46
votes
How are you dealing with AI generated results in your searches?
I'm finding it more and more frustrating having to parse the things I'm actually looking for from what feels like a deluge of rubbish. Are there any strategies, extensions, add-ons, etc that people are using to filter results?
I don't have a great solution, but this is not just a problem, it is a rapidly worsening problem.
Just recently on YouTube I noticed a video on my homepage about current conflict/politics from a channel I'd never heard of. I look over the video and it's full of suspect, likely false information. Looking through the history of the channel reveals it started posting this kind of video as of 3 days prior, 1 video every 5 hours or so, with a 1 month gap of silence before that, and a years-long history of turkish-speaking cooking videos before that. Hacked channel most likely.
Then there are the comments. I appear to be the only one calling out how it's fake. Tons of comments discussing the video content - I look into the users posting those comments: blank pages, no playlists, no user picture, no history of actually existing - bots. Hundreds of bots filling in fake conversation on a fake video to substantiate it and push it in the algorithm.
Not long after having opened that video, another channel I'd never seen before is on my homepage. Also fake crap, but not a hacked channel - this one gives a better appearance of being a real (if incredibly slanted / biased) channel producing conflict / world politics content. No history in the videos of it being a different channel... but then I notice the pace of video release: a 12 to 25 minute video being released EVERY HOUR, 24 HOURS A DAY FOR MONTHS IF NOT YEARS. I literally got tired of scrolling further into the channel history to see how far back in time it went. Generative AI channel set up to absolutely FLOOD fake content, all with surface-believable comments in each video from hundreds to thousands of different user accounts, I'd guess 98%+ of which are bots. Literally no point commenting 'hey, this is fake/AI' when it will just generate a new video in less than an hour, with another after that forever.
This has very much become an extension of warfighting operations for adversarial nations, polluting the signal to noise ratio in our information, creating false narratives, attacking the credibility of legitimate sources and more.
Between disinformation-AI, warfare-AI, and all the development in drone warfare, it really is feeling like a dystopian slide towards a mix between Black Mirror and Terminator 2 these days.
Edit 2024-May-30 T:0800: if anyone is interested to look over the examples I was talking about, I made a couple videos about it, in part to document things in case they disappeared (which the first channel kind of did, they took their 3 days of fake videos down after I commented in each of them calling them out).
Video 1: Hacked YT channels: be skeptical of sources you don't recognize.
Video 2: Hacked YT channels and AI disinformation part 2
Edit 2024-May-30 T:0808: looks like that first channel has resumed posting fake/bias-slanted videos again, I just refreshed their page and 2 new conflict/politics videos are up. So apparently they took their videos down and went silent for a couple days as an evasion method. It's probably a reasonably effective thing to do - if I hadn't checked back I might have assumed that specific channel was no longer being used.
A lot of what you're describing is the Dead Internet Theory and AI is rapidly accelerating the problem.
That was a depressing thing to read, and yes, it does seem to have a significant overlap with what I was seeing.
FYI You can sort a channel's YouTube videos by oldest to show the oldest listed one without any scrolling
I realized I had missed something: the channel information shows how many videos the channel has. The one that was producing a video per hour has over 40 thousand videos posted. FORTY THOUSAND VIDEOS.
At that point I start wondering how many such channels there are, then I wonder if it is, in part, an attack on the ability of the platform (YouTube) to operate. Storage and bandwidth are both plentiful and cheap, but they are not infinite and not free towards infinite scalability.
I keep remembering Tom Scott's tentative prediction that YouTube won't exist by 2030
Pretty far fetched. If that were on a betting market, I’d lay pretty substantial money against it.
Was just about to mention this. I used up my 100 searches in ~4 days and it asked me to subscribe. One of the best $10 per month expenses I’ll make. Have been very happy with the service.
I’m really tempted to try Kagi. I love how straightforward their business model is. For now I’m using DuckDuckGo and generally happy with it. Of course it’s largely just a frontend for Bing, which is headed down the same path as Google so who knows what’s coming around the next corner.
If you end up trying it out, it's worth it to look at the instructions because there are many useful features that you won't necessarily realise are there. I was also using DDG before making the switch. For my searches, Kagi does a lot better geographical filtering (when I need it) and the results are a lot more diverse.
I love the 'quick response' feature, the ability to customise how the results are displayed, and the fact that I can use different Lenses to find academic articles, forum posts, or a host of other types of results. These can even be combined into a customised mix of result types. I haven't needed to do so yet because I'm already getting such a nice selection of diverse results.
I can also easily check from a panel, without clicking the link, how many trackers that site has and how much traffic, and other info. Although I'm using ad blockers, I still prefer not visiting the sites that are infested with trackers because I have less trust in the quality of their content.
I highly recommend it. Using it made me feel like it was 2016 again.
Are there any services similar in quality to Kagi that offer enterprise packages? Seems that Kagi's enterprise offering is AI powered internal search. I work at a tech company so web searching is an important part of almost every role, and as Google's search results decline it would make sense for a company to want to invest in improving web searches for their employees.
