Ask ChatGPT how to check for malware on your Mac, and it will almost certainly tell you to open Terminal. Not System Settings. Not Activity Monitor. Terminal.
This pattern plays out across every AI chatbot, for nearly every troubleshooting question. Mac blogger Howard Oakley at The Eclectic Light Company decided to investigate why, and his findings say a lot about a fundamental blind spot in how LLMs work.
The Training Data Problem
The core issue is straightforward: LLMs process text tokens. They learned from text. And describing how to click through a graphical interface in text is surprisingly hard.
Think about it - writing "click the Apple menu, then System Settings, then General, then Software Update" requires careful step-by-step verbalization of something most people do on autopilot. Meanwhile, a single terminal command like softwareupdate --list is already text. It's compact, unambiguous, and perfectly suited to how language models represent information.
This bias gets worse as more instructional content moves to YouTube. Video tutorials showing GUI workflows generate zero usable training text for LLMs. The written blogs that used to describe these steps are declining. So the training data skews even further toward command-line solutions that were already overrepresented.
ChatGPT's Malware Advice: A Case Study
Oakley put ChatGPT's terminal-first instinct to a practical test by asking it to help detect malware on a Mac. The AI recommended five separate procedures. The results were rough:
One command that was supposed to verify XProtect (Apple's built-in malware scanner) was up to date instead produced 6,528 lines of output with no version information anywhere in it. Other suggestions were either flat-out incorrect for current macOS versions or pointed users toward checks that wouldn't actually detect anything.
None of the five procedures worked as described. A knowledgeable Mac user would have gotten better results from opening System Settings and checking two screens.
The Real Risk
This is not just an inconvenience. Terminal commands bypass the safety guardrails that Apple builds into its graphical apps. When someone copies and pastes a command they don't understand into Terminal, they're running it with whatever permissions their account has - no confirmation dialogs, no undo button, no "are you sure?" prompt.
This is exactly how clipboard-hijacking malware works: trick users into pasting a malicious command. AI chatbots are training millions of non-technical users to do precisely this - paste opaque strings into Terminal and hit Enter.
The deeper problem is that following terminal instructions teaches users nothing. You don't build a mental model of how your system works by copying defaults read com.apple.SomeFramework. You just get a wall of text you can't interpret and a false sense that you've accomplished something.
For anyone using AI assistants for tech support: if the answer is always "open Terminal," that should be a signal to question the advice, not follow it. The GUI solution might take more clicks, but at least the AI would have to describe something you can actually see and understand.