Study finds that 52% of ChatGPT answers to programming questions are wrong

https://futurism.com/the-byte/study-chatgpt-answers-wrong

by abunuwas • 2 years ago

61 89 2 years ago

jghn2• 2 years ago

This is looking at the wrong metric. I'm not expecting it to be 100% correct when I use it. I expect it to get me in the ballpark faster than I would have on my own. And then I can take it from there.

Sometimes that means I have a follow on question & iterate from there. That's fine too.

avg_dev2• 2 years ago

you're giving your fellow programmers way too much credit. most people will never bother doing that.

parpfish• 2 years ago

That’s how you, an experienced programmer, use it.

What does this do to beginners that are just learning to program? Is this helping them by forcing them to become critical reviewers or harming them by being a bad role model?

bdcravens• 2 years ago

The same is true of things like code generators and scaffolds.

jghn2• 2 years ago

How is this any different than the age old "googling stack overflow" method everyone's been using for years?

nwoli• 2 years ago

Same could be said of wrong stack overflow answers or random google results. Clearly they’ll become critical of the results if the code simply doesn’t compile, same as our generation sharpened our skills by filtering bad from good from google results

paulsutter2• 2 years ago

If you’re learning to program pretty much the best approach is to write and debug programs. There isn’t a shortcut

It’s like the saying “the fog of war” (at best you have incomplete and flawed information). Programming is just like that

raghavtoshniwal2• 2 years ago

If this increases iteration speed for beginner devs and they learn about code quality post it goes into the real world, it’s not a bad bargain to strike imo.

I think we all partly learnt about code quality by having our code break things in the real world.

ModernMech• 2 years ago

I've been saying from the start that this is not a tool for beginners and learners. My students use it constantly and I keep telling them when they go to chat GPT for answers, it's like they are going to a senior for help -- they know a lot but they are often wrong in subtle and important ways.

That's why classes are taught by professors and not undergrads. Professors are at least supposed to know what they don't know.

When students think of ChatGPT as their drunk frat bro they see doing keg stands at the Friday basement party rather than as an expert they use it differently.

pornel3• 2 years ago

Beginners tend to write awful code without GPT's help, so I don't think it makes things worse.

Answers don't exist in a vacuum. The chat interface allows feedback and corrections. Users can paste an error they're getting, or even say "it doesn't work", and GPT may correct itself or suggest an alternative.

tivert3• 2 years ago

What does this do to beginners that are just learning to program? Is this helping them by forcing them to become critical reviewers or harming them by being a bad role model?

Harming them.

I told a new grad employee to write some unit tests for his code, explained the high level concepts and what I was looking for, and pointed him at some resources. He spun his wheels for weeks, and it turned out he was trying to get ChatGPT to teach him how to do it, but it would always give him wrong answers.

I eventually had to tell him, point blank, to stop using ChatGPT, read the articles, and ask me (or a teammate) if he needed help.

crpietschmann• 2 years ago

Exactly! Just because part of the answer isn't right, doesn't mean the entire answer is useless. It's much faster than only doing a Google search when working out the solution to a problem.

greatpatton• 2 years ago

Sometimes you are going to loose a lot of time trying to make a ChatGPT solution work when Google would have provided right away the right answer... Just yesterday I asked ChatGPT for an AWS IAM policy. ChatGPT-4o provided an answer that looked ok but was just wrong, tried to make it work without success. Just Googled it and the first result provided me the right answer.

Veuxdo2• 2 years ago

I expect it to get me in the ballpark faster than I would have on my own.

This is great if you are an experienced developer who can tell the difference between "in the ballpark" and fixable and "in the ballpark" but hopeless.

nkozyra5• 2 years ago

This is great if you are an experienced developer who can tell the difference between "in the ballpark" and fixable and "in the ballpark" but hopeless.

While true, Stack Overflow wasn't much different. New devs would go there, grab a chunk of code and move on with their day. The canonical example being the PHP SQL injection advise shared there for more than a decade.

panarky2• 2 years ago

If 52% of responses have a flaw somewhere, then 48% of responses are flawless.

That is amazing and important.

The headline should be "LLM gives flawless responses to 48% of coding questions."

