Don't credit the LLM(isaacsu.com) |
Don't credit the LLM(isaacsu.com) |
In a perfect world such a disclaimer would help anyone reading the output by setting expectations. We each have different experiences with this, and different levels of acceptance. Knowing that something is using generated content can help anyone either accept / skim / thoroughly review something.
The problem is that we're seeing lots of backlash, often for ideological reasons, or for outdated somewhat technical reasons, so people have less of an incentive to be honest, and tons of incentives to hide their use.
No one minds when I credit a fuzzer for finding a bug. I hope we can use the same pragmatic attitude for LLM help.
Also, I don't hear a lot of people credit fuzzers. I do hear a lot of people say things like "we are fuzzing the code"'just like they say "we are hammering the nails" and not "wow hammer great job getting that nail in place"
Case in point, the recent Hank Green incident. That was an own-goal of being too honest about LLM use and I suspect the industry learning will be "don't talk about the process." Larian Studios got in hot water earlier over a similar "we may use AI in preproduction" incident.
it didn't. it produced an output, and you hopefully invoked some human judgement before believing it was worth sharing, and you are hopefully expecting to be accountable if that judgement was incorrect (whatever the failure mode)
A disclosure is appropriate whenever the reader might feel differently about the result of armed with this information. For example, "this research on the non-addictive nature of tobacco, funded by the tobacco industry".
Moreover, not citing the LLM is plagiarism. Maybe people outside academia have an oddly narrow understanding of plagiarism, and (for example) cannot understand how it is possible to plagiarize yourself. It is hard to explain the importance of citing LLM contributions to people with this view.
Cheating isn't the right word here. Crediting LLMs is about giving everyone involved accurate information which can be used for later decision making (e.g. if something unexpected happens as a result of a LLM-ism).
Yes, if you are writing a longer piece of communication and you just verbatim ship whatever the LLM spits out, it is important that you note that in the communication. This obviously applies to deepfakes as well.
But this quickly gets slippery. If an LLM generates an initial draft of a paragraph, and I edit it 10 times making mostly stylistic changes, do I need to cite the LLM for that passage? How much personal touch is enough to absolve one of LLM citation? We don't require filmmakers to present a "this is CGI" banner during films (only occasionally in trailers), if a filmmaker uses an LLM to generate 20% of the contents of a film, much in the manner of CGi usage, do they need to add a disclaimer?
The text produced by an LLM really doesn't function the same way text produced by humans does. The cited human is a stable entity that exists, not a stateless machine. You can ask them about their quoted work, look up their history, etc. There is a stable participant in the field with a stable identity that you can talk to. Not so for LLM produced text.
At the end of the day whatever it is your present is coming from you. Own it. If you feel the need to hedge that such and such was from an LLM it probably means that it's bullshit that you'd better rework or at least understand much better before you pass your slop off to some poor coworker
I did in fact "bother to review" the tests, and I thought they were good enough to merge. Indeed, they were better than what I would have written on my own: they were lower-quality on a line-by-line basis, but for tests, quantity is a quality of its own, and I would not have bothered to write nearly as many tests. Even so, it was worth flagging that Claude wrote them.
If I trusted LLMs more, would I still feel the need to credit them? Maybe. They are improving very quickly, and I think they will eventually improve to the point where a mostly-LLM-written PR will be as good as mine in every way. Then it will no longer be necessary to apologize for submitting LLM code. But if and when that point is reached, I wouldn’t get used to taking credit for said code! Because it probably won’t be much longer until the human in the loop becomes redundant.
Going back to today, I don’t feel the need to credit LLMs when I use them for tasks that don’t involve submitting their output. In particular, they have recently become quite good at finding bugs in my code.
No it’s not. A spell checker tells me if I misspelled any of my own words.:!it doesn’t write for me.
> or a pilot announcing their version of avionics software over the PA while taxiing to the runway.
No, it’s nothing at all like those things. It would be more like your pilot letting Claude direct his taxiing.
People disclose LLM involvement because they know they didn’t really* author that output, and they want to disavow responsibility for it being totally wrong. Even if I just need a quick figure and I google it, I still disclose that I googled it rather than pulling it out of my own expertise.
