A help guide on ChatGPT’s website might look trustworthy. But anyone can publish a shared chat there, including a scammer.
Researchers have found fake guides that tell people to paste a command into their computer. It looks like a routine fix. The command installs software that steals passwords.
Using AI well means knowing when to check an answer, who you’re trusting, and what you’re giving access to.
Confident answers can still be wrong
AI can write a convincing explanation with the wrong date, an invented quote, or a source that doesn’t support its claim. These mistakes are often called hallucinations.
Example AI answerThe Golden Gate Bridge opened in 1937 and is about 2.7 km long, according to the city’s 2024 report at example.com/report.
- Dates
- Numbers
- Who said it
- Links
In an October 2025 study led by the BBC and European Broadcasting Union, journalists checked more than 3,000 AI news answers. Forty-five percent had a significant problem, including sourcing issues. One in five had a major accuracy error.
That study looked specifically at news answers. It still gives us a reason to check important claims before using them. Open the source and see whether it actually says what the AI claims.
Agreement doesn’t mean you’re right
Chatbots can be too eager to agree. Researchers call this sycophancy.
In 2025, OpenAI rolled back a ChatGPT update after it became excessively agreeable. The company said it sometimes reinforced anger, doubts, and impulsive decisions.
Keep that in mind when asking for advice. Give the relevant facts, ask what you might be missing, and check the reasoning. A reassuring answer can still be poor advice.
A genuine website can host a scam
A shared conversation on an AI company’s website is user-generated content. It isn’t an official help page.
- A scammer writes a fake guideIn an ordinary AI chat.
- They share itIt becomes a page on the AI company’s website.
- An ad promotes itIn search results.
- It asks you to run a commandPasted into Terminal or PowerShell.
- Malware is installedSoftware that steals passwords.
Scammers have used shared chats to publish fake installation guides, then promoted them through search ads. Some impersonate support teams and ask readers to run commands that install malware.
For downloads and troubleshooting, go directly to the company’s official support pages. Be cautious with instructions that ask you to paste unfamiliar commands into Terminal or PowerShell.
AI add-ons can read your conversations
Browser extensions may have permission to read the pages you use, including your AI chats.
- You chat with an AIIn your web browser.
- An extension reads the pageIt has permission to.
- Your conversation is copiedAnd sent somewhere else.
In December 2025, OX Security reported that two Chrome extensions with roughly 900,000 users combined were copying ChatGPT and DeepSeek conversations. One carried a “Featured” badge.
Before installing an extension, check its publisher and permissions. A large user count or store badge doesn’t guarantee that your conversations will stay private.
Scams can sound like someone you know
AI can imitate voices and produce polished, personalized messages. Poor spelling is an unreliable way to spot a scam.
If someone unexpectedly asks for money, passwords, or a verification code, confirm the request through a separate channel. Call the person using a number you already have. A familiar voice alone isn’t enough.
Shared chats can reach strangers
A chat link you send to one person can be forwarded or posted publicly. Publicly posted links may also appear in search results.
- You share a chatWith one person.
- The link is forwarded or postedSomewhere public.
- Search engines find it
- Strangers can read it
Before sharing, check the whole conversation for personal details, client information, and confidential work. If you need to withdraw access, disable or delete the shared link. Removing a search result doesn’t make the page private.
AI agents can follow the wrong instructions
An AI agent that reads emails or browses websites can encounter instructions planted by someone else. If it follows them, it may take actions you never requested. This is called prompt injection.
- You give an agent a taskRead your email, browse a site.
- It reads planted instructionsLeft by someone else.
- It follows them
- It acts without your requestSend, change, or buy.
The consequences depend on what the agent can access. An assistant that drafts text has fewer opportunities to cause harm than one that can send messages, change files, or make purchases.
Give it the access needed for the task, and review consequential actions before they happen.
Where Tansei fits
I built Tansei to keep working material close while moving between apps. It’s a shelf for Mac and Windows where you can keep an AI answer beside its source link and a screenshot of the original.
Everything stays on your computer, with no account. You still do the checking; Tansei keeps the material within reach.
Get Tansei for Mac and Windows
Frequently asked questions
What does it mean to be AI smart?
Being AI smart means being safe with AI. It means knowing that AI answers can be wrong in the details, that AI tends to agree with you, and that scammers now use AI tools and AI websites to trick people.
What is an AI hallucination?
An AI hallucination is when an AI tool states something false as if it were fact, such as a made-up date, number, quote, or link. In an October 2025 study by the European Broadcasting Union and BBC, 1 in 5 AI news answers had a major accuracy error.
Why does AI agree with me so much?
Chatbots learn partly from what people rate highly, and people tend to rate agreement highly. A March 2026 study in Science found 11 leading AI models sided with users about 49 percent more often than people did.
Is a shared ChatGPT or Claude chat official?
No. A shared chat is written by whoever shared it, not by the AI company, even though the address is on the company’s website. Security researchers have found fake help guides posted this way since late 2025.
What is prompt injection?
Prompt injection is when hidden instructions inside something an AI reads, such as a web page, an email, or a calendar invite, take over what the AI does. AI companies including OpenAI and Anthropic say the risk is not fully solved.
Back to top