I recently had a moment where I realized I needed an AI filter. Not a filter that blurs fine lines and flaws, but a filter that spotlights errors and brings them to the forefront of any response.
Not many people know this, but ChatGPT is wrong 1 in 4 of the time. The AI summaries we often rely on for summarization are also not always accurate.
For a long time, I thought that to get more accurate answers, I needed to ask better questions. For me, that meant assuming that if the answer was wrong, the correction would be a sharper prompt, more detail, or clearer instructions. I treated ChatGPT, Gemini, and Claude like machines that needed the right input to produce the right output.
What most people (including me) get wrong
I found myself making the same mistakes over and over again. I didn’t blindly trust the AI, but if the response sounded confident, organized, and well-written, I usually accepted it. It seemed okay so I moved on.
But being “good enough” can cause a lot of problems, especially in the workplace. For example, a neat-looking timeline can ignore real constraints, or a clear-cut explanation can obscure important facts.
None of these seem catastrophic at the moment, but small failures can add up. If you’re using AI to increase productivity, the last thing you want is to end up with more work because the AI messed up.
Most ChatGPT users treat AI a bit like a magical search engine. They ask questions, get answers and move on. Although this approach works in most cases, it is flawed in more complex cases.
AI is optimized for confidence and fluency, not attention. You can easily:
- fill in the missing context
- make reasonable assumptions
- Smoothly overcome uncertainty
- Skip the inference step
But the problem is that while the answer is clear, it’s not always reliable. The problem isn’t that AI is “bad.” That is, we often read it passively.
That’s what I wanted to change. My little filter (the one that changed everything)
Now, whenever I receive an AI response, I pause for about 10-15 seconds and run through three simple questions in my head.
- What does this assume?
- What could I be missing?
- What must be true for this to be false?
- Do I need to check sources or facts?
To be clear, I don’t always enter these into ChatGPT. This is not a prompt, but a mental filter that you apply after receiving the answer.
Sometimes the answer appears. In some cases, weaknesses may become apparent right away. Either way, I’m no longer passively consuming AI, I’m intentionally evaluating it.
why this works so well
This little habit does three things:
- There is a limit to how slow it becomes. Don’t overthink it. Enough to avoid rubber stamping a sophisticated response.
- Hidden assumptions come to the surface. AI often assumes things you haven’t said (deadlines, budgets, priorities, constraints, etc.). My filter makes me aware of that.
- Move AI from oracle to thinking partner. Instead of asking, “Is this true?” we’re asking, “Under what conditions is this true?”
As AI becomes more and more integrated into our workflows, it’s important to have this filter, or one like it, in your mental tools. Filters don’t make AI smarter; they help it understand better.
real world example
I now naturally use this filter every time I use a chatbot. For example, I recently asked ChatGPT to help me map out a multi-week timeline when planning a big project. The first version was beautifully composed. In fact, it was so beautifully constructed that I knew it was missing something.
So I ran a filter and asked myself: What does this assume?
As it turns out, there are a lot of them. Everything from the chatbot didn’t take into account how easy it would be to schedule a meeting, how quickly everyone would approve my ideas, whether I had a dedicated team, etc.
After noticing these deficiencies in the response, I adjusted the constraints. The revised plan is now much more realistic and much easier to use. Thanks to filters, there is no need to build on unstable ground.
Similarly, I asked Gemini to explain a concept in simple English. The explanation was clear, but I felt it was a little too neat.
My filter worked: What could I be missing?
This question made me realize a useful but technically misleading simplification. I followed up and got a more nuanced version and actually learned a lot more in the process.
The filter did not prove the answer “wrong”. It’s now more reliable.
Finally, I tested it with Claude to conclude the paragraph. It suggests shorter sentences and a cleaner structure, all reasonable. However, my filter made me ask the following question. What must be true for this to be false?
I realized that clarity must always trump audio in editing. That wasn’t my goal. We kept the structural improvements but restored some of the tone.
Instead of blindly accepting the AI’s edits, we cooperated with it.
conclusion
This filter can be used with any AI or chatbot. No new tools, configuration, or subscriptions required. Try this the next time you get an important answer from an AI. Read the response once as usual, ask yourself what’s missing, and prompt based on what you notice.
You don’t have to do this every time. Only do it if you’re working on a complex project or if the response feels too general or completely inaccurate. And keep in mind that this filter has its limits. AI errors cannot be ruled out. Even if you end up using a prompt like “Cite your source,” you still need to use critical thinking and fact-checking.
This little mental filter is now part of every AI conversation I have, and it’s making the biggest difference in my results.
to follow Tom’s Guide to Google News and Add us as a preferred source Get the latest news, analysis, and reviews in your feed.
