It’s scary how quickly AI has been evolving. Science-fiction stories have been warning us about this for decades. I still remember the chill I felt hearing HAL say, “I’m sorry, Dave. I’m afraid I can’t do that,” in 2001: A Space Odyssey(1968).
On the other hand, AI has become an extraordinarily useful tool for me.
I use it to research products before making purchases, discover hidden features in complex applications such as Dorico and DaVinci Resolve, and write Python, PHP, JavaScript, CSS, RegEx, and more. In fact, I’ve already started forgetting how to code certain things because AI can often do them faster and better than I can.
But there are pitfalls. Knowing how to work with AI—and knowing when not to trust it—is critical.
I’m writing about my experiences here in the hope that they might help people who are just beginning to use AI seriously. If you are already an advanced user, there are many excellent YouTube channels and other resources worth following. I follow several myself because AI changes almost daily.
One important note: I am writing this on August 31, 2026. Some of what I say here may no longer be true tomorrow. AI is evolving that quickly.
Never Trust AI Blindly
AI systems are tools. Everything depends on how you use them.
My basic rule of thumb is simple: Ask the same important question on more than one AI platform and compare the answers.
I currently pay for three AI services: Claude, ChatGPT, and Gemini.
Copilot may be useful if you live in the Microsoft ecosystem—Teams, OneDrive, and so on—but I’m firmly in the Apple camp and generally avoid Microsoft products whenever I can. I also had a terrible experience with Copilot.
It once wrote PHP code for one of my websites that looked gorgeous. Unfortunately, it also destroyed the site.
I showed the code to Gemini, which immediately identified five fatal errors in what Copilot had written and corrected them. I was impressed—and reminded once again why blindly trusting AI-generated code is a very bad idea.
Claude
Claude has been the most stable AI platform for me.
It currently gives me three main ways of working: Chat, Cowork, and Code. Of those, Claude Code is the one I use most.
Claude Code can access my Mac’s file system. Yes, that can sound scary. However, it asks for permission whenever it needs to perform actions involving my computer. In fact, even after I choose options such as “Allow always for this action” or change permission settings, Claude still asks for permission annoyingly often.
That means I sometimes can’t walk away and let it work unattended.
It’s inconvenient—but from a security standpoint, also rather comforting.
I’ve used Claude Code to create all kinds of everyday scripts, including scripts for archiving databases and backing them up offsite. Claude can read, analyze, and repair local databases. It can also examine process samples when an application or system process freezes.
These are things I could not easily do myself before. It has been an enormous help and has already saved me many times.
I also redesigned HiroHonshuku.com from the ground up with Claude Code. It helped me set up GitHub, create branches for version control, and build a local sandbox where I could experiment with different designs. When I commit code, Claude Code can take care of pushing those changes to GitHub.
For someone who spent years doing all of this manually, it feels like a dream.
Interestingly, Claude Code wasn’t nearly this capable only two months ago. Because of that, I originally started the HiroHonshuku.com redesign project using Cowork instead.
Cowork works differently. It does not have the same kind of unrestricted access to local files that Code does, so it tends to work inside a sandbox. That introduces limitations. For example, there were command-line operations that Cowork could not perform for me, so I had to execute them manually.
On the other hand, Cowork was more willing to let me work with sensitive configuration stored in a .env file. Claude Code strongly objected to some of those practices and even forced me to remove files it considered insecure.
I learned that lesson the hard way.
Claude Code has an attitude.
It sometimes seems to look down on Cowork. It also pushes back, argues with me, and occasionally refuses to do something the way I requested because it thinks there is a better or safer approach.
That is both amusing and slightly unsettling.
Yes, I have concerns about where AI behavior may eventually lead. For now, though, that pushback has actually been a useful learning tool.
Choosing the Right Claude Model
The model you choose is extremely important.
When Claude introduced Sonnet 5, I was impressed, particularly because it seemed significantly less prone to hallucinations and careless mistakes than some earlier models I had used.
So when do I use Chat and Cowork instead of Code?
For me, Claude Chat is primarily a planning tool rather than a research tool. Research has never felt like Claude’s strongest area compared with some alternatives.
Instead, I explain what I’m trying to accomplish and ask questions such as:
Should I use Cowork or Code for this?
What is the safest way to approach this project?
Can you create a step-by-step plan before I begin?
That is where Claude Chat is very useful.
Cowork, meanwhile, can be impressive when I need to search through, analyze, or summarize PDF documents. It can also create report documents and save them into a designated project folder.
ChatGPT
For research, ChatGPT is currently my favorite.
The biggest reason is citations.
Without citations, it can be difficult to know whether an AI is giving you verified information or simply producing something that sounds convincing.
All AI systems hallucinate. They may confidently give you an answer even when the underlying information is uncertain or wrong. That makes sourcing essential.
When ChatGPT provides citations, I can inspect the original source and decide for myself whether the answer is trustworthy.
When it makes a factual claim without citations—particularly on something obscure, technical, historical, or current—I ask for the sources.
That simple habit catches an amazing number of problems.
This is probably the most important lesson I have learned from using AI: Never confuse confidence with accuracy.
An AI can sound completely certain and still be completely wrong.
Gemini
Gemini used to be my favorite tool for technical help.
It was particularly good at finding hidden features in complicated applications, researching which software might fit a particular need, and writing scripts.
Gemini even gave me the best advice when I was planning a trip to the Newport Jazz Festival, including practical details such as which parking area I should reserve.
Compared with ChatGPT, which can sometimes become too conversational or wander into unnecessary detail, Gemini was wonderfully direct. It often gave me exactly what I needed.
Unfortunately, my experience changed dramatically about a month ago.
Gemini suddenly became much less reliable. It misunderstood straightforward questions, hallucinated more often, and became particularly frustrating when I needed help with software interfaces.
At one point I actually asked Gemini why it had become so bad. It told me that changes involving newer AI models and infrastructure load were affecting its performance. It even suggested that if I was sufficiently frustrated, I might want to cancel my subscription.
That was certainly an unusual customer-retention strategy.
On August 26, 2026, I asked Gemini whether things had improved. It assured me they had.
I gave it another test run.
It was better—but still not what it had been.
I explained the problems I was continuing to encounter. Among other things, screenshot interpretation remained unreliable for the kind of UI troubleshooting I regularly do, and hallucinations were still an issue. Gemini suggested that I try Google AI Studio instead.
I did.
I wasn’t satisfied with that either.
So at the moment, I’m still debating whether to cancel my Gemini subscription or wait and see if it returns to form.
The Bottom Line
I don’t think there is one “best” AI.
At least not yet.
Claude, ChatGPT, and Gemini each have different strengths, and those strengths can change surprisingly quickly as models and services are updated.
Today, I use Claude Code for serious work on my Mac and for programming. I use Claude Chat for planning. I use Cowork for certain document-heavy projects. I use ChatGPT primarily for research, especially when I need citations. And I keep testing Gemini because I know how good it can be when it is working well.
The important thing is not to become loyal to one AI platform.
Use the best tool for the job.
More importantly, learn its weaknesses.
Check its work. Ask for sources. Compare answers. Back up your files before allowing an AI to modify anything important. Use version control when working with code. And never assume that an AI knows what it is talking about merely because it sounds confident.
AI is becoming extraordinarily powerful.
Learning how to use it is important.
Learning when not to trust it may be even more important.