RicoElectrico• 2 years ago

Absolutely, it's especially useful when it suggests which libraries to use if you're not familiar with the ecosystem. Or writing boilerplate for popular frameworks, step by step. It can, to a degree, repair errors if you paste it the output.

balder1991• 2 years ago

Might not be a good idea for people not security aware: https://vulcan.io/blog/ai-hallucinations-package-risk/#h2_4

pgm8705• 2 years ago

I agree this is the correct way to use it, and it is incredibly useful in that case, but I think a study like this is valuable in the face of all the hype/fud about how AI Agents can program entire complex applications with just a few prompts and/or will replace software engineers shortly.

thatjoeoverthr• 2 years ago

This. For inexperienced developers, I advise thus; don't consume answers you don't understand. If you can't read it, interrogate it, and find a question at your own level. When you accept its emission, you're taking responsibility for it, and beyond a certain low level, it can't do your thinking for you.

NBJack• 2 years ago

From the article:

What's especially troubling is that many human programmers seem to prefer the ChatGPT answers. The Purdue researchers polled 12 programmers — admittedly a small sample size — and found they preferred ChatGPT at a rate of 35 percent and didn't catch AI-generated mistakes at 39 percent.

alphazard• 2 years ago

This is how we've all adapted to use these tools, but it's not what was originally pitched.

Gigachad2• 2 years ago

Every time I’ve tried it it’s sent me to completely the wrong ball park and after a while whacking it’s solution I end up completely dumping it and doing it myself.

avg_dev2• 2 years ago

That article links to the actual paper, the abstract of which is itself quite readable: https://dl.acm.org/doi/pdf/10.1145/3613904.3642596

Q&A platforms have been crucial for the online help-seeking behav- ior of programmers. However, the recent popularity of ChatGPT is altering this trend. Despite this popularity, no comprehensive study has been conducted to evaluate the characteristics of ChatGPT’s an- swers to programming questions. To bridge the gap, we conducted the first in-depth analysis of ChatGPT answers to 517 programming questions on Stack Overflow and examined the correctness, consis- tency, comprehensiveness, and conciseness of ChatGPT answers. Furthermore, we conducted a large-scale linguistic analysis, as well as a user study, to understand the characteristics of ChatGPT an- swers from linguistic and human aspects. Our analysis shows that 52% of ChatGPT answers contain incorrect information and 77% are verbose. Nonetheless, our user study participants still preferred ChatGPT answers 35% of the time due to their comprehensiveness and well-articulated language style. However, they also overlooked the misinformation in the ChatGPT answers 39% of the time. This implies the need to counter misinformation in ChatGPT answers to programming questions and raise awareness of the risks associated with seemingly correct answers.

ggddv• 2 years ago

Can there be some sort of mechanism on HN for criticism of an unsubstantiated headline?

dang• 2 years ago

You can always email hn@ycombinator.com if you think a headline is misleading, since the site guidelines call for changing those ("Please use the original title, unless it is misleading or linkbait; don't editorialize." - https://news.ycombinator.com/newsguidelines.html)

resource_waste• 2 years ago

Can someone email the author and explain what a LLM is?

People asking for 'right' answers, don't really get it. I'm sorry if that sounds abrasive, but these people give LLMs a bad name due to their own ignorance/malice.

I remember having some Amazon programmer trash LLMs for 'not being 100% accurate'. It was really an iD10t error. LLMs arent used for 100% accuracy. If you are doing that, you don't understand the technology.

There is a learning curve with LLMs, and it seems a few people still don't get it.

mrweasel2• 2 years ago

LLMs arent used for 100% accuracy.

I think you're wrong about that. They shouldn't be, but they clearly are.

51Cards2• 2 years ago

The real problem is that it's not marketed that way. WE may understand that but most people, heck even in my experience a large percentage of tech people, don't. They think there is some kind of true intelligence (it's literally in the name) behind it. Just like I also understand that the top results on Google are not always the best.. but my parents don't.

mrweasel2• 2 years ago

ChatGPT was released one and a half year ago. It basically duct tape code together from a probability model, the fact that 52% of it's coding answers a correct is amazing.

I'm still on the fence about LLMs for coding, but from talking to friends, they primarily use it to define a skeleton of code or generate code that they can then study and restructure. I don't see many developers accepting the generate code without review.

ObnoxiousProxy• 2 years ago

Misleading headline and completely pointless without diving into how the benchmark was constructed and what kinds of programming questions were asked.

On the Humaneval (https://paperswithcode.com/sota/code-generation-on-humaneval) benchmark, GPT4 can generate code that works on first pass 76.5% of the time.

