Princeton Digital Laboratory’s Ethical AI agenda examines how AI is actually being used. Specification-driven development is an emerging method for putting AI to work in software, and it makes a two-part claim: that people stay in the lead, and that the result is better. This mapping study establishes where research on SDD currently stands, and sets it against the earlier development methodologies that made related claims.
Specification-driven development arrived fast. Between June 2025 and April 2026 it went from a conference keynote to an open-source toolkit from GitHub, an IDE from AWS, and an end-of-support notice for the product that IDE replaced. What it did not acquire in that time was evidence. This study maps 138 sources across the academic literature, the industrial tooling and the practitioner argument, and finds the three barely touching — the industrial strand contains no peer-reviewed sources at all, and not one of the named tools has been evaluated in peer review. Neither half of the claim has been tested. That better specifications yield better software has no measurement instrument anywhere in the reviewed literature. That people stay in the lead turns out to be a more modest proposition than it sounds: five of twenty-four current industrial sources treat the specification as the program, against thirty-two of forty-four in the movements that came before. Those movements — automatic programming, formal specification, Cleanroom, literate programming, model-driven architecture — were each told by their own advocates, writing afterwards, what had gone wrong. The argument that writing an adequate specification is as hard as writing the program is thirty-eight years old, and in the literature reviewed here it goes unanswered.
Published under CC BY 4.0. Paper: 10.5281/zenodo.22731487 · Extraction sheet and search log: 10.5281/zenodo.22731354“