This post is a verbatim copy of two LinkedIn posts I wrote today.
The first explores the recent concerns about watermarking LLM outputs (which is a good thing).
The second explores the ethics of disclosure. I don’t mention LLMs at all in the second post, but they’re obviously linked.
LLMs and watermarks
Why are some people so upset that LLM companies have agreed to watermark their outputs in order to access the EU market? No one seems to remember that Google’s been doing it for two years already.
If you use an LLM, it’s good that we know you’ve done it. If it does that by weighting certain word choices as a sort of signature, why is it bad that it’s doing that?
If models themselves are training on unwatermarked AI output, it’s a race to the bottom. We need this feature rolled out broadly for the sake of our general knowledge base and it’s already four years late.
If you’re not reading the machine’s output and making it your own, is your day job just performative at this point? If you’re just summarizing a document, you do credit* the robot, right?
(It’s not like we couldn’t tell when someone’s writing got a whole lot more verbose and folksy in the last three years.)
I don’t have a problem with people using these things as tools. I have been required by my current employer to use LLMs and indeed there have been some excellent use cases from time to time. But I still go and read what it did and improve it where I can. It’s not making me work any faster. I spend more time reviewing than writing (words, code), but it’s exactly the same speed when you average out where the thinking happens. (And if you think it’s making processes faster at creating things, you’re mistaken or oblivious to where the time payment happens.)
And it’s not just at work. I see countless former good presenters use common LLM phrases in their scripts (there’s one creator I used to enjoy who phones it in with “It’s like X in a trenchcoat” or “It’s just Y wearing a moustache”).
I see you. We all see you. We can’t tell what changed but we know it changed. Just admit you’re doing it.
At least the LLM companies can provide tooling so that we can see it for ourselves. Thank the EU for that. They got USB-C wrong, but they’re doing the right thing here.
P.S. I’m not talking about machine translations and copy edits and document summaries and the conveniences of doing busywork that LLMs give you. If you use an LLM to do that, so what? We already know you didn’t write that monthly newsletter with all the charts that no one reads. If you port a program from one language to another, so what if it’s watermarked?
On the other hand, if you use an LLM to do that for a school assignment or a paid gig and don’t reveal it, and then get flagged for cheating, my sympathy glass runs low. I’m going to have a word with you anyway. (Editors exist for a reason.)
- Crediting an LLM is fraught. I’m not saying it does original work. AI art is theft. Read the room.
The ethics of disclosure
In its simplest form, “ethics” describe a pattern of behaviour. Those ethics are codified as an ethos, or “character”. This is your social capital.
There are social pressures on everyone at all times. We generally agree to a set of laws, and morality can differ between groups and even individuals. Ethics are a social contract to demonstrate trustworthiness across those boundaries.
So, your moral code is environmental, but you get to choose your character, and it’s how people judge you.
If I find myself in an ethical dilemma, I might ask myself “what is the right thing to do?”
(The right thing is not always the legal thing, or repeals and amendments wouldn’t exist. It’s more a vibe of what would be seen socially as just or fair. A significant chunk of humanity sees that as a religious vibe. I’m not one of those people, but we often have similar goals.)
Remember when we made deals on a handshake and giving our word? It’s not the handshake that’s important.
I try to be like those people: my word is my bond. If I promise to do something, I will try my best to get it done (on time and under budget).
So what is “disclosure”, and what makes it an ethical decision?
There’s a long line of human behaviour where we borrow ideas from others and build on them. It’s a very long line, reaching before recorded history. (Copyright is a recent development. In my opinion, it exists because we need it in a capitalist society. And capitalism has abused what it what meant to solve: if you make something, you get to support your family from your efforts.)
An obvious example is in music. There is a glorious tradition of sampling, interpolating, and remixing existing work. Jazz embodies this.
Art is art. You create a new thing, that’s definitely yours, but is clearly inspired by someone else. It’s so easy to say “I was inspired by X”. So say it.
Royalties being what they are, you might have to pay to use samples. Many artists do. Around half of the royalties for Gotye’s song “Somebody That I Used To Know” go to other artists.
What about writing? If I write a three act novel framed as a hero’s journey, it’s obviously based on the Odyssey. But most of our stories are that shape. It’s now “common use”. But when I write technical documents, it’s ethically correct to cite my sources. There happens to be legal precedent to do this so I don’t get sued, but the choice is an ethical one.
We thrive when we live and work together. There is no shame in sharing that you didn’t do it all yourself.
There’s no shame in sharing.
I’m asking you to acknowledge when you use work created by someone else. You not only improve your social capital by doing so, but you boost the work and raise awareness of the people whose work you used. A high tide raises all boats.
Besides, “works well with others” looks a lot better on your resume than almost anything else.
Give credit copiously. Share the love. Share. We’re in this together.