[Paper] Feasibility of AI-Assisted Programming for End-User Development
Source: arXiv - 2512.05666v1
Overview
Irene Weber’s paper investigates whether large‑language‑model (LLM) copilots can replace—or at least augment—today’s low‑code/no‑code (LCNC) visual builders for end‑user development. By running a small user study where non‑programmers built a web app with an AI assistant, the work shows that “AI‑assisted coding” is not only doable but also welcomed by participants, hinting at a shift in how organizations let non‑technical staff create digital tools.
Key Contributions
- Empirical evidence that non‑programmers can successfully complete a real‑world coding task using an LLM‑driven assistant.
- Comparison of AI‑assisted development against traditional visual LCNC platforms, outlining benefits such as flexibility and reduced vendor lock‑in.
- Design guidelines for integrating AI copilots into end‑user development workflows (prompt framing, iterative refinement, debugging support).
- Discussion of broader impact on software engineering education, tooling ecosystems, and organizational digital transformation strategies.
Methodology
- Participant recruitment – 20 volunteers with no formal programming background were selected from a corporate environment.
- Task definition – each participant had to create a simple CRUD web app (form entry, list view, basic validation).
- Tooling – a chat‑based LLM assistant (similar to ChatGPT/GPT‑4) was provided, together with a code editor and a one‑click deployment sandbox.
- Procedure – participants interacted with the AI via natural‑language prompts, iteratively refining generated code until the app ran correctly.
- Metrics collected – time‑to‑completion, number of AI‑assistant interactions, error‑rate, and post‑task satisfaction surveys.
- Analysis – quantitative results were complemented by qualitative feedback (open‑ended comments, observed strategies).
The study deliberately avoided any visual drag‑and‑drop builder, forcing participants to rely solely on the AI’s code generation capabilities.
Results & Findings
| Metric | Outcome |
|---|---|
| Task success rate | 85 % (17/20 participants delivered a working app) |
| Average time to completion | 27 minutes (± 8 min) – comparable to typical LCNC build times reported in prior work |
| Prompt iterations | Median of 6 interactions per participant (prompt → code → refinement) |
| Error correction | 92 % of bugs were fixed after the first AI‑suggested fix, indicating effective conversational debugging |
| User sentiment | 78 % expressed confidence that AI‑assisted coding could replace visual LCNC for many routine tasks |
Key take‑away: non‑technical users can leverage LLM copilots to produce functional code quickly, and they appreciate the conversational, “ask‑and‑receive” workflow more than fiddling with visual blocks.
Practical Implications
- Tool vendors can embed LLM copilots directly into their platforms, offering a hybrid UI that lets users switch between visual blocks and natural‑language code generation.
- Enterprise IT can reduce reliance on proprietary LCNC stacks, lowering licensing costs and vendor lock‑in while still empowering business users.
- Developer teams may shift from building “all‑the‑things” themselves to curating prompt libraries and validation pipelines that guarantee code quality and security.
- Rapid prototyping: product managers can spin up proof‑of‑concept features in minutes, then hand off the generated code to engineers for polishing, accelerating the innovation cycle.
- Training & onboarding: new hires or citizen developers can learn programming concepts through conversational guidance, shortening the learning curve without formal courses.
Limitations & Future Work
- Sample size & diversity – the study involved a relatively small, homogenous group; broader demographics (e.g., varying technical literacy, different industries) need validation.
- Scope of tasks – only a basic CRUD web app was tested; more complex workflows (integrations, performance‑critical code) may expose current LLM weaknesses.
- Security & correctness – generated code was not audited for security vulnerabilities or long‑term maintainability; future work should incorporate static analysis and compliance checks.
- Human‑AI interaction design – optimal prompting strategies and UI affordances (e.g., inline explanations, version control) remain open research questions.
The paper concludes that AI‑assisted end‑user coding is a promising, feasible paradigm, but scaling it to enterprise‑grade applications will require deeper studies into robustness, governance, and seamless human‑AI collaboration.
Authors
- Irene Weber
Paper Information
- arXiv ID: 2512.05666v1
- Categories: cs.LG, cs.AI, cs.SE
- Published: December 5, 2025
- PDF: Download PDF