If this is talking about what I think you meant, you might want to check out https://help.kagi.com/kagi/plans/team-plan.html which is still sufficiently web search-focused, I’d say. And I’ll believe it is among the, if not the single best offer available for that. :D
Thanks so much, that's what I was looking for :) Interesting that they let you hook into the API directly, but I guess it makes sense for what you're paying.
I subscribe as well. Beyond the search being really good, I actually use their Fastgpt front end quite a bit as well (which is basically using an Anthropic model on its search data).
I'm in a technical field, but it's been pretty good at answering very specific scientific questions. While it isn't perfect, it does provide reference links (many times to specific scientific papers) which are helpful in narrowing down what I am trying to find.
Edit: For clarity, their FastGPT is strictly optional and isn't contaminating your search results.
Where I’m having a problem is finding appropriate imagery and artwork to use in my ancient history videos. Now that so many of the results are swamped with AI generated visuals, I have to be extra careful.
I am extremely tempted to someday use AI artwork. These subjects are so esoteric and literally no imagery exists of, say, Hittite city life or reasonable estimates of how Yamnaya children looked.
But our channel is more focused on academic accuracy and scholarly consensus than most and the technology isn’t quite there yet.…
I'm extremely curious to see what AI would do if I told it to make a scene of a Neo Assyrian conquest. Would it generate the mountains of skulls, rivers of blood, and human skin flags???
But yeah, I can see that. Late Bronze Age is already not an incredibly popular topic, let alone for art.
It's an extremely popular topic for a certain subset of infotainment producers. YouTube offers me Bronze Age Collapse videos daily, despite there not being really anything new to add to the discussion for like a decade.
I think some people find the whole "Sea Peoples" thing too evocative to pass up.
Depends what's in the training data. If it doesn't have any pictures using those terms if won't do much with it. I'll try running it tonight when I get a chance and see what I get.
You also could always try describing what you want to see in detail and let it draw that even if it doesn't know what that is.
Finally got around to doing this. Ran it through stable diffusion with a few different models. Just used "Neo Assyrian conquest" as the prompt and nothing else.
https://imgur.com/a/gTSOXSc
A few models gave me mostly maps. A few others focused on this flag and tended to give vaguely history bookish images. Last bunch were from an anime model.
The anime ones came out of nowhere, and I think did psychic damage! No one's being flayed alive, so I guess this is a peaceful tribute tour.
What an odd decision, flags. I would've expected it to know the history of when they were used!
I don't know what you use to search, but you could potentially just set the search to only include things from before 1/1/2023, for example.
Kagi allows you to train its search algorithm to display less results from particular sites (or none at all, if you so wish) and more from sites that you've deemed worthy. Over time, this might make your searches a lot more efficient if they are often around similar topics.
That's one of the things with generative AI, though, isn't it? Do we want the results to show stereotypes, or reality? Maybe either is acceptable, depending on our purposes and motivations.
If no imagery exists of it, how would you asses the accuracy of the generated image?
That’s a good question. I’m basically training myself as an art historian, although I’m sure a real art historian would roll their eyes at such a claim.
But what I mean is there’s a whole host of cultural context clues with which to judge any piece. Not only the style of the art (and this is currently the biggest giveaway with AI art is that one overly airbrushed digital style), or the inaccurate hands and faces, but there are a bunch of other clues too.
For example, in my latest video on Charlemagne, there are a host of recognizable artistic masterpieces from throughout the ages illustrating his reign. I try to stick with what is known instead of new works.
Perhaps I could do a video on my channel that would be a “making of an ancient history video” episode that showed the current AI examples and what makes for a valuable image for educational and entertainment purposes.
Very impressive research and video work. I wish you luck in your efforts.
I'm in a bit of similar position as you, so I think I relate to the draw of AI as a generative tool.
If you are establishing an audience who trust you, this has got to be the most valuable type of connection to create in the future because of the floods of AI-generated images and low credibility pixel brain spam.
Consider the example of the vocoder in music. When the vocoder first became a viable production tool "instant cool robot voice makes your track unique" it was briefly a shortcut and a replacement for the other time-tested parts of musical composition that go into making a hit. When the novelty of the vocoder voice wore off we were left with a few OUTSTANDING vocoder hits and piles of less than mediocre vocoder songs, the latter being those songs that bet big on the novelty effect of the vocoder without actually imbuing their tracks with typical and timeless musical vitamins (chord changes, arrangement, song concept, lyrics).
It could just be me in my little silo here, but I think already discerning internet users (such as your audience) are relegating anything that smells like AI generated stuff into the mental junk pile, and hunt out the stuff where there's a credible human behind it, a host/creator taking their audience seriously.
AI as a content tool will fail in part because the productivity it enables is not in the interest of the recipient audiences. As a reader/audience member I'm just gonna go to the place where trust and authority has been established. Tildes is a great example of this.
I'll just suggest you hold your own credibility and authorship in the very highest esteem, and let your competitors erode their audiences trust by getting sucked up in the current AI-fad.
