Literature

AI-Assisted Programming for PhD Researchers

A curated reading list behind the course. You do not need to read any of this to follow the workshop. It is here so you can go deeper, check the evidence for claims made on the slides, and build a reference shelf for your own research software work. Statistics are quoted with the same hedges the sources use.

Primary guides

The practitioner documents the course is built on.

Research on AI-assisted programming

Peer-reviewed and preprint evidence on productivity, learning, and quality. All figures below match the source wording.

Productivity and its limits

Learning, judgment, and over-reliance

  • Bastani et al. (2025), Generative AI without guardrails can harm learning, PNAS. In a field experiment, unguarded GPT access cut later unassisted exam scores, while a guardrailed tutor eliminated the harm. (High-school mathematics; transfer to PhD programming is an argument, not a given.)
  • Prather et al. (2024), The Widening Gap, ICER. Generative AI can create an illusion of competence in struggling novices.
  • Lee et al. (2025), The Impact of Generative AI on Critical Thinking, CHI. Across 319 knowledge workers, higher confidence in GenAI predicted less critical thinking.
  • Sarkar & Drosos (2025), Vibe coding: programming through conversation with AI, PPIG. The first empirical study of vibe coding: developers rely on impressionistic scanning, not systematic review.
  • Denny et al. (2024), Computing Education in the Era of Generative AI, Communications of the ACM. The era’s shift from writing code to reading and evaluating it.
  • Smith & Zilles (2024), Code Generation Based Grading, ITiCSE. Explain-in-plain-English as a check on genuine understanding.
  • Bouvier et al. (2025), The Rest of the Robots: Generative AI in Post-introductory Computing Education, ITiCSE-WGR. The most PhD-relevant working-group report; the generate-then-evaluate discipline.
  • Osmani (2026), Comprehension Debt. The growing gap between how much code exists and how much any human understands, paid back with interest during later revisions.

Security of AI-generated code

  • Veracode (2025), 2025 GenAI Code Security Report. About 45% of AI-generated code samples carried OWASP-class vulnerabilities, and Veracode’s Spring 2026 update reports the same ~45%, a security pass rate stuck near 55% for two years.
  • Pearce et al. (2022), Asleep at the Keyboard?, IEEE S&P. About 40% of 1,689 Copilot-completed programs were vulnerable across 89 CWE scenarios.
  • Perry et al. (2023), Do Users Write More Insecure Code with AI Assistants?, ACM CCS. Participants with an AI assistant wrote less secure code and were more confident it was secure.
  • Spracklen et al. (2025), We Have a Package for You!, USENIX Security ’25. Package-hallucination rates of 5.2%–21.7%, the basis for “slopsquatting” supply-chain attacks. A 2026 re-run on frontier models (Churilov) finds the range compressed to 4.6%–6.1% (lower), but 53 hallucinated package names remained registrable, so the risk persists.

Git & research software practice

The reproducibility and version-control canon for researchers.

AI in research & integrity

Evidence on AI in scholarly work and the policies that now govern its use.

Courses & further material

Sibling and complementary courses on coding with AI.

Tool documentation

The reference docs for the tools used in the labs.

  • OpenCode, docs. Install, configuration, providers, permissions, MCP, and agents.
  • Zed, documentation. Editor, edit prediction, and external agents via ACP.
  • Mistral, documentation. Models, pricing, tiers, and data-privacy controls.
  • Model Context Protocol, specification. The open protocol behind the tool integrations in Lecture 07.