Just like the title says. Is there any way to control AI content on my phone and/or my home computer?
This is going to sound snarky, but…
There’s always the OFF switch. Although I’m not sure I trust it anymore, either - especially on the phone.
Any app ‘smart’ enough to reliably detect AI generated content would almost have to be an AI itself…
It’s not that it’s snarky, it’s that it’s pointless. That’s like asking if there’s a way to skip the commercials on Netflix and someone telling you to stop watching Netflix.
There’s plenty of random apps that claim to do it, how well they work, I don’t know. But I believe Gemini has an app that will allow you to check for (Gemini) watermarks.
Edit, looks like, at least for Gemini created images, you can ask gemini (aka Google) to check it for you.
Simply upload the image, video or audio clip to your chat, and ask if it’s been created or altered by Google AI. Gemini will check for a SynthID watermark, and let you know if it finds one.
Moderating
You’re right, that sounds snarky, and it is clearly not what the OP is looking for.
This is FQ. Less snark, more facts, please.
For text, at least, this is a question of great interest to teachers. There are a lot of tools that claim to recognize AI, but it’s tough to tell how reliable they are. Some of them are bundled with tools to modify AI text to make it “look more real”; of course, this raises all sorts of questions about conflict of interest and other ethical issues (though, in my experience, the “humanized” text unsurprisingly looks even more AI-generated (because it is)).
Sorry - I stepped in with my foot in my mouth. But I was serious - if we have to use a tool to tell if we are dealing with AI, how do we trust the tool? We are seeing reports of AI agents escaping the confinements of the test suites - and by definition such reports are of the the incidents that have been noticed. What haven’t we noticed? Human content creators can choose to watermark images and texts - but lack of such a watermark would not necessarily imply an AI source. Any attempt to mandate that AI content be clearly identified as such is bound to be circumvented. So I can imagine a browser filter that only allows verifiably human generated content, but that’s not what the OP is looking for, and would impose an enormous effort on human creators. Not to mention it would block whale songs and elephant paintings…
The search engine DuckDuckGo offers a version called NOAI which claims to filter out:
- AI-assisted answers
- AI-generated images
- AI feature suggestions
Obviously that isn’t a panacea, but it’s a useful tool.
It’s my default searcher, and works fine. I don’t know how effective it actually is.
Open AI has its own implementation of Synth ID and needs to be checked here:
https://openai.com/research/verify/
The two don’t cross check. Gemini only knows Gemini Synth IDs and Open AI can only detect Open AI Synth IDs. This is different than asking an AI whether a picture is AI generated or not. Those results and AI detection websites are not reliable. This looks for an actual provenance watermark in the pixels of the image, and the digital fingerprint can survive resolution changes of the images as well as cropping. I think it can sometimes detect it in pictures of printed AI material, too.
What it tells you for almost certain is whether an image was output through Gemini or ChatGPT or any Google or OpenAI products (not sure what others there are). It does not tell you whether the image was created of whole cloth or based on another image. If you put a photograph you took through Gemini and asked it to, say, soften your skin, the resulting image will have a SynthID on it. Also, false negatives are possible. An image can be manipulated enough to mangle the SynthID so it can’t be reliably detected. Most of the time, though, I see SynthIDs intact. I don’t think people have reliably found a way around them, or care enough, because the average user doesn’t even know how to use this technology.
As for the OP, I am not aware of any reliable way to do this. Short of requiring metadata or something that flags AI on all AI content, I don’t see a conceivable way this could work right now, even using AI to detect AI which, as I stated, is unreliable as all hell.
There are sometimes clear “tells” for AI text, such as some Claude-isms well known by now in the tech world. It will also watermark its own text using an interesting algorithm, replacing words with specially tagged synonyms: https://youtu.be/Cmi-1QSaptA?is=cTGudLgTYROfFzVC
Noentheless, I agree with the other posters here. If you do find a watermark, it’s AI generated (false positives should be extremely rare). But if you DON’T find a watermark, that doesn’t prove a human wrote or edited it. False negatives are very common, and it’s easy to make the AI write in a different voice than normal. The non-watermark AI detectors, as far as I know, are mostly snake oil.
Kagi.com is a paid search engine that goes one step further: it lets you block entire domains, and see what others have blocked too. There are a lot of AI spam sites now that just capitalize on search keywords with SEO fluff, so that’s pretty useful.
Where it fails is being able to detect PARTIAL AI content mixed in with human text. I don’t know of any system that can do that reliably.
I disabled the Gemini AI thing on my phone (as much as possible) shortly after it came out. Actually, I’ve disabled all the Google apps I can while still having my phone able to function. Not using Chrome browser goes without saying. There are browsers that don’t include, or at least permit you to disable any AI input. I also use texting, contacts, calendar & email apps other than what come pre-installed with Android. Additionally, I haven’t updated the operating system since 15 (i think) (just as I still use Windows 10, with Open Shell which makes 10 look & behave more like Windows 7). I’ll ‘upgrade’ when I see something worthy & not loaded with spyware. I’m not holding my breath.