Or buy a vocoder if you must! :-)
Thanks for taking a look! And I very much agree with you. The community that our channel’s founder established is what makes the entire thing work, regardless of the quality of the content.
I’m kind of in a unique situation since he died and left very clear guidance of what kind of channel he wanted. I have no need to go against his deathbed wishes so the channel will continue as it always had and I will keep giving most of the profits to his widow and young daughters. This very human story makes it far more meaningful to many of the subscribers.
But I agree in a wider sense. I can learn discrete facts from a non-human source quite easily but in terms of larger works and artistic projects, I just don’t see the point in consuming synthetic media.
There was a bit of a chat on this several days ago with this post about adding a term to remove AI overview from results. Though that doesn't solve the issue of AI generated junk articles coming up in results.
One word: Kagi!
I'm on my free trial period but will be subscribing as soon as it ends. I've already saved multiple days worth of my time just with a couple searches that Google has never gotten even close to right.
Editing to add that I'm someone who normally avoids subscription fees like the plague. This is just too good to pass up, and such a healthy alternative to all the predatory Big Tech out there that it would be worth paying the monthly fee even just as an act of charity. I want people to continue having access to ad free information.
Also, I found out about Kagi from Tildes, so thank you everyone who've been mentioning it on here.
My biggest pet peeve since the LLMpocolypse is when I, or someone else asks a question on a forum, and either a bot, or even worse, a real person just copy and pasted an answer from chatGPT as a reply. Usually without any additional prompting, they just plopped my question into chatgpt. It's very easily recognizable, and usually worse than useless, because the answer is usually both plausible sounding and wrong, because of the answer could easily be given by an LLM, the person asking the question would have just asked an LLM. They're always easily recognizable, because chatgpt always gives very excessively wordy, similar sounding answers. It's just garbage noise.
I always wonder, do some of the people that do that somehow believe they're actually providing some sort of useful service, or are they all just upvote farming to sell the account later on?
As others have pointed out, I've subscribed to Kagi. I've been using it for a few months now and I couldn't be happier. Before that, I was kidding myself into thinking I was using DuckDuckGo, even though more than half the time I ended up using
!gto bring up slightly better but not great Google results.The whole shenanigans regarding the recent Google stuff made me even more confident that I've made the right choice. Kagi does have AI stuff as well, but it doesn't get in the way and is pretty intuitive.
Ooh, I feel really called out. I do this, too.
Yeah, the text search results for DDG are subpar, but for some reason their image search returns far less AI generated works than Google does. Perhaps the AI compilation websites are not hitting its SEO, since it's a more niche search engine.
For recipes where the search results are absolutely infested with them, I've pretty much given up and just lean to a handful of sites now to grab recipes from, some specific YouTube channels and relatives who were professional cooks/bakers for technique.
Back then I used site-cleaning extensions but plenty of the recipes turned out bland or mediocre even if salted to taste and require extensive adjustments. Turns out many are just recipe blogs for the sake of recipe blogs.
Now I go with a simple rule that if I can't see the ingredients list below the fold, I'm out.
So that becomes mainly BBC Good Food because they're in metric and have baking recipes by weight. I'll maybe occasionally spend the mental bandwidth to translate recipes to metric and country standard cup sizes for proven known quantities, like Chef John's stuff on Allrecipes.
I survive using Google with uBlacklist Chrome extension https://chromewebstore.google.com/detail/ublacklist/pncfbmialoiaghdehhbnbhkkgmjanfhe which allows me to easily block garbage results.
I do three things that kind of work:
A good example is my recent search for dress shoes. Just searching for dress shoes to find reputable companies nearby will not get you good results. Adding reddit as text will get you better results. Asking the AI to name a few companies gives you something more specific to search for.
But honestly, the results still suck. I’m curious what others do.
I think I've given up and am returning to real life references. I'm going through the process of buying a GPS watch and it's just a clusterfuck of information out there, even for a very niche little product. I've had more luck reaching out to a guide we used from a few seasons ago and chatting with folks I run into wearing them when I swim in the morning. I think the problems that the internet used to solve has now come full circle and is just adding to the noise.
I guess YouTube is still relatively safe from AI, you can still find tiny channels that review niche products.
I feel like most of the Youtube channels I frequent have become more popular - even little kitbashing ones with 30k subscribers - and now have sponsors. In the case of GPS watches, I'm worried that if a company sponsors a channel they will be more generous with reviews vs competing offers. Garmin sponsors almost everyone I watch for outdoor sports so I don't feel super confident in their reviews.
Is this a specific extension that keeps its own blocklist of AI sites or just a general blacklist you populate yourself?
I use uBlacklist, add sites to my personal blacklist with the little link that is displayed with the search results and subscribe to multiple blacklists that are updated on GitHub.
From other tildes questions I've gleaned some useful alternative search engines you might try:
wutsearch
qwant
mojeek/
AI does not give me good results for the things I'm searching for, so I turned it off following instructions in this Ars Technica article.
You may find this Tildes post helpful. It is about adding
&udm=14to your searches to strip AI content away.That's only going to work for so long.