Whatever the original intention, this is clearly a bug and should be fixed. But should ChatGPT sign its cartoons with its own name or leave them unsigned?

▲

kzsh 7 hours ago | parent | next [-]

I think what you're seeing is the probability of a particular signature or style of signature appearing on a particular style of cartoon, not an intent to sign.

▲

dotancohen 7 hours ago | parent | next [-]

I do many voice transcriptions. Many times I've had empty silence at the end of recordings transcribed as "Thank you" or even "Like and subscribe".

▲

ctippett 7 hours ago | parent | prev | next [-]

This. It's also while you'll sometimes get a mangled Getty Images watermark on some image generations, or a logo in the bottom left corner. If it's a prominent feature in the training dataset it'll show up, exactly how these models are supposed to work.

The 'bug' here is whatever post-processing step or system prompt is in place to steer the model away from doing this.

	▲	zahlman 5 hours ago \| parent \| next [-]
		Maybe it would have worked better to preprocess the training data…
	▲	miltonlost 5 hours ago \| parent \| prev [-]
		Maybe don't use stolen images in the first place.

▲

Muromec 7 hours ago | parent | prev [-]

Yes, but that's just the pixel generation layer. The harness or whatever infrastructure around could nudge it towards common sense.

▲

saalweachter 8 hours ago | parent | prev | next [-]

I mean, the bug is, "it generated the most likely cluster of pixels in the corner of a New Yorker cartoon".

▲

DonsDiscountGas 7 hours ago | parent | prev | next [-]

SynthID watermark but no visible signature

▲

Forgeties79 8 hours ago | parent | prev | next [-]

No one should be allowed to claim they drew something when they didn’t draw any part of it and LLM’s are not people/can’t work without a person. We don’t credit pens and paintbrushes after all.

One could argue nobody should be allowed to claim it. It just exists.

▲

zahlman 5 hours ago | parent | next [-]

For the record, the original tweet doesn't explicitly claim anything at all; it just shows the generated image.

▲

_carbyau_ 6 hours ago | parent | prev [-]

But it didn't just blink into existence.

Even if the person can't claim the copyright of the image produced they ARE responsible for the use of their tools and what they do with the output.

In this case, they released an image with someone else's signature on it. That is wrong, the person should take the blame for that.

The person releasing the image may take it up with the AI service that their tooling led them into making such a mistake. But good luck with that in court...

	▲	Forgeties79 6 hours ago \| parent [-]
		i’m not saying it’s airtight, this is a pretty unbaked idea. But I think it pretty clearly has some legs to stand on. Can’t be any worse of an argument than somebody prompting an AI for a (facsimile of a) photograph and going “I made a photo.” And mind you I’ve heard people argue that that absolutely constitutes taking a photograph (which I wholly disagree with) here on HN. Somebody “made something.” But just because you do something doesn’t mean you get to claim whole ownership of it and get to sign it with your name. Plenty of examples in life.

▲

johnnyanmac 7 hours ago | parent | prev | next [-]

We call it a "watermark" when a machine "signs" its work. And I wouldn't be surprised if its required as a part of AI disclosure in the coming years.

	▲	zahlman 5 hours ago \| parent [-]
		That's a different kind of signature, for a fundamentally different purpose.

▲

dorkwood 7 hours ago | parent | prev [-]

Why is it a bug? If other parts of the generated illustration are similarly taken from an artist, why not the signature as well? Why is a signature crossing the line but the rest of the image isn't?

▲

DonsDiscountGas 7 hours ago | parent | next [-]

For the same reason I'm allowed to draw, paint, or write things very similar to what others have drawn, painted, or written but I have to sign my own name not theirs.

▲

pessimizer 7 hours ago | parent | next [-]

You are a person, LLMs are not. You know this, which is why you know that if you signed someone else's name it would be forgery, but when you see the machine do it you call it a bug.

If the machine is like you, the machine is a forger. The machine is not like you, it is simply blending the work of others to order. Adding someone else's signature is simply part of that statistical process.

	▲	retsibsi an hour ago \| parent [-]
		I think this is a tangent. The obvious problem with signing someone else's name to work they did not create and do not endorse is that, if you publish the image, you will be falsely attributing authorship to them without their consent. The intent (if any) behind the addition of the signature is a separate issue.

▲

dorkwood 7 hours ago | parent | prev [-]

[dead]

▲

none_to_remain 5 hours ago | parent | prev | next [-]

I don't even believe too much in Intellectual Property but the signature amounts to a false representation, rather than a resemblance. The pixels are not special in the image file, but they're special when we look at them. This behavior should be trained out of the model.

▲

Maken 7 hours ago | parent | prev [-]

Because it removes plausible deniability, duh