impact on Software Security and Cybersecurity 2026 report (supported by OpenSSF, ActiveState and the Linux Foundation AI & Data Foundation). And it turns out the 2026 report "found a growing gap between the pace of AI-driven software development and organizations' ability to secure the code being produced." 58% of surveyed teams ship code faster than security reviews can keep up, while 60% say AI coding assistants have increased the number of external packages in their codebases. This data underscores the importance of trusted, reliable guidance as AI deployment continues to grow.
Meanwhile the Linux Foundation announced at this week's Web Summit that it was releasing a new study on "how commercial open source companies generate revenue, convert users into customers and build their businesses," emphasizing the question's new relevance in the age of AI. "AI now writes code almost for free, and that raises a hard question for the whole software industry: if code costs next to nothing, where does the value of software lie?" Commercial open source [COSS] companies have been answering that question for more than 30 years. They have built large businesses on code they give away, which makes them a good guide to how software will make money in the age of AI... Seventy-two percent of respondents report a net positive impact from AI, against 8% negative. So far, the gains show up mostly inside companies. Internal productivity is the most cited benefit (73%), ahead of market expansion (46%) and new monetization opportunities (39%). AI roles are the only hiring category growing on net.
COSS strategy is mostly unchanged by AI. But when companies do adapt, they are nearly twice as likely to open more of their core product rather than become more closed. When asked what sets them apart, 57% of companies name brand recognition and trust, well ahead of enterprise features, integrations and community size. They turn that trust into revenue through support and services (63%), managed cloud (49%) and open core (43%). Most combine these: 65% run two or more revenue models.
Companies stack models as they grow, rather than switching between them: managed cloud rises from 35% of companies below $5 million ARR to 66% above $20 million. Venture-backed companies lean toward open core (53% vs. 38% for others).
The report draws on a survey of 225 commercial open source founders and enterprise leaders (published by Linux Foundation Research in partnership with Serena, and with sponsorship from AWS, Red Hat, Futurewei and the Cloud Native Computing Foundation. And when it comes to open source, "over half of founders point to brand recognition and community trust as their greatest asset," said Chris Aniszczyk, CTO of the Cloud Native Computing Foundation. So "The findings from this year's report prove that AI isn't a threat to commercial open source software — it's an accelerating tailwind. With 72% of companies viewing AI as a net positive primarily driving internal productivity, the ecosystem is leaning further into collaboration."
The Linux Foundation's announcement includes an interesting perspective from Matthieu Lavergne, partner at the venture capital company Serena. "Open source won the developer. Next, it will win the agent. The 225 founders we surveyed mostly see AI as a productivity gain today, but that's only the start: agents are set to become open source's largest users, and they choose what is open, proven and trusted. The founders who build and price for these new users will define the infrastructure of the AI era."
Using AI bug discovery and AI patches is great... (Score:4, Interesting)
... but if you don't incorporate AI into your project manage system, you're going to get swamped by it all. AI needs to also handle the first-tier of bug reports, patch submissions, and feature requests - reviewing them, evaluating them within the project goals, filtering out or improving submissions with issues, seeking and taking into account user feedback on controversial issues, etc. The limited human hours that most OSS projects have need to be insulated from the deluge, leaving them to focus on the high-level decisions, with content that they receive to consider being pre-reviewed and pre-considered for suitability to the project.
Otherwise, I don't see how most OSS projects can deal with this in the age of AI. Unless there's going to be some big increase in funding from somewhere?
Re:Using AI bug discovery and AI patches is great. (Score:5, Insightful)
Unless there is a big increase in funding across the board, how do you expect "most OSS projects" will pay for the bottomless bucket of tokens necessary to have "AI" handle project management, bug reports, patch review, etc etc?
What I expect to actually happen here is that the folks that have been maintaining the non-corporate F/OSS stuff will just walk away altogether and find something more personally rewarding to do with their limited time than subsidise commercial exploitation of their work.
Re: (Score:2)
how do you expect "most OSS projects" will pay for the bottomless bucket of tokens necessary to have "AI" handle project management, bug reports, patch review, etc etc?
By getting one of the subsidized subscriptions at 20 or 100 US$ from Anthropic / OpenAI (or whoever).
That's not bottomless, but a manageable amount of money. But of course it's still money that must come from somewhere and I could definitely understand any OSS developer who doesn't want to pay that.
Re: (Score:2)
Deepseek 4.1 Flash, GLM 5.3 Flash, Qwen 3.8 Flash, Muse Ember, Claude 5.5 Haiku, GPT 5.6 Luna, MiMo V2.6 Flash (or even MiMo V2.6 Pro), etc.
You don't need to use, say, Fable or Astra to handle bug reports :P
Re: (Score:2)
Also, handling bug reports / patch review is far less token-intensive than coding as well. It's really hard to compete with coding when it comes to token consumption.
