I use language models almost every working day: to draft code, to explain a datasheet, to clean up a paragraph, to turn a messy error log into a plan. They are already good at some of that. They are still reliably bad at other parts, often in a confident voice.
This is not a league table. I am not going to invent a benchmark, and I will not tell you that "AI is about to replace" a profession. I want a practical split: work I am willing to start with an assistant, and work I will not trust without a check.
The companies that build these products say the same thing in their own help pages: answers can be wrong. That warning is easy to skip. It should be the first thing you remember.
What I already trust it to start
Boilerplate and structure. If I need a argparse skeleton, a CSS layout I have written twenty times, an n8n-style step list, or a first outline for an article, a model is fast. I still edit the result, but I no longer start from a blank file as often.
Explaining a thing I already have in front of me. If I paste a function, a stack trace, or a paragraph of a datasheet and ask "what is this doing, in plain language?", the answer is often a useful starting point. The important part is the paste. Without the source text, the model will happily explain a typical function that is not yours.
Translation and tone shifts. Turning a rough technical note into a calmer email, or a Lithuanian draft into careful English, is work it handles well — as long as I check that no meaning drifted, especially around numbers, warranties and safety.
Rubric work. "Here is a messy meeting note; give me decisions, open questions, and owners." "Here are five log lines; group the errors." These are compression tasks. The model is good at compressing, and I can see immediately if a point was dropped.
A second pair of eyes on a draft I wrote. "What is unclear? What did I claim without evidence?" is a use I like, because I still own the text. The assistant is a reviewer, not the author of record in my head.
Where it still fails, in ways that look like success
Invented facts, with a straight face. This is the classic failure, and it has not gone away. A model can invent a function that does not exist, a paper title, a helpful-looking URL, or a "typical" voltage for a part. OpenAI, Anthropic and Google all warn that their assistants can be inaccurate. I treat any fact I did not paste in as unverified.
I am especially careful with citations. If I need a source, I open the source. "According to a 2024 study" is not a source.
Fresh or local facts without a search tool. Training data is not a live view of your city, your repository, or this week's price list. Even with browsing, the model can summarise a page badly or miss a date. If the answer depends on "as of today," I check the primary page myself.
Arithmetic and precise quantities, unless a tool does the work. I do not let a chat "just calculate" a battery runtime, a resistor network, or an invoice total without either running the numbers or using a tool that actually computes. The prose around the number can look perfect.
Long, unstated context. If a codebase has a hidden invariant — this flag must stay in sync with a hardware pin, this email must not go out without approval — the model will not know unless you say so. It will produce a plausible refactor that breaks the invariant.
Safety-critical or irreversible actions. I do not let an assistant be the last check on a charging threshold, a payment, a public statement of fact, or a destructive command. It can draft the procedure. A person signs off.
Knowing when it is lost. The most awkward failure is not a blank reply. It is a fluent reply that skips the hard part. If I ask a vague question, I often get a vague answer dressed as a complete one. That is why a clear prompt helps, and why "say what you do not know" belongs in the instructions.
A simple rule I use at the desk
I ask myself one question before I accept a reply: can I check this, quickly, against something real?
- If yes — a compiler, a datasheet, a git diff, a calculator, an official help page — I use the draft and I check it.
- If no — a legal claim, a medical claim, a market statistic, a "study shows," a hardware rating I cannot measure today — I slow down or I throw that part away.
This is also why I like assistants inside a loop I already trust: a test suite, a linter, a dry-run flag, a staging server, a second measurement on the bench. The model proposes. The loop disposes.
What this means for how I prompt
I do not ask "is this a good idea?" if I need "what are the trade-offs, and which of them can you not see from here?" I do not ask "write the production script" if I need "write a script with --dry-run and no hard-coded secrets." I do not ask for a literature review if I am not going to open the papers.
And I do not use an assistant as a witness. "ChatGPT said so" is not a source, including when I am the one who asked.
The honest middle
The useful story is not "AI can't do anything" and it is not "AI can do my job." The useful story is narrower: these tools are already strong at drafting, structuring and explaining material you give them, and they are still weak at truth, at current facts, and at responsibility.
If you keep that split in mind, the tool gets cheaper. You stop arguing with it about things it cannot know, and you stop being surprised when a polished paragraph contains a hole.
I'm Saulius, an independent full-stack, automation and electronics developer in Kaunas, Lithuania. You can also find me on Dev.to and LinkedIn.
Top comments (1)
That split between compression tasks over pasted material and open-ended factual recall matches day-to-day engineering reality. Requiring the source text to be pasted in before asking for explanation eliminates the subtle failure mode where a model confidently describes a generic textbook version of a function instead of the actual code in front of you.
Treating arithmetic and precise quantities as mandatory tool calls rather than text generation is another boundary that holds up well. Even strong models still slip on multi-step unit conversions when left to predict digits inline.