An analysis of GPT-3.5-based ChatGPT's answers to 517 programming questions on Stack Overflow: 52% of answers contain incorrect information, 77% are verbose
https://futurism.com/... Charlie Stross / @cstross@wandering.shop : Study Finds That 52 Percent of ChatGPT Answers to Programming Questions Are Wrong — https://futurism.com/... Shafik Yaghmour / @shafik@hachyderm.io : “Our analysis shows that — 52% of ChatGPT answers contain incorrect information and 77% — are verbose.”: https://dl.acm.org/... If you have to basically verify everything then what is the point? I can see how it may point you in the right direction and that could be helpful but it could easily do the opposite as well. … X: Armand Domalewski / @armanddoma : no one doing any actual programming work is using GPT-3.5, this is like writing a study about how much the iPhone sucks and burying in the footnotes you used the iPhone 3 Professor Erwin Loh / @erwinloh : Analysis of #ChatGPT answers to 517 programming questions finds 52% of #ChatGPT answers contain #incorrect information. Users were unaware there was an #error in 39% of cases of incorrect answers. https://dl.acm.org/... [image] Gary Marcus / @garymarcus : Important work, by @SamiaKabir12 and others: https://dl.acm.org/... Gary Marcus / @garymarcus : Brutal. Absolutely brutal. So much crap code is probably being written. [image] LinkedIn: Rafael Brown : AI is unreliable, but people like the misinformation because its easy. And corporations like the misinformation because they can make money off it. … Kris Jenkins : Is ChatGPT and / or other LLMs or GenAI the elevators to success? — Last night Joseph Turner shared a paper that did some empirical research around … Saeed Al Dhaheri : This study from Purdue University researchers shows that 52% of #chatgpt answers to questions about programming are wrong … Jason Arbon : 😱be careful when generating test code—in general I don't recommend it FWIW 🤷 — “....52% of ChatGPT answers contain incorrect information … Tyler Jensen : Seriously flawed. Purdue? Shouldn't we be able to expect better from such an august university? — “The Purdue researchers polled 12 programmers — admittedly a small sample size.” … Dr. Jeffrey Funk : team of researchers from Purdue University showed that 52%" of programming answers generated by ChatGPT are incorrect, 77% are more verbose than human answers … Pranab Ghosh : A correctness of 52% is almost as good a model based on random coin flipping. A Transformer based auto regressive model like GPT will never learn programming logic. …
Context & Ripple Effects
The Purdue evaluation puts a concrete programming-use case behind earlier concerns that generative systems can produce fluent output without dependable grounding, echoed in the critique of GPT-2's superficial knowledge. It also sharpens the error-tolerance question raised when ChatGPT-style tools became broadly usable across new AI use cases.
Because the test concerns GPT-3.5-based ChatGPT and uses a small poll of 12 programmers, it is evidence about a particular model and task rather than a universal measure of coding assistants. Its value is in showing that persuasive presentation can obscure mistakes: evaluators missed errors in a meaningful share of incorrect responses.
First-order effects
- Developers using this version of ChatGPT for Stack Overflow-style questions must treat answers as leads to test and verify, rather than drop-in code or authoritative explanations.
- The combination of incorrect answers and high verbosity raises the review burden: users may spend more time locating the actionable claim and checking whether it is sound.
Second-order effects
- Teams adopting coding assistants are pushed toward guardrails such as tests, code review, trusted documentation, and narrower task scopes; the value proposition shifts from answer delivery to assisted investigation.
- Providers of coding tools face pressure to distinguish model capability from answer reliability and to improve evaluation on practical programming questions, not just fluent output.
Third-order effects
- If coding assistants continue to generate plausible but uneven guidance, software organizations may treat verification capacity as a core complement to AI use—limiting gains where users cannot cheaply test outputs.
- The result reinforces a broader split between general-purpose chat models and tools designed around constrained workflows, retrieval, or validation; whether that split persists depends on measurable reliability improvements.
The trend: Generative AI is moving from novelty-driven adoption toward workflow-specific deployment, where the economics depend on the cost of detecting and correcting confident errors.