When I see it, it always makes me want to respond “weird, I asked Claude too and it said <different answer>”.
Sometimes the person asking just needs an estimate, a ballpark number. It doesn't really make a difference if 17 or 18 support requests were received.
a. You bring no expertise beyond asking an LLm for the answer. b. They can just do that themselves.
If someone is asking you something and you tell them "this is what the LLM said" you're doing it wrong. If they are asking you they assume you possess some kind of acumen you can use to do better than they can blindly following an LLM and if you can't, well, why are you even involved in this interaction?
If you use AI chat as a search tool and then read whatever web pages it finds to check them for yourself (as you should), then you can cite that instead.
Citing an LLM is a sign that you couldn't be bothered to do that, but at least it's honest.
By not crediting and personifying the LLM, one (re)takes full human responsibility over the produced works and stated claims. As a result one won’t become an “agent” for the LLM.
The title could have been Don’t blame the LLM.
When we first started using LLMs professionally it was fine to hide behind disclaimers. Now with widespread adoption it’s time to step out of the LLM’s shadow and take credit/blame again.
Unfortunately, it conflates what's meant by co-authoring. Because so far, Claude has not claimed sovereignty as an artificial intelligence, so actually, mentally crediting Claude is like crediting Vim or Emacs.
We don't have to think of our tools as being authors just because Anthropic did.
The Co-authored-by had a real meaning related to accountability.
Putting our tool in front of us is like letting our cat drive the car. I mean, it's fine if it sits in our lap with the paws on the wheel wearing sunglasses. Just don't pretend it is the one deciding the direction, taking critical turns. Because when we get pulled over, the cops aren't gonna buy it. And neither is anyone on GitHub. We know Claude is still just a program.
Isaac Su seems to be irritated and mildly triggered that people don't think or act like LLMs aren't just dumb, mute, simple tools.
Instead of talking about THAT, he is doing a weird combination strawman/bikeshed-like post about the different imagined, nuanced reasons why he imagines people credit LLM input in attribution.
A more honest take would be:
"Some of you are ascribing more sentience and agency to LLMs than I feel comfortable with. Stop doing that. I'm obviously and fundamentally right that they are dumb tools that can never threaten my intellectual and emotional security or human superiority."
Whether they are right or wrong — one reason that many people are giving credit is that they feel that (at some at level at least) that Claude/DeepSeek/Gemini/etc. is acting enough like a participant or partner to deserve being given credit.
And even if you believe it's just a tool, a text generator that takes in and spits out the Internet, you should cite it like citing a website.
(You don't have to mention LLMs every time you use them because sometimes the credit really doesn't matter.)
> But for the first reason, I’d challenge the notion of sharing credit with a tool, let alone feeling like a cheat for using one.
That's the main reason LLMs aren't "just a tool."
> Credit and accountability are two sides of the same coin. A tool can only take as much credit as it can be held accountable for, so crediting it inadvertently dilutes your own accountability for the work.
Huh? No, credit and accountability are pretty orthogonal. I use tons of open-source tools, and everyone agrees that the respective authors get credit for what they made, but only I am accountable for my use of them. In fact, all open-source licenses spell this out explicitly.
You piloted the LLM, you are responsible for the code. I don't sign my commits with "autocompletions by Intellij". This whole thing is nonsense.
As for the possibilities listed in TFA, I personally think the most common one is people just blaming the tool or hoping to use citation as cover for slop work. It's basically an admission of "I didn't look into this as thoroughly as I myself think I should have to really understand what I'm doing". Don't be that person. If you don't understand what you are doing prompt the LLM with questions until you do understand. It isn't difficult.
If I have a large amount of capital to burn, and an idea that begins and ends with 'build me a Groupon clone for x market', then just pour money into a team of outsourced humans at a software sweatshop to build it. Is it still my work? Again, not many people would object if I were the face of that company, it is after all my app.
I'm struggling to see what the difference is if humans were swapped with LLMs in each case? Is the issue that the LLM models were trained on plagiarized code? I agree we should take issue this aspect, but when it comes to being honest about credit for the 'doing' part, this has never been the case, pre or post LLMs.