Why AI “Fixes” Things That Aren’t Broken (and How to Prevent It)
Yesterday, I explained what to do when AI forgets about past conversations. Today, Eugene wrote in with this gripe about AI: "I hate when you, for instance, write a story or blog and you…ask the AI for stuff like grammar corrections or just some simple editing, and then it sometimes changes your story. First up, checking grammar and spelling with AI: I’ve tested various AI models extensively for grammar and spelling fixes. The results are…variable. ChatGPT-5 is clever, but the latest LLMs still struggle with specific grammar and spelling fixes. Here’s a fun example: Ernest is rolling in his grave right now. Try that test for yourself, and ChatGPT or Claude might count the letter M correctly. I found a similar errant test on X about counting Bs in the word blueberry that ChatGPT aced for me. But AI can and will still trip up. If you must use ChatGPT for proofreading jobs, a prompt like this can help: “Identify critical grammar and spelling errors in bold. Do not change any words, facts, or story elements. If you’re unsure about a correction, highlight it for my review instead of changing it. Explain your reasoning clearly.’” You can also fire up ChatGPT Canvas mode and edit your draft in real-time with AI. It struggles with specific grammar fixes because of pattern-matching bias and an overreliance on tokenization. In short, it’s either trained itself on incorrect answers or misreads the arrangement of letters and numbers. Earlier this year, I compared ChatGPT to Grammarly. The result? A dedicated grammar checker has a far better hit rate than a default AI chat. You can get a Grammarly discount here. For longer content, I’d still hire a human editor. Second — preventing AI from changing your content: If you care about your authentic content or stories like Eugene, it’s yours too. I do, however, use AI to provide editorial feedback. I give AI a set of talking points to evaluate my content and drafts. I’ll craft a specific prompt like: “You are my editorial assistant. Propose edits and suggestions for each of these talking points: what can I improve, what’s clear, what’s unclear, suggest a metaphor, etc. Do not change the core meaning or facts. The ideal output looks like…” I’ll ask for a few variations, and then I’ll act on whatever feedback is appropriate. AI is clever, but I don’t count on it for verbatim rewrites for anything critical or personal. I’m in charge, not the machines! For the next few days, I’m answering your questions about AI. Write in and ask, or email me your gripe! I’ll write a reply. If you need more help using AI, check out Prompt Writing Studio. |