Re: (Score:2)
That doesn't change the bottom line that tokens cost money that is nearly entirely coming out of maintainers' pockets.
Re: (Score:2)
I don't think you understand how cheap these models are.
Combined with the low token usage compared to normal coding, it's a total non-issue. I'll pay for any major OSS project's token usage for project management myself if they ask, that's how little it is.
Re: (Score:2)
> I don't think you understand how cheap these models are.
It doesn't matter how cheap they are when their use is effectively unbounded. Slop purveyors open 1,000 bug reports or PRs? That's a whole lot of tokens that get burned through automatically evaluating/reviewing everything, resulting in a DoS for "legit" stuff until you insert more tokens.
And it the flow isn't automatic, what's the point of adopting it?'
> I'll pay for any major OSS project's token usage for project management myself if they a
Re: (Score:2)
These models typically cost on the order of 10-20 cents per MILLION tokens in (assuming cache misses!) and ~50 cents per MILLION tokens out. 1/100th the cost of Fable or Astra, for a task that's far less token-hungry than coding. 1000 bug reports or PRs is perfectly handleable. It simply is not an issue.
Secondly, if there's evidence that a bad actor is involved, a model is perfectly
Re: (Score:2)
Also, there's a number of literally free [openrouter.ai] models you can use. They're not as good and generally rate limited, but they're better than nothing.
Re: (Score:2)
Couldn't they use the free models? It's not really a problem of "giving away your data" when it's all in the open anyway, is it?
FWIW, deepseek is actually a pretty good coder. It's also not bad at review. And it's free to use. I'm not going to talk to it about my personal problems, but for a bit of hobby code it's great. If it's not to your liking, OpenCode has a bunch of other free models you can use instead.
Re: (Score:1, Insightful)
Simple... they don't cave in and use AI... thus, maintaining the FOSS moniker... soon as you incorporate some "Clod" code into it, it's not FOSS.
Oh, and the free usage for the AIs is going to go away once everyone is addicted enough to it... it already has all your information, so it can just setup a recurring subscription. Not a big deal.
You _could_ refuse all AI junk and just write and release , and figure out an AI-proof bug-reporting system.
Re: (Score:2)
Indeed. And on top of that, using LLM code may well compromise the FOSS license used as well. Not good.
My take is that FOSS projects at this time need to reject contributions written in full or major parts by LLMs. In a few years we will know more and can revisit that.
AI is more for QA, not for coding ... (Score:3)
Using AI bug discovery and AI patches is great...
For QA and testing, yes, for coding, no. We are increasingly learning that AI coding often does not really save time on serious projects. The time saving of code generation is lost by an experienced human having to carefully review all the code. AI patches will be like a surge of inexperienced junior devs submitting updates. And then there is the copyright problem, AI generated code is not copyrightable in various jurisdictions. Copyrights are extremely critical for FOSS.
but if you don't incorporate AI into your project manage system, you're going to get swamped by it all. AI needs to also handle the first-tier of bug reports, patch submissions, and feature requests - reviewing them, evaluating them within the project goals, filtering out or improving submissions with issues, seeking and taking into account user feedback on controversial issues, etc.
Patch submissions are the exception
Show your proof. (Score:2)
Demonstrate proof that the amount of patches will be an ever-increasing number.
Linus just said they seem past the hump
He seems like an FOSS lead.
Huh? (Score:2)
For the service the software provides???
I mean, you don't pay a guy for the sole act of swinging a hammer, you're paying him because he's building you a house to live in
Re:Huh? (Score:4, Insightful)
But open source is often practically oriented. The author's aren't after money or fame, it has close to zero marketing and many projects are "I built a nice tool for me, if you find it helpful you can use it as well". If the helper I put online is getting out of fashion because someone else built the same with more bells and whistles I don't have any disadvantage from it, because I didn't put it online for paying rent.
The same way I don't care too much if you can now clone it with a hand full of LLM prompts. Maybe today I would have built it with a hand full of LLM prompts myself. I may even decide not to put it in online as it is most likely that people will rather create it with AI themselves tailored to their needs than having to deal with how to apply my tool to their use cases.
So no matter if the creation is now free, the product is useful if you need the program. It is only no longer useful to make money of it. That's mostly a problem for closed source. And maybe finally open core dies, when people take the GPL core and say fuck you to the CLA and rather vibe code the missing parts the company keeps gated.
AI or die (Score:2)
Basically, open source projects that want to retain their relevance will alow and even encourage the use of AI, or they will die or be forked off.
Refusing to use AI will be akin to refusing to write in C and sticking to assembler.
Re: AI or die (Score:4, Interesting)
F/OSS doesn't need to "retain relevance".
It's about scratching an itch.
