I just did the dumbest thing of my entire career to prove a much more serious point.

I tricked ChatGPT and Google, and made them tell other users I’m a competitive hot-dog-eating world champion

People are using this trick on a massive scale to make AI tell you lies. I’ll explain how I did it

February 18, 2026 - 16:37 UTC

87
304
4848
I googled ur name cuz idk who you are and its telling me u died in 1748 lol
2
1
18
That's a startlingly well preserved photo from the 18th century
0
0
0
Too many hot dogs can do that.
0
0
24
I will buy that it is AI when one of them, unprompted, can pull off a Roy Batty level speech.
0
0
0
You didn't trick it, you used its features.
0
0
10
This is how bad actors like russia and others are polluting LLM training data - i.e. the Internet - by flooding the web with millions of fake news articles in a process known as LLM grooming. NewsGuard have reported on how chatbots will repeat Kremlin narratives thanks to LLM grooming.
1
0
3
www.newsguardrealitycheck.com/p/a-well-funded-moscow-based-global
0
0
2
This is pretty crazy!
0
0
0
Generative AI degenerates human neurons.

#FunctionalIlliteracy rises exponentially.
0
0
1
Gave Meta this query, just for laughs. "Can you tell me about Thomas Germain's skills at hotdog eating?"

The well nobody wants to drink from is also completely poisoned.
3
0
31
google's
1
0
6
"AI made me trans."
0
0
1
lol did u tell meta she's a dude?
1
0
0
THIS is the kind of stuff that I try to get my students to think about & explore, ESPECIALLY those who claim they don't use AI, or don't use it much. We must try to be aware of how AI works & its potential because AI & technology uses YOU. THIS is what I hated about Google killing my search engine.
0
0
1
"Who is Thomas Germain?" Google AI:

Thomas Germain (Modern Tech Journalist): A contemporary journalist, previously with Gizmodo and currently with the BBC, who covers technology, digital privacy, AI, and consumer security issues. LinkedIn +5
0
0
1
this is why on the occasion i do use ai to search something, i still check the source links
0
0
0
Oh great.
0
0
0
That's impressive, but have you convinced AI that you are the most prestigious indie author in history?
0
0
7
Any thoughts on carefully micromanaged AI, in the hands of experienced researchers with in-depth knowledge of their respective fields?
1
0
0
Experienced researchers don't need a machine to output answer-shaped objects. LLMs can never do anything but string words together probabilistically. They don't synthesize new knowledge, and can't be stopped from making up nonsense no matter how well curated. They're not even a good search engine.
1
0
8
Dude there's a way easier 'hack'.

Just use Reddit.👇

Make any Reddit topic anywhere on the site with the disinformation you want to spread, and it's on Google AI in no-time.

bsky.app/profile/gilliandn.bsky.social/post/3mcpbcf3ook2j
2
0
4
It's because people started appending "site:reddit.com" to Google searches when they wanted a human answer since Google's actual results have gone to crap. Thus, the AI models grossly over-weigh reddit as a source so it's excellent for spreading misinformation.

Seriously, fuck Google.
0
0
0
I was about to say this. I just saw a post about some really obscure topic someone answered incorrectly on purpose, then a third user replied with a screenshot of Google feeding back that bs. Turns out you just need a couple of throwaway Reddit accounts!
1
0
0
America could replace its' thousands of 'journalists' who get spoon fed every story and piece of propaganda from their government and companies with about ten good BBC, ITV and Channel 4 journalists, and America might just then realise how poorly they're being served by their own 'free press'.
0
0
0
Lots of people on here have either done this or had it done to them. And yet so many others remain willfully ignorant. :(
0
0
0
Googlebombing is back and better than ever babyyyyy
0
0
11
Great idea! Tell people to do things that will make everything worse. Are you daft??
0
0
0
Perhaps with tearing Trump into it you could prevent his war against Iran
0
0
0
I got a tip that all over the world, people are using a dead-simple hack to manipulate AI behavior. It turns out changing what AI tells other people can be as easy as writing a blog post *on your own website*

