Three Years of AI: Five Things I Wish I'd Known on Day One

Three Years of AI: Five Things I Wish I'd Known on Day One

I started with ChatGPT in February 2023. Since then, I've used most of the serious models that came after — Claude, Gemini, Copilot, and others — almost every day, across writing, research, coding, and decision-making.

Three years in, here's what nobody tells you: AI doesn't save time. It moves it. The first draft is faster. Deciding what's true, what's useful, and what's yours — that got harder.

This is what I'd tell anyone starting out today.

1. The early wow was misleading

The first time a model wrote a polished email in five seconds, it felt revolutionary. For a while, that was the story everyone told.

Then I tried to depend on it. Models invented citations. They forgot what they had just said. They wrote fluent code that didn't run. They delivered the right answer and the wrong answer with the same level of confidence.

Demos are not workflows. The novelty wore off the moment real work was on the line.

2. No single model is best at everything

One thing using several models in parallel teaches you quickly: they are not interchangeable. Each has a personality, a set of strengths, and a set of failure modes.

One is sharper at structured reasoning. Another is better at long, careful writing. A third is faster but more confident than it should be. A fourth integrates tightly with the tools you already use, which sometimes matters more than raw capability.

The lesson isn't to pick a winner. It's to stop expecting one model to be the right tool for every job. The serious users I know rotate between models the same way a craftsman rotates between tools — deliberately, by task.

And that leads to a more important point: the skill isn't in the prompt. It's in the context you give the model. These systems have no memory of who you are, what you're working on, or what good looks like in your field — unless you tell them. Vague input produces vague output. Sharp input produces sharp output. The bottleneck isn't the model. It's the clarity of your own thinking.

3. Where AI genuinely helps

After three years, this is where I find AI most valuable, regardless of which model I'm using:

• Getting past the blank page. The first draft is the hardest part of any writing. AI removes that friction almost entirely.
• Translating between formats. Notes into a memo. Memo into slides. Specification into test cases. The mechanical reshaping of information is exactly what these models do well.
• Thinking out loud. Ask it to argue against your position, list counterexamples, or stress-test your logic. It's a tireless sparring partner.
• Routine code. Boilerplate, parsers, scaffolding, schema migrations. Fast, usually right, and easy to verify.

Where it isn't good: anything that requires being right when nobody else will check.

4. Where AI is dangerous

A real example. Early on, I asked a model for a list of references on a research topic. It produced six — author names, journal titles, page numbers, all impeccably formatted. Three of them didn't exist. Invented authors. Invented papers. Same fluent confidence as the real ones.

This is what people mean by "hallucination." It's not a bug being fixed. It's a property of how these models work — they are trained to produce plausible language, not correct information. When you ask for a fact, the model generates a fluent guess.

Newer models do this less. Connecting them to real documents reduces it further. But the risk doesn't disappear — it shifts from obvious wrongness to subtle wrongness. And subtle wrongness is more dangerous, because it survives a quick read.

After three years, my single most reliable habit is this: I never trust an AI answer on a fact that matters until I've verified it independently. Names, numbers, dates, citations, technical specifications. Every one gets checked outside the model. This applies equally to every model I use.

5. The five lessons that changed how I work

If I compress three years into the things that actually changed my behaviour:

• AI is amplification, not substitution. It makes good thinkers faster. It makes weak thinking sound more confident.
• Structuring problems beats writing prompts. There is no shortcut around clear thinking.
• Verification is part of the work now. If your workflow doesn't have a checking step, you don't have a workflow — you have a risk.
• Domain knowledge matters more, not less. AI raises the floor for everyone. It raises the ceiling far higher for people who already know their field.
• The first draft is never the final answer. The moment you accept it as final, you've outsourced judgment to a system that doesn't have any.

Three years of daily use, across multiple models, have left me with one durable conclusion.

AI hasn't made my work easier. It's made my judgment matter more.

That's the real shift. The hype cycle will keep rising and falling. The fundamentals won't.

If you've been using AI heavily — what's the one habit that's saved you, and what's the one mistake you've stopped making?