A Student's Guide to Writing with ChatGPT

Original Article ↗ Hacker News Discussion ↗

Uses of LLMs for Learning and Work

  • Many see LLMs as powerful assistants, not replacements: for “rubber ducking,” brainstorming, generating outlines, alternative phrasings, counterarguments, and feedback.
  • Common technical uses: writing regexes, basic parsers, debugging config files, producing boilerplate code, and quickly prototyping UI components.
  • Some students and professionals report large productivity gains (e.g., “90% of my best work in a fraction of the time”) when they already understand the material and use AI to speed formatting, citations, and grunt work.

Limits, Failures, and Quality Problems

  • Models often hallucinate, ignore corrections, or repeat the same broken code; 4o is repeatedly described as worse than earlier GPT-4 or Claude in this regard.
  • Domain coverage is uneven: well-trodden coding tasks work; niche topics, specialized frameworks, and some disciplines produce fabrications or brittle solutions.
  • Several posters tried to learn from LLMs (e.g., sockets, DBMS) and found answers verbose but conceptually wrong, leading them to conclude LLMs only help once you already know what you’re doing.

Impact on Students and Learning

  • Heavy concern that students use LLMs to generate entire homework, reports, and even take-home exams without understanding, mirroring earlier “calculator dependence” or Google-copying.
  • Teachers report obvious AI-written reports with glaring analytic errors, yet many still pass through.
  • Worry that offloading initial ideation erodes “productive struggle,” critical thinking, confidence, and deep understanding; risk of intellectual atrophy and overreliance.

Educator Responses and Assessment Design

  • Some instructors embrace LLMs but raise the bar: larger system projects, CI pipelines, documentation, oral explanations, and in-class defenses.
  • Others emphasize “teaching against” AI early (no-LM basics), then “teaching with” it later (analyze, test, critique AI-generated code or essays).
  • Proposed tactics:
    • Shift grading to in-class work, oral exams, and niche topics where LLMs fail.
    • Require students to expose their AI use (shared chat links, acknowledgments) rather than ban it.
    • Avoid unreliable AI-detectors; several teachers mistakenly used LLMs to accuse students of cheating.

Philosophical and Future-of-Education Debates

  • Strong split between “adapt like we did with calculators/compilers” vs. “ban or tightly constrain LLMs for students.”
  • Disagreement whether AI will raise the bar (more advanced work faster) or lower it (widespread shallow competence, lost fundamentals).
  • Some argue education should pivot to teaching AI use itself; others insist core skills and manual practice remain essential for real understanding.