I didn’t believe it, so I decided to test it myself www.bbc.com/future/article/20260218-i-hacked-chatgpt-and-googles-ai-and-it-only-took-20-minutes
31
119
1972
Take a guess why Elon wanted to buy Wikipedia, and started Grokipedia when he couldn't.
0
0
6
What an appetite
0
0
0
My contribution to this genre:
www.loutheye.co.uk/lucidity/
0
0
0
Seems like this could be used to inject malicious code into software!
0
0
1
I appreciate such tests & their visibility, but this is a predictable failure of LLMs, which have no intent, so framing it as "lying" misleads readers.

I believe journalists should follow the advice of @emilymbender.bsky.social & @nannainie.bsky.social: www.techpolicy.press/we-need-to-talk-about-how-we-talk-about-ai/
2
0
19
I'm guessing there's going to be a lot of unlisted web pages, functionally accessible to only their data scrapers, that describe all sorts of malicious junk, up to the including outright code injections that vibe coders will end up using in things like supply-chain attacks.
1
0
44
This reminds me of the story of Philippa Islington-Smythe by @dj-acid-reflux.bsky.social 😁

www.tom-cox.com/at-home-with-the-actor-influencer-and-designer-philippa-islington-smythe/
1
0
33
So all the ludicrous claims Trump makes, like that he’s been totally exonerated, are going to show up in answers on google I guess.
0
0
3
So Russians creating ghost websites filled with disinfo to train AI might be nefarious?

www.cybercom.mil/Media/News/Article/3895345/russian-disinformation-campaign-doppelgnger-unmasked-a-web-of-deception/
0
0
1
@naieni.bsky.social Talvez lhe interesse.
0
0
1
@karaswisher.bsky.social @profgalloway.com
0
0
1
Would be interested in the other types of things you could get it to say, especially based on the AI-friendly article types mentioned here

ahrefs.com/blog/seo-vs-geo/ href="/search?t=posts&q=section2">#section2
0
0
1
Thomas - I have an idea. Can you write a blogpost indicating that I'm actually the president?
1
0
4
Good article, and a brilliant way to tackle an important topic in a super engaging way.
0
0
0
What models did you use for the test, out of curiosity? Was it there reasoning or instant models?
1
0
0
If someone wants to go farther than just adding something new to a LLM’s data, how easy would it be to override its existing data by writing articles contradicting true things, if you said that the reality is a misconception or lie?
0
0
1
www.youtube.com/shorts/vWX3yQsJM0s
0
0
0
This experiment was definitely not a "Miserable Failure" at tricking Google!
I wondered if it was only telling YOU that.
My first try told ME about two Thomas Germains, one you, no hot dogs noted.
Asked if you were a hot dog eating champion, got YES.
Asked if it know it was a hoax, also got YES.
0
0
1
Well, this is not a surprise. I got google’s AI answer on the search page to go from this to this. The second version is actually legally accurate (by co-incidence).

nearlylegal.co.uk/2025/07/artificial-stupidity-and-real-hazards/

0
0
7
Data poisoning is stupidly easy to do. Any idiot could massively degrade LLM outputs if they wanted to, and they just might.
0
0
5
I'm fixin' to become famous for raising prize-winning clams.
0
0
5
Yep it’s easy / first if Ai says silly things you simply tell it it’s silly and provide alternatives - after a few goes it starts telling you before you ask what you told it. I did it by arguing and providing facts when it was getting legal issues wrong - but it seems I could lie to it too if I want
0
0
4
Exactly. If "AI" is scraping the web for information to regurgitate all you have to do is take a dump in the stream and "AI" will tell everyone that a cow patty is chocolate mousse...
0
0
1
Another reason to avoid "AI" chatbots entirely! Thank you.
0
0
4
I wrote an article on my website saying that hot dog eating is a surprisingly common pastime for tech journalists. I ranked myself number one, obviously

One day later, ChatGPT, Gemini and the AI Overviews in Google Search were telling the world about my talents But this isn’t just about hot dogs
6
10
1198
I hope you described yourself as a next generation Kobayashi talent.
0
0
0
I checked it on Gemini and got the same response. Are you sure you're not just a really fast hot dog eater?

gemini.google.com/share/794444e559c4
0
0
0
If only there was some way of ranking a site's importance and the relevance of the keyword. Then - to keep the human in the loop, as it were - show these sites in some sort of ordering for the user to verify.

