Oof what a silly naming decision. Deepfake already has an established coherent meaning - it's specifically about impersonation with video/audio/images. Just writing in collaboration with a tool is not deep or fake necessarily.
I'm cool with using the word Deepfake for text. Generated text should generally be referred to in the derogatory, considering it very rarely adds anything of value.
So you're cool with using just any words to describe just anything, so long as the vibe is right. Great. I'm sure that'll help a lot with everybody's understanding of the world.
Deepfake is a pejorative not because of the artistic value of the AI but because of the ethnical concerns that spring to mind around one's ownership of their own likeness, and how the tech can be used for deepfake porn and political misinformation. Muddying that conversation with a personal dislike for any AI seems harmful especially to the people most affected by it.
The closest analogue for text I can think of is when you use an LLM to write in the style of someone else, like Shakespeare. But "deepfake" is a stretch because AI-generated text in someone's style is not particularly convincing that it actually came from them, and wasn't too hard to do manually before. Plus doesn't raise the same ethical questions.
Ah, yes. Text generation trained on nearly the entire corpus of preserved human literature, what a useless invention. Yes, let's call it derogatory terms, and maybe it will leave us alone.
Don't most AI either incorporate or surpass GAN architecture, making good software detector impossible to build?
If I understand right, GAN is a co-trained AI similar in size to main AI that discriminate and identify generated AI data apart from sample human data. When the main AI has gotten so good that GAN becomes a dice roll, training is considered finished.
Human brains are completely different implementation and nothing is stopping us from instantly spotting AI, at least for now, but computers shouldn't be able to. Isn't that the case?
> Don't most AI either incorporate or surpass GAN architecture
No.
Most generators don't bother because it is time consuming. Maybe they will if there is a popular detector like this. But it sure increases the cost of training and do people really care? Probably not unless you're explicitly trying to create subversive content. It's also important to remember that tuning a model can also have adverse effects. Sure, you get better in one way, but that can often mean you get worse in another. There's no free lunch.
To "Um AcKtUaLlY" you, GAN is actually a training technique, not an architecture, though we all know what you mean.
Maybe worth noting, that our GAN based networks tend to result in much smaller and faster generative models which can thus produce higher resolutions with much lower resources. The best face generators are still GAN based btw. But you still need big architectures to match diffusion models and you tend to lose your advantages (not all, but it gets nuanced).
> making good software detector impossible to build?
It is a cat and mouse game. I'm not sure there is anything wrong with this, but detector systems tend not to be continually updated, aimed at specific models, and I think their communication is very unclear. You should never say anything is 100% generated, it should be "suspect." So "low suspect" to "high suspect" but never "not AI generated" or "AI generated." The communication is problematic and leads to kids getting flagged for plagiarism when they just wrote something generic.
Though tbh, with a lot of the LLM based detectors, I am really confused by the methodologies. Like why is there this big focus around perplexity? LLMs are trained to be entropy minimizers, so it really doesn't make sense to me that detectors are often looking for... entropy (since perplexity is scaled entropy).
They appear to basically RLHF a model on a bunch of examples of human/AI output on the same prompt. Not sure how well it works, but I'm guessing Mozilla is doing something similar here.
That's the first time I've seen "deepfake" used to refer to generated text, isn't that term usually used for generated imagery/audio which is passed off as real?
No, text is not fake in the same vein. The whole point of deepfakes is that they're images of real people in situations they were not in. Unless you're actively claiming that some victim wrote the text, it's not similar.
Mh fair point. I guess the terminology / actual use of the words will evolve over time as AI becomes ever more prevalent. I'll probably have the deepfake - image/video connotation for some time.
Seems interesting. Also there should be mention that this works for English text only somewhere. Non-English text is not supported. Also If I open the extension and try non-English text, then the "x" button for that page won't close the extension sidebar.
edit: It is even worse, if I'm visiting non-English page then the extension sidebar will be loaded and you cannot close it. You can't control this behavior. This is not a problem with English pages.
Detection algorithms can be less useful in general if they're public and open source, eg. website operators sometimes mass ban on a delay so they can trap a lot of misbehaving users and don't immediately tip off the suspicious behavior.
Although in this case it's already a publicly available extension so even without the algorithm details you could easily check your text against the detector and tune it until it passes... but maybe having the method be public would make it easier for users to prompt engineer the AI to write in ways that wouldn't be caught by it.
>People should be careful with these extensions. It's one thing if they have a GitHub or GitLab repo, where people can easily check out the source, developer history, or any issues presented. It's another thing when that's not the case. Many of its reviews are not promising either.
Someguy posted this immediately deleted his post. It only had a 62% chance of being human.
Comical because the ai didn't realize that the Dev is mozilla.
Yes, this is a lost battle. Ask for a specific way of writing to defeat most existing system. Gain access to the review system to defeat them entirely.
I lost faith in this company, Mozilla has been distracted by chasing hypes, taking money from a major competitor, dropped servo development, the list goes on and on. I had to kagi for "Mozilla Foundation" NPO to confirm that it's still a nonprofit.
In today's view, I'm looking at a Ad Company depends on Google happens to own a browser. This contradicts one of it's main sale point: diversifying the browser market. For me what's left for Firefox is just uB0 support.
There's no remorse in switching to and strongly soliciting Firefox forks like LibreWolf or Waterfox.
Mozilla is cooked because of the DOJ vs Google lawsuit. Google has been overpaying Mozilla to keep Google as the primary search engine in Firefox, but once that goes away they'll have basically no revenue.
