Selected writing · Research data management
FAIR is a gate, not a badge.
Generated code is research data. The 2025 FORGE paper reads it against all four FAIR principles and finds the same answer at each: the tool does not decide, the review does.
Why it exists
Vibe coding lowered the barrier and raised the volume.
When a model writes the technical layer, people who do not program can produce working software. That is a real gain. It also means more code enters research projects without a history, a licence or a test.
The paper, written with Leona Elisabeth Münzer for FORGE 2025, asks a narrow question: can that code be FAIR. Not whether it is useful, not whether it is original — whether it can be found, accessed, reused and interoperated with as research data.
The frame matters. Code in a repository is not a by-product of research. It is the evidence for a claim.
The argument
Four letters, four failure modes.
-
Findable
More code means more metadata and more persistent identifiers, not fewer.
-
Accessible
Licensing is unresolved; training data is opaque, and generated code can duplicate protected work.
-
Reusable
Quality declines without review: code smells, duplicated blocks, conventions ignored.
-
Interoperable
Hallucinated packages, poisoned dependencies, prompt injection, private API keys.
Findable is a volume problem. A productivity gain of the kind GitHub reported in 2023 produces more source than a team can describe. Undescribed code is not findable code.
Accessible is a legal problem. It is unsettled how far generated content falls under copyright, the training data of many models is unknown, and generated code often contains fragments of third-party work. A licence cannot be granted confidently over material whose provenance is unclear.
Reusable is a trust problem. Assistants are poor at removing redundant code and inconsistent with conventions. A codebase that grows faster than it is cleaned loses the quality that makes reuse rational.
Interoperable is a supply-chain problem. A model can name a package that does not exist, or one that does and should not be installed. Private keys and local paths make code neither reusable nor interoperable.
The conclusion is not that generated code is bad. It is that FAIR is a property of a curated object. The gate is human review, with binding rules for quality and licensing.
What holds in 2026
The tools improved; the gate did not move.
Assistants now refactor and document better than they did in 2025. Benchmarks improved. Attribution did not: a diff still arrives without a reliable account of where its parts came from.
The supply-chain risk grew, not shrank, as agents began installing their own dependencies.
What I would add: provenance as a fifth requirement. Which model, which prompt, which date, which sources. Without it, the other four are assertions.
What I would keep: the sentence that a person signs the commit. It is the only part of the argument that cannot be automated.
The record
The deposit is the conference paper: four principles, one gate. The German title is the deposited one.