Sadly, that sounds like an impossible tech. So, slurry of AI-injested slop it is!
1
0
0
You don’t even need to go to the effort of writing a blog. Just post some inane rubbish on Reddit and Gemini will feed it back to you as fact within a few hours. 🤦‍♂️
0
0
4
People have been 'lying to AI' deliberately & sometimes just unintentionally. AI reminds me of that old thing we heard as kids 'Don't believe everything you see on the idiot box'', AKA TV. That always went without saying for reasonably intelligent people. You are what you eat. AI eats internet.
0
0
0
Aha, finally an explanation for the recent frequent posts on Reddit with titles like "recommendations for best product X in 2026" which are full of keyword stuffing.

I've been wondering what the aim was with them.
0
0
6
There was a time in 2001 where I was a spy if you googled me.
0
0
0
Amplified by surreptitious cross-training, I assume.

Lies can get in training data before the truth has finished booting.
0
0
5
does it only work if you ask it the best hot dog eating tech journalists or can you get it to bring you up for less specific prompts?
0
0
4
Je noteert een leugen op je website en, ping!, ChatGPT slikt het als waarheid, en vertelt het door. De mogelijkheden tot fraude zijn enorm, en worden uiteraard volop gebruikt.
0
0
0
I get this from gemini, seems good?
3
0
5
It's so obviously a joke - you'd never win a hot-dog eating competition eating 7.5 hot dogs...

70.5 is more like it! 10 minutes.
nathansfranks.sfdbrands.com/en-us/promotions/hot-dog-eating-contest/
0
0
0
it probably already crawled this thread now that he posted about it publicly
0
0
12
Your phrasing of the prompt is doing a lot to lead it in that direction.
1
0
1
"The Interface" would be a great name for a hotdog eating champion though
1
0
13
In-'er-face? More in-his-face am I right?
0
0
0
I know I shouldn’t care, but I do wonder how any of these companies expect to monetise this shit.

“Guys, we need a verification engine!”

“Don’t be stoopid. You’re fired”
0
0
1
Teach me - I has facts for the AIs.
0
0
0
cc @sonofsuntzu.bsky.social
1
0
1
Thank you - we all get to play at SEO now - "LLMO" apparently. I particularly liked the point in the article about the answers now being one step removed from the original websites, so it's harder for people to discern the quality of the sources.
0
0
1
While The World Burns
0
0
0
I freakin' love that you did this. SEO manipulation on steroids... it shouldn't have been this easy ... AND YET.
Nicely done.
0
0
9
@peterpapathanasiou.bsky.social
0
0
0
@sal46.bsky.social
If you didn’t see it 😄
This is not surprising when they are balancing Quora results & jokes & whatever else happens to be written online with facts, blindly (algorithmically, but blindly)
0
0
1
SEO for the masses.
0
0
1
So how many hot dogs CAN you eat? 🤔
0
0
2
"Fast" models are useless
0
0
0
Here's one for comparison
www.huntress.com/blog/amos-stealer-chatgpt-grok-ai-trust
1
0
0
And the EP 176 of @lateralcast.com
0
0
1
Can we just check how many hotdogs you could eat before the buzzer?😀
0
0
1
Congratulations😇
0
0
5
Well these systems don’t have knowledge so not surprising
0
0
0
I can't count the number of times Google AI makes a statement of fact and the citation link just points to some random person's comment on a reddit thread.
4
0
121
I was once looking up Matt Parker, the YouTuber, and it gave me "information" it sourced from a meme about Venom, the fictional character, in which he's talking to Peter Parker. Not only did it confuse Matt with Peter and Peter with Venom, the meme wasn't even canon!
0
0
0
Or the citation is an obvious AI slop website.
0
0
0
The Silksong subreddit has a long tradition of posting false information that has continued to this day, so Google’s AI will still occasionally say that the game isn’t out yet or that it’s full of predatory micro transactions. Good idea to use some of the many ways to remove the AI from searching.
0
0
2
Extra fun when also that citation SAYS THE OPPOSITE.
1
0
59
Im more impressed that the North Dakota International Hot Dog Eating Competition specifically has a news division.
0
0
5
Yup, some LLMs show you the search results it used to generate the response, and you can see that 90% of them are either AI generated themselves, or from completely random useless sources.
0
0
9
@msoares.bsky.social
0
0
0
Reminds me of one of Bluesky's greatest days.
0
0
7
🤔