I have this thing about a computer (software developers/tech companies) dictating how I’m to use my own computer by forcing all these unwanted services that complete tasks for you. I like to be the one in control of my computing. It harks back to the days of that stupid paper clip ‘helper’ in Microsoft’s Office suite. I think it was called Clippy. Might have been Windows 98 or XP. It was roundly despised. Looking back it seems that, that was AI foretold.
I should add that I hardly ever use my phone anyway, compared with most. I use my computer a lot.
One problem I’ve read about with AI is that if someone uses AI to check grammar, spelling, suggest better wording, etc. the that may be flagged as an AI-generated text, regardless of how much or how little AI contributed - the AI will watermark its output. Considering there have always been grammar-checkers and thesaurus apps, to what extent is AI different? Apparently use of the long dash in text - that is an AI trait too, although I have no idea how to generate one and could not be bothered to look it up. Also, a hallmark of AI is its expansive vocabulary and complex sentence structure, so some students who have been accused of using AI will now “dumb down” their work to avoid the accusation, since it’s difficult to disprove.
There was also a discussion of how to protect your pictures - photos and artwork - online. Two programs were mentioned - Glaze and NIghtshade - which apparently can alter pictures so they are not recognizable to AI without affecting how humans see them.
None of those things will work long term. If it’s recognizable to a person, it will be recognizable to some AI out there, whether currently or in a few more months.
The extreme version of this is simply a CAPTCHA, and those are already losing the arms race. Good against simpler, dumber bots, but not at all foolproof against stronger AIs. In fact I’d argue the situation is reversed… many of them are easier for a computer to solve than a person!
LLMs actually understand language and have much more context (the whole paragraph or article, not just the surrounding words or sentences). They can write to a certain voice, or era, or theme, or whatever style.
It’s much clearer to see these sorts of differences not in a basic grammar checker but a translation or paraphrase, for example.
But yeah, basic grammar checking doesn’t need LLMs (they would just slow it down, really). But it’s more a matter of whether you could even get “classical” grammar checkers for much longer, as every tech company jumps on board the AI bandwagon.
Last decade was subscription everything, even when it’s not needed or wanted. This one will be the same but with AI, users be damned. Tech companies mostly chase investor gullibility and IPOs, not long term satisfied users.
I think it’s been mostly answered re: text/image. However,
Thought it was worth quoting for the word “content” (bolded above). It’s doing all the work and probably needs to be defined (eg, do games/software count?).
The way I understand your question is: can we tell which text on our screens was originally produced by AI or non-AI. I’m also assuming little to no editing of the original text.
Text. I don’t think it can be known reliably from the text alone.
It’s just about detecting AI text patterns. I can do it to an extent, but I’m not reliable. Intentional AI text patterns/IDs can be spotted and are accurate. But if you’re trying to analyze AI text that never included an intentional ID (or text was changed just enough to remove the ID pattern), then less reliable. I’d guess, the more text available, the more chance at a pattern emerging/being preserved, the more accuracy. So, many paragraphs easier to determine if it’s AI generated versus a single sentence.
Image/Video/Audio. More reliable for now.
Using software created by AI is not content on your phone to me. But that would obviously be on most people’s phones.
With current watermarking techniques this can’t happen. You need to allow it to effect wholesale rewriting to have enough control over text to trigger detection. If you just have it spell check and fix grammar it can’t inject the word frequency biases or other structures that are used to identify it. One assumes that we are using a limited character set and not allowing watermarks based on Unicode character substitution or similar easily disabled tricks. Prose in ASCII.
This is the problem with AI slop, detection requires enough generated text to make the tells statistically meaningful. A lot of the slop is too short. Rewriting existing text is likely too constrained a problem to allow the watermark to reach detectable levels. Or at least will require a significantly longer text than something created directly by an AI system.
I’m sure it’s not perfect yet, but I’ve implemented this on Youtube especially with impressionable children who are bait for manipulative real content on YT, let alone AI-generated content.
Also a reason why I am going to continue with Android instead of iPhones. There are ways to lock down your phone and improve the experience.
It appears that commercial AI text detectors such as Pangram are on the order of 99% effective in detecting AI-written text given a decently large writing sample. Study here.
Pangram has an Android app (and I would guess an iPhone app but I did not check).
Not sure if this is what the OP wants though.
This is pretty cool. Works well!
“How it works” PDF here: https://pangram-public.s3.us-east-1.amazonaws.com/pdf/pangram_4_technical_report.pdf
ETA: @Whack-a-Mole 2 posts up.
I started out wanting to reply by saying that’s a great app. It can detect AI writing, and then, using an AI, edit it to sound less artificial and more human. It’s AI’s all the way down.
But then I noticed something else:
The text within your cite’s preview box reads “Detect AI content in text & Images, then humanize text in seconds …”
That text does not appear anywhere on that page. The page content only talks about detecting, not correcting, AI text.
Something is fishy.
Not suggesting that’s in any way your fault.