I promised you I’d share some of the ways I’ve been using ChatGPT. As I’ve found out more about using the tool, I’ve noticed a few things that seem to crop up regularly, no matter what you are using it for. I thought these might be worth listing.
I’m going to offer the following disclaimer throughout this series. Although I’m talking about ChatGPT, most of my observations are also true for Gemini, Perplexity, Claude and other leading-edge AIs. I suspect that over time their capabilities will become increasingly similar. Whether that’s good or bad, I don’t know.
Observation 1: ChatGPT has become much better at understanding what you mean, not just what you say.
When I first started using ChatGPT several years ago I found it very useful for helping me clean up manuscripts and offer suggestions for improvements – now it’s infinitely smarter.
Here are four changes I’ve noticed.
Better reasoning. It can handle more complicated questions, compare alternatives and revise an answer when you point out a problem.
Better conversational awareness. You can say, “No, that’s not quite what I meant,” or “make it less formal,” without having to start again from scratch. It is much more capable of treating the exchange as an ongoing conversation.
Better understanding of context. If you’re working on something over several conversations — a trip, an article, a novel or a purchase — ChatGPT may be able to carry relevant information from earlier discussions into later ones. That can occasionally be a little unnerving. You might ask for restaurant recommendations on a trip, for example, and find that ChatGPT takes into account preferences you’ve mentioned in an earlier conversation.
Better responsiveness to nuance. Modern ChatGPT is better at recognizing the difference between, say, “give me information about this” and “help me decide how to approach this.” It can often infer the purpose behind a question.
The early versions of ChatGPT often felt like an extraordinarily knowledgeable person who wasn’t always listening very carefully. Today’s version feels much more as though it understands what you’re trying to accomplish.
Observation 2: ChatGPT can be surprisingly insightful — but you get more of that insight when you ask it to think with you, rather than simply answer questions.
ChatGPT does not have intuition or lived experience in the human sense. What it can do extremely well is synthesize large amounts of learned patterns and apply them to the situation you describe. That can produce answers that feel surprisingly perceptive.
For example, it can often help by:
spotting an assumption you hadn’t realized you were making;
identifying the trade-off hidden inside a decision;
suggesting a different way to frame a problem;
connecting two ideas that you had been treating separately;
noticing inconsistencies in a plan or draft;
asking, in effect, “Have you considered this angle?”
How do you harness that?
The answer may be: don’t use ChatGPT only for questions with known answers. Give it problems to think about.
Instead of:
What are the advantages of X?
you might ask:
Here’s what I’m trying to accomplish, and here are the constraints. What am I overlooking?
Or:
Challenge my thinking on this.
Or:
What assumptions am I making?
Or:
Give me three different ways of looking at this problem.
Or even:
Based on everything I’ve told you, what stands out to you?
Observation 3: ChatGPT can be a very persuasive co-pilot. The trick is not to let it become the driver.
The same sensitivity that makes ChatGPT seem insightful can also make it over-interpret what you say. A slight emphasis, an example you happen to mention, or the way you frame a problem can cause it to infer a preference or direction that you didn’t really intend.
There’s another danger here. If you reveal what you’re hoping the answer will be, ChatGPT may sometimes follow your lead rather than challenge it.
In other words, ChatGPT is very good at following signals — but sometimes it follows them too enthusiastically.
AI can take a hint and run too far with it.
For example, if you say:
I’m leaning toward option A, but I’m not sure.
ChatGPT may start helping you justify option A, when what you really wanted was an impartial comparison.
Or if you say:
I’d like this article to be funny.
it may begin pushing every suggestion toward humor, even if you only wanted a light touch.
That leads to a useful lesson: you have to stay in the driver’s seat. ChatGPT can suggest, interpret, and elaborate, but you should periodically check whether the conversation is still heading where you want it to go.
A few habits help:
Say explicitly when you want neutrality: “Don’t assume I prefer either option.”
Correct direction early: “You’re taking this further than I intended.”
Ask for alternatives: “What would the opposite approach look like?”
Separate brainstorming from decision-making: “Give me possibilities, but don’t recommend one yet.”
Occasionally restate the goal: “Remember, my main objective here is…”
Observation 4: Tell ChatGPT what perspective you want, not just what role to play.
Specifying a role can help — but usually less than people think, especially with modern ChatGPT.
For instance if you paste a set of medical notes into the prompt box and you simply say,
Interpret these medical notes for a layperson.
ChatGPT will already infer that it should explain the medical language clearly and probably organize the important findings. You usually do not need to preface it with a specialization, for example
You are an oncologist.
Where role instructions become useful is when you want a particular perspective or standard of analysis. For example,
Interpret these notes for a layperson. Focus especially on anything an oncologist would consider important, and separate routine findings from anything that might need follow-up.
is more specific than merely assigning the role. The value comes from telling ChatGPT what lens to use, not from pretending it literally is that professional.
Similarly, instead of saying,
You are a financial advisor.
try:
Look at this the way a retirement planner would. What risks or assumptions might I be overlooking?
Framing the question really can steer the answer, so it’s worth being deliberate about the lens you give it.
The Better It Gets, the More Judgment Matters
What I’ve gradually realized is that getting good results from ChatGPT isn’t mainly about learning clever prompts. It’s about learning how to have a productive conversation with it.
Give it enough information to understand what you’re trying to accomplish. Ask it to challenge you, not merely agree with you. And pay attention when the conversation begins heading somewhere you hadn’t intended.
ChatGPT may be getting smarter, but that doesn’t make our judgment any less important. If anything, it makes it more important.