Dries Buytaert Calls for Shared AI Infrastructure in Open Source

Access to artificial intelligence tools, shared AI skills, and review practices should become part of open-source contribution infrastructure, Drupal founder and project lead Dries Buytaert argues in a new essay. The essay frames AI as a possible way to reduce the privilege of free time in open-source contribution, provided communities make the tools and knowledge available beyond those who can pay for or master them alone.

In The privilege of AI in Open Source, published 30 June 2026, Dries revisits his 2019 argument that open source is not a pure meritocracy because contribution depends on time, income, and schedule flexibility as well as talent. He argues that AI can help contributors understand unfamiliar codebases faster and do more with limited time. The Drupal significance comes from practical work he later highlighted in a LinkedIn comment: the AI Best Practices for Drupal project and Angie Byron’s proof of concept for shared Drupal AI evaluation infrastructure.

Dries identifies two gaps that open-source projects need to address if AI is to reduce rather than reinforce unequal participation. The first is cost, because capable models and coding agents can be expensive. The second is skill, because using AI well requires judgement, experience, and shared practices. He argues that public AI guidance should be embedded into contributor workflows rather than left for contributors to find on their own.

The AI Best Practices for Drupal project attempts to provide a canonical place for opinionated Drupal guidance for AI agents and developers. Drupal.org describes its scope as universal guidance that applies across Drupal core and contributed projects. Its current audience includes contributors and Drupal developers, with possible future expansion to site builders. The project was created by Angie on 26 March 2026 and is seeking co-maintainers.

Angie’s 20 May 2026 post, Toward a shared eval infrastructure for Drupal AI: A proof of concept, addresses a related trust problem: how the community can tell whether AI-generated Drupal advice or code is correct. The post defines AI evaluations, or evals, as tests for language model behaviour rather than tests for whether code runs. One example from the project checks whether an AI-generated Drupal kernel test uses KernelTestBase, avoids UnitTestCase and BrowserTestBase, and passes a php -l syntax check.

The proof of concept builds on a five-layer Drupal Eval Commons umbrella proposal by George Kastanis, known in the Drupal community as zorz, of Point Blank. The proposed layers cover eval cases, result envelopes, storage and distribution, community submission, and domain-specific bundles. Angie’s implementation includes eval cases across six Drupal skills, a standard result format carrying pass or fail details, latency, token usage, and estimated cost, plus a live dashboard for comparing model behaviour. It also explores connecting eval data to production observability through OpenTelemetry.

That work makes Dries’ broader argument more concrete for Drupal. AI-assisted contribution may make it easier to produce patches, but Dries cautions that more contribution is not automatically progress if maintainers face greater review pressure. He argues that the real test is whether AI helps more contributors move from issue to tested patch while making the result easier for maintainers to trust and merge.

The LinkedIn discussion around the essay also raised concerns about capital intensity, vendor dependency, standards, governance, and whether expensive frontier models could reinforce existing barriers. Those responses show why shared infrastructure is not only a tooling question. It is also a participation question for communities that depend on volunteer time and maintainer review.

For Drupal, the practical test is whether shared AI guidance and evals can help contributors produce work that maintainers can verify more easily. If those practices are embedded into contribution workflows, AI-assisted development becomes less dependent on private subscriptions and individual experimentation. The emerging work gives the community a way to examine AI as shared contributor infrastructure, not only as individual developer tooling.