Mozilla has released another Firefox extension after Orbit. Deep Fake Detector is a free beta extension for Firefox that determines whether text has been created by AI or humans.
It is part of Fakespot, a service that Mozilla acquired some time ago. Fakespot checks reviews on popular shopping sites to determine whether they are legitimate.
The new browser extension uses Fakespot’s ApolloDFT Engine, but it works with text on any website.
Using the extension is quite simple: You select some text and either right-click on the selection and select “Scan for Deepfakes”, or activate the add-on’s icon in the Firefox toolbar to do so.
The extension needs an active Internet connection and the text that you have selected is passed on to the detection engine. The result tells you the AI’s assessment in regards to human or AI authorship.
It is interesting to note that the extension queries multiple services. It does rely on Fakespot’s own engine, but also three additional ones (ZipPy, Binocular, and UAR). These may not always come to the same conclusion, as you can see on the screenshot below.
You may disable models in the settings of the extension. There you find ratings for each of the used models and options to disable them.
Ratings include the speed of analysis, domain versatility, short and long text handling. If you have a preference, you can disable the others to focus the analysis on that one engine.
What is it good for?
The Internet is flooded with AI-generated texts already. Be it reviews on shopping sites, blog posts, comments, posts in forums or on social media sites, and in many other places.
Finding out if a text has been generated by AI can be useful to determine trust. While that should not be the only factor, it may be useful to weed out AI-generated content to focus on the human-generated content instead.
The developers have big plans for the extension. They want to add image and video analysis in the future, which would make it more versatile and useful.
Closing Words
All in all, Deep Fake Detector is a useful extension that may help you detect AI generated text on the Internet. While it may be wrong at times, use of multiple models may decrease the chance.
Most models rely on Internet connectivity as computations are run on servers and not on user systems. This is the case for the extension as well. This may change in the not-so-near future, but for now, it is the norm and little can be done about it.
What is your take on the extension? Would you use it to check text? Feel free to leave a comment down below.
How to establish conclusions on the basis of probabilities?
– For one, IA is increasingly sophisticated to the point of combining efficiency and a look, a style of modesty when the aim is to achieve a human-like style.
– For two, deep fake detectors deliver their analysis in terms of probability. For my concern I’d require all detectors applied to a given text to agree the text is, or is not, a fake to trigger the start of the beginning of the idea the text may likely be, or not, fake.
– Mozilla’s Deep Fake Detector, as stated by the article, includes 4 engines (Fakespot’s own engine, ZipPy, Binocular, and UAR) . If all 4 engines deliver the same analysis, I’ll start shifting doubt to one side or to the other, otherwise doubt will remain neutral, so to say, that is with the same status as if I had not used any anti-fake engine.
– Conclusion : I’m unlikely to use a tool that could be an incentive to fall in the trap of certitudes with the lack of bullet-proof evidence.
I tried original Fakespot before for a couple of years on Amazon. And I have no clue if it worked properly. Fakespot’s grade was always low. It was never A. Few times it was B. The lion share of grades was C. And there were quite a few D’s. Like a previous commenter said, these evaluations are highly subjective and unreliable, even before AIs. So even though I tried their previous product, I cannot recommend or reject this extension. But I probably will not use it. My method is go to reviews that 1/5 or 2/5 stars and see most common problems. If the same problems mentioned a lot, product probably is not reliable.
AI policing AI, what could go wrong?