While on SWE bench (https://www.swebench.com/) GPT4 with RAG can only solve about 1% of github issues used in the benchmark.

tasuki2• 2 years ago

Does that mean that 48% of ChatGPT answers to programming questions are correct? If so, that's amazing!

haolez2• 2 years ago

Kind of, but I'd guess that Google searches and Stack Overflow might have a higher success rate than this.

tasuki2• 2 years ago

No way. I send ChatGPT my Haskell code and the unreadable compiler error message, and it tells me what the error means in plain human terms or at least points me in the right direction.

Google and Stack Overflow are useless here, people have different situations than I do.

I find it's worse at providing working code (much less good code), but pretty good at telling me why my code doesn't compile, which is 80% of the work anyway!

Foivos• 2 years ago

This is way better than I thought. A follow-up question would be for the times that it is wrong, how wrong is it. In other words, is the wrong answer complete rubbish or it can be a starting point towards the actual correct answer?

jrvarela563• 2 years ago

Similar to how programmers work, the AI needs feedback from the runtime in order to iterate towards a workable program.

My expectation isn’t that the AI generate correct code. The AI will be useful as an ‘agent in the loop’:

- Spec or test suite written as bullets

- Define tests and/or types

- Human intevenes with edits to keep it in the right direction

- LLM generates code, runs complier/tests

- Output is part of new context

- Repeat until programmer is happy

jrvarela563• 2 years ago

This workflow is very close to being possible. I gave it a try last year by adding exceptions and test output to clipboard automatically (requires custom code for your stack). The context has increased considerably since my last attempt and agents are now a thing (ReAct loop, etc).

This should be feasible this holiday season.

jrvarela563• 2 years ago

This requires:

- function calling: the LLM can take action

- Integration to your runtime: functions called by the LLM can run your tests, linters, compiler, etc

- Agents: the LLM can define what to do, execute a few tasks, and keep going with more tasks generated by itself

- Codebase/filesystem access: could be RAG or just ability to read files in your project

- Graceful integration of the human in the agent loop: this is just an iteration of the agent but it seems useful for it to ask inputs from the programmer. Maybe even something more sophisticated where the agent waits for the programmer to change stuff in the codebase

drewcoo• 2 years ago

To those who constantly claim ChatGPT is "like an intern," just how low are the standards for interns?

Turing_Machine• 2 years ago

I've heard anecdotal claims that most applicants (not just interns) can't even write fizz buzz, so...pretty low?

odyssey7• 2 years ago

Extrapolation:

Odds of correct answer within n attempts =

1 - (1/2)^n

Nice, that’s exponentially good!

meindnoch• 2 years ago

Wrong. The probabilities are not independent.

cjonas3• 2 years ago

I scanned the paper and it doesn't mention what model they were using within chatgpt. If it was 3.5 turbo, then these results are already meaningless. GPT-4 and 4o are much more accurate.

I just used GPT-4o to refactor 50 files from react classes to react function components and it did so almost perfectly everytime. Some of these classes were as long as 500 loc.

haolez2• 2 years ago

I'd guess that React code is a lot easier for a LLM, since it's a frequent occurrence in its training dataset and frontend code tends to be repetitive and full of boilerplate.

I believe that AI will be a perfect programmer in the future for all niche areas. My point is that frontend will probably be the first niche to be mastered.

cjonas3• 2 years ago

I agree, but would say it different:

AI will be a perfect programmer in the future for all NON-niche areas

There's going to be a positive/negative feedback loop that makes it hard for new languages and frameworks to gain popularity. And the lack of popularity means lack of training material being generated for the AI to learn.

When choosing a tech stack of the future, the ability for AI to pair will be a key consideration.

Hackbraten• 2 years ago

It says they used GPT-3.5.

cjonas3• 2 years ago

Ya if it's GPT-3.5, I'm actually surprised the accuracies were so high!

I've been pairing with GPT since 3.5-turbo. I run 20-100 queries a day (have an IDE integration). The improvements for GPT-4 over 3.5 are significant.

So far GPT-4o seems like a step-up for most (not all) queries I've run through it. Based on the pricing and speed, my guess is it's a smaller, more optimized model and there are some tradeoffs in that. I'm guessing we'll see a more expensive flagship model from OpenAI this year.

But honestly, these details don't really matter... Regardless of the performance and accuracy of the models today, the trend is obvious. AI will be the primary interface for writing all but the most cutting edge code.