Does it still itch? Scratch ahead. Did it stop, because someone did better scratching with AI? Great, enjoy your Netflix dinner on the couch! :-)
Re: (Score:3)
What do you mean by use? For code generation, or code analysis? Or do you mean for other maintenance tasks? All three?
Using AI for code generation is problematic for many reasons, some of which are a fundamental threat. Using it for the other things is problematic for fewer reasons, but more defensible.
Re: (Score:2)
Indeed. But there are a ton of idiots with no insight into how things actually work but deep awe for LLMs. Obviously, these two factors are connected. These morons probably believe we can revolutionize the whole world of software overnight, because LLMs are so powerful. In actual reality, LLMs can (badly) do something like 20% of a software creation process and still nobody major has seen any actual revenue increase from LLM use. That tells us the whole thing is like all other AI hypes before: Only a small
Re: (Score:2)
That is obvious nonsense. There is no need to use AI in FOSS projects. Many are in a good state and move slowly. That can nicely be continued with the non-LLM development model. As to "forking off", yes, some people will do that. They will fail, because it takes a bit more than last-stage code production to keep a FOSS project going.
I guess we will see a lot of AI forks failing in the next years because too many people think what they do not understand is easy to do and LLMs are all-powerful.
I'm not sure you understand what open source means (Score:2, Offtopic)
I just don't see people doing open source work as much anymore. Doing anything publicly on the internet right now means you get overflowed with AI agents spamming every stage of the cycle.
Re:I'm not sure you understand what open source me (Score:4, Interesting)
and a lot of people, especially juniors, were contributing to open source because if you look for a job, recruiters want to "see your work". But obviously since when you work for a company it's always proprietary and NDA, none of your work is "available to see". The only way to do this was to contribute to open source. A green activity grid on Github made it somehow look good.
Nowadays, open source contributions mean absolutely nothing because AI is doing all the work. So you have two types: idiots who keep pushing AI commits to have "public activity" on Github, and people who realized that the "Github activity hack" doesn't mean anything anymore.
In the end, you get a net decrease in quality.
Re: (Score:2)
There if a fix for that, called "closed contributor groups". You have a fix? Here is a list of people that can get it in for you. If you spam them with low-quality slop, you will just be added to a filter. Do not do it. If you want to contribute, prove competence and experience and you can apply.
open prompt? (Score:2)
Based on that litle experience, I'd say it is good for open source code. I am not sure if the source code is that valueable anymore. The prompt may be more worth...
So the only potential positive outcome I see (Score:2)
So for example you could point in AI at the Microsoft office file formats and it could quickly figure them out at all the nasty little tricks Microsoft uses to break open Office file format compatibility. It makes the old, it's not done until lotus 1 2 3 won't run thing much harder for monopolies to do.
On the other hand I would expect very quickly the open source
In the short term, or? (Score:2)
Re: (Score:2)
You copied that nonsense right from the claims of the LLM pushers, did you? It is so much without insight, it is staggering. No, "AI" cannot do most steps of software developement, and it can in particular not do the most important ones.
Re: (Score:2)
Some people seem to have a hard time coming to terms with the fact that AI can compete with humans in abstract fields such as software development and maths.
If AI is of any use at all (Score:1)
Re: (Score:2)
Very unlikely. Software is 20% or so implementation. The rest is design, architecture, etc. and LLMs cannot do any of that competently. Think of it like this: You can have generative AI paint you a picture. But unless you understand visual art _and_ prompting, it will be slop. Same for software, where the requirements are a bit higher and a bit more abstract.
How does (Score:3)
Why the future tense? It already is affecting it. People used it to find thousands of bugs in Linux and to create a Photoshop clone which even though its still unfinished looks closer to photoshop than its nearest fully human-made alternatives. We're well beyond the philosophical question "How will it affect" and the interesting question for most users is "How do I start also using it".
Re:How does (Score:4, Insightful)
and to create a Photoshop clone which even though its still unfinished looks closer to photoshop than its nearest fully human-made alternatives.
You can put a body kit on a yugo to make it "look like" a Ferrari, but it sure as hell ain't going to drive like one.
Re: (Score:2)
And yet they are closer than the other open programs. I am still skeptical about how much will really work and very cautious about the claims to soon be at feature parity, but please compare their current state with how gimp looked a few month after project start. This proof of concept has more working features than many traditionally engineered programs had after years.
Re: (Score:2)
> And yet they are closer than the other open programs.
No, they merely *look closer*, and you keep forgetting that "looks like photoshop" was never the point.
Functionally this slop is a dumpster fire, and is utterly worthless for the "real work" that everyone supposedly needs Gen-u-ine Photoshop for.
> but please compare their current state with how gimp looked a few month after project start.
First, GIMP never attempted to be a clone of the Photoshop UI, and Secondly, that was what, 30 years ago, and b