For decades, there was a small, patient corner of research called Requirements Engineering, the discipline of capturing what a system is actually supposed to do. From 1983 to 2019, people tried to point language technology at it: automatically spotting contradictions in a spec, classifying requirements, finding defects. Useful work, but slow-moving and mostly analytical. The machine read your requirements and told you what looked wrong.

Then ChatGPT arrived, and the field changed character almost overnight.

The first studies on large language models for requirements appeared in 2023, not a year earlier, not a year later. Seventy-four have followed since. In two years, this new wave has produced roughly a fifth of everything the previous thirty-six years managed. What unlocked it wasn't a breakthrough in capability so much as a breakthrough in access: the moment researchers could simply type instead of building pipelines, they did.

But the volume isn't the interesting part. The interesting part is what these studies attempt.

Models are being used to elicit requirements from messy, informal material (user feedback, issue trackers, half-finished notes), to validate them and to build models from them. Tasks that were considered out of reach for traditional NLP are suddenly on the table.

And then there's the detail that should make every software engineer sit up: about one in five of these studies isn't really about requirements at all. It's about generating code and tests directly from them.

That's a quiet inversion of how we build software. Today, requirements are documentation: written once, drifting out of sync with the code within weeks, consulted mainly during arguments. If code and tests can be derived from the specification, the specification stops being a description of the system and becomes the system's actual source. Code turns into an output, something regenerated rather than maintained by hand.

Which shifts where the craft lives. Not in typing the implementation, but in describing intent precisely enough that the implementation follows. Anyone who has tried to write a truly unambiguous requirement knows this is not the easy half of the job.

Software engineering may be about to re-center itself on the thing it always claimed to start with.

The numbers behind it

74 studies on LLMs for Requirements Engineering, all since 2023
+136% publications in 2024 versus 2023
~20%: how much of 36 years of prior research (1983–2019) this wave matched in just 2 years
20% of studies generate code or tests from requirements, not just analyze them
66% / 23% / 11% conferences / workshops / journals: a field moving faster than journal cycles
28% appear at the flagship RE conference; 57% are scattered across other venues, including AI, NLP and HCI communities
2026: the year output could equal all prior NLP-for-requirements research, if growth holds

Source: arxiv.org/abs/2509.11446

SoftwareEngineering #RequirementsEngineering #LLM #AI #GenerativeAI

Originally published on LinkedIn.