Two years ago, I thought an AI writing code was 50 years away. Yesterday, I took a picture of an invoice on my phone, and asked GPT to recreate it in HTML and it did so perfectly.

asadotzler2• 2 years ago

Not meaningless when 99% of the people use the free version which apparently has license to lie to them far more than the paid version. What a fucking sick joke, pay up or we lie to you even more.

mnk47• 2 years ago

The new GPT4o model is free. Plus users will get better rate limits and the voice feature, but everyone has access to the best model right now.

happypumpkin2• 2 years ago

From the paper:

"Additionally, this work has used the free version of ChatGPT (GPT-3.5)"

ryanwaggoner• 2 years ago

This is a critical detail. GPT-4 is much better than 3.5 for programming, in my experience.

happypumpkin2• 2 years ago

Yeah I really don't understand why research is still being published that uses GPT3.5 rather than GPT4 or both models. ~500 programming questions is maybe a few bucks on the API?

Last5Digits• 2 years ago

Here's hoping that the average HN commenter will actually read the paper and realize that the study was performed using GPT-3.5.

f0e4c2f7• 2 years ago

This study uses a version of ChatGPT that is either 1 or 2 versions behind depending on the part of the study.

It cracks me up how consistent this is.

See post criticizing LLMs. Check if they're on the latest version (which is now free to boot!!).

Nope. Seemingly...never. To be fair, this is probably just an old study from before 4o came out. Even still. It's just not relevant anymore.

123yawaworht456• 2 years ago

iirc, I saw some other study (or an experiment some random guy had ran) where original GPT4 had vastly outperformed its later incarnations for code generation.

current openai products either use much lower parameter models under the hood than they did originally, or maybe it's a side-effect of context stretching.

jononomo• 2 years ago

I use ChatGPT for coding constantly and the 52% error rate seems about right to me. I manually approve every single line of code that ChatGPT generates for me. If I copy-paste 120 lines of code that ChatGPT has generated for me directly into my app, that is because I have gone over all 120 lines with a fine-toothed comb, and probably iterated 3-4 times already. I constantly ask ChatGPT to think about the same question, but this time with an additional caveat.

I find ChatGPT more useful from a software architecture point of view and from a trivial code point of view, and least useful at the mid-range stuff.

It can write you a great regex (make sure you double-check it) and it can explain a lot of high-level concepts in insightful ways, but it has no theory of mind -- so it never responds with "It doesn't make sense to ask me that question -- what are you really trying to achieve here?", which is the kind of thing an actually intelligent software engineer might say from time to time.

cubefox• 2 years ago

From the paper:

For each of the 517 SO [Stack Overflow] questions, the first two authors manually used the SO question’s title, body, and tags to form one question prompt1 and fed that to the free version of ChatGPT, which is based on GPT-3.5. We chose the free version of ChatGPT because it captures the majority of the target population of this work. Since the target population of this research is not only industry developers but also programmers of all levels, including students and freelancers around the world, the free version of ChatGPT has significantly more users than the paid version, which costs a monthly rate of 20 US dollars.

Note that GPT-4o is now also freely available, although with usage caps. Allegedly the limit is one fifth the turns of paid Plus users, who are said to be limited to 80 turns every three hours. Which would mean 16 free GPT-4o turns per 3 hours. Though there is some indication the limits are currently somewhat lower in practice and overall in flux.

In any case, GPT-4o answers should be far more competent than those by GPT-3.5, so the study is already somewhat outdated.

ChrisArchitect• 2 years ago

Related presentation video on the CHI 2024 conference page:

https://programs.sigchi.org/chi/2024/program/content/146667

nijuashi• 2 years ago

I guess I know how to ask the right programming questions, because my feeling about it is it’s about 80-90% correct, and the rest just gets me to correct solutions much faster than a search engine.

ph4• 2 years ago

I view it as an e-bike for my mind. It doesn't do all the legwork, but it definitely gets me up certain hills (of my choosing) without as much effort.

MrSkelter• 2 years ago

ChatGPT isn’t the best coding LLM. Claude Opus is.

Also as you can always tell if a coding response works empirically mistakes are much more easily spotted than in other forms of LLM output.

Debugging with AI is more important than prompting. It requires an understanding of the intent which allows the human to prompt the model in a way that allows it to recognize its oversights.

Most code errors from LLMs can be fixed by them. The problem is an incomplete understanding of the objective which makes them commit to incorrect paths.

Being able to run code is a huge milestone. I hope the GPT5 generation can do this and thus only deliver working code. That will be a quantum leap.

Study finds that 52% of ChatGPT answers to programming questions are wrong

Related Stories