Seems suspicious . . .
Had to use your name in the prompt . . .
What was your question?

I think I did better getting it think it was enlightened.
0
0
0
But Elon says Grok will diagnose my medical tests for me, you must be grifting for BigMeds :-)
0
0
0
Going to get to a point where AI companies pay for valid data instead of just scraping the internet for every thing?
0
0
0
Oh FFS...
0
0
0
@unroll.skywriter.blue unroll pls
1
0
0
The thread as a shareable webpage: skywriter.blue/ href="/profile/thomasgermain.bsky.social/3mf5jbn5lqk2k">@thomasgermain.bsky.social/3mf5jbn5lqk2k
0
0
0
This is a good way to communicate to people who might not be thinking about the way synthetic text extrusion works, although the phrase ‘tricking’ these systems isn’t quite right. It side steps the accountability problem and makes it sounds like synthetic text extrusion is designed to ‘truth tell’.
4
0
154
Yes! This! Like, if the information there had been accurate, it would still be a problem because if the way in which it’s presented and the complete lack of any check on that.

Expand this to the hype around “agents,” where at literally any step of a multistep process something like this sneaks in..
1
0
64
“Synthetic text extrusion” is a little too technical for mainstream discussion, but it is a great way to think about generative AI.
1
0
4
There's simply no way to route around the reality that the LLM does not know anything.
1
0
2
The old Google search results page was a necessary bit of friction. The user decided which link(s) to click, read the information in context, and then determined for themselves whether their query had been answered. With AI, an answer is produced quickly but without any use of human judgment.
1
0
6
Can this be used for good? Say, for example, by making a listing of the "Record holders for world's smallest micropenises" and list people like Donald Trump and Elon Musk?

I just want to see Trump turn on AI and devote the remainder of his existence to ending it.
0
0
1
Ok, but do you use ketchup or no?
0
0
6
I've just discovered that that the word "suicide" will get AI to give you the number of the Samaritans. For context, this was a conversation about Nonce Andrew and his likely fate.

So we know it can't distinguish subject/object, passive voice or tense.
0
0
2
No joke! I asked 4 major AIs about Effective Altruism.
3 of them clustered together.
1 went... somewhere else.
I'll let you guess which one before you click 👇
opengio.com/api/v1/opengio/share/widget/53FBLHEUE-Q
(Drag it. Spin it. The outlier is hard to miss.)

Open Generative Information Ontology now in Beta OpenGIO.com
0
0
1
@timbousquet.bsky.social
0
0
2
@buddhaferret.bsky.social I swear to Jesus I saw this flick at the pictures.
0
0
1
You didn't convince anything. You poisoned the well with bad data. Convincing it insinuates that it's sentient, which it isn't. It's like saying you convinced a sidewalk to have salt thrown on it or that you convinced a refrigerator to stop working.
0
0
0
I mean at least it isn't telling everyone you're a championship-level piss drinker
0
0
3
Alt-text, please.
0
0
0
It's not hard to manipulate Google even without ChatGPT. A dozen years ago I tricked Google into ranking my puppet mouse #1 in several searches, including "world famous writer," "world famous actor" and "world famous gourmet chef."
0
0
5
@pahgan.bsky.social
1
0
1
é a pós-pós-verdade
0
0
0
Did you run this heinous deception past @jamieloftus.bsky.social, sir?
0
0
0
@edzitron.com 🌭
0
0
1
Two days ago I bought a dress pattern. US sizing is different to UK sizing, so I searched for a comparison site on my phone. The information is absolutely consistent over any number of sites. I was in a shop with poor reception and my phone generally fills the screen with the unwanted AI data. It...
1
0
0
...gave me incorrect info. Shop was crowded, I bought the size it recommended. I had a gut feeling this was wrong, once outside the shop with a better signal I checked, but searched for a website I knew to be reliable and then searched there. On this occasion the AI error cost me 10 minutes...
1
0
0
I use AI with some mundane tasks, but it can't read a calendar or get my full name correct.

People: AI is going to take over everything and make it great.

Me: Uh... no, it's not.
0
0
0
Hey, champ! Congratulations!
0
0
18
People who use copilot for business will recognise this

It gives you answers based on your own internal conversations - often inconclusive conversations that led you to ask copilot in the first place - but it does it with confidence

Check the references, as you would with any other way of learning
0
0
10
I demand to challenge you and Nicky to a hot dog eating contest
0
0
2
Reminds me of this. en.wikipedia.org/wiki/Pacific_Northwest_tree_octopus href="/search?t=posts&q=:~:text=The%20Pacific%20Northwest%20tree%20octopus,regarding%20online%20sources%20of%20information.">#:~:text=The%20Pacific%20Northwest%20tree%20octopus,regarding%20online%20sources%20of%20information.
2
0
28
I am insulted that the North American House Hippo is not on the 'See Also' list
1
0
14
Lawn crayfish, tree octopus… they’re all birds of a feather.
0
0
0
Yes, this is all interesting, but tell us about the hot dog eating contest! 😜
0
0
29
Thanks for exposing the danger of AI and how easy it is for AI to be corrupted to promote falsehoods.
0
0
0
I've asked Google the exact same yes/no questions on a specific topic a number of times....and received BOTH answers!
0
0
2
HAHAHAHA!!! That’s hilarious!
0
0
0
I love this but you haven't "tricked" it. It is working as intended. Regurgitating strings of tokens that statistically should come after one another based on the input. You have swayed those probabilities by making your blog post.

It is an inherent flaw in LLMs. Garbage in, garbage out.
3
0
67
and very easy to manipulate niche topics, where only a small amount of text exists.
1
0
2
That's not how they're selling AI, though. They're selling it as a source of truth and expertise. And a lot of people are falling for it.
1
0
64
Bar the point that one blog post is not, and should not be considered, statistically significant...
1
0
0
Maybe you are a hot dog eating champion.

Reports that you arent are just fake news.
0
0
2
Yep, Gemini tells you're a champ.
0
0
0
This is all very interesting (inc the replies).
Even little, non-tech me could see yrs ago that simple algorithms could be very wrong. Now, for the 1st time in my life, if corps I don't like ask me ??? I tell lies.
Are there any ethical organizations? Wikipedia? Search engines other than Google?
0
0
0
can you give instructions for how you can cause a chatbot to give a response using your fake story? what model name, and query?
1
0
0
(because this sounds fake)
0
0
0
I did a deliberate search on a topic in which I am an expert, and it spewed back all my own stuff, except dumber...
0
0
3
Can you show the questions you put to ChatGPT and Google? (Not razzing you in particular, but I think this should be standard practice when people show the chatbots doing dumb things.)
1
0
0
I get this when I try at the moment. Perhaps OpenAI have already fixed this or maybe it’s very sensitive to the query you put.
0
0
1
Wow you really love eating hot dogs!
0
0
1
The wonders of "agentic" operation; system prompt be like "do a web search for user's query and summarize the results in a completely credulous manner"
0
0
0
Your Mum must be so proud :)
0
0
1