Most people use ChatGPT as a search engine that talks. That is the single job it is worst at, and it explains almost every disappointing experience people report.
It is excellent at transforming text you provide — summarising, rewriting, translating, explaining, drafting, restructuring. It is unreliable at recalling facts from memory. Move every task you can from the second category into the first, and the quality of what you get back changes completely.
The one distinction that matters
There are two kinds of request, and they behave nothing alike.
Transformation — you supply the material, it changes the form. "Summarise this contract." "Rewrite this email to be shorter." "Explain this paragraph simply." The information is in front of it, so it cannot invent.
Retrieval — you ask it to produce facts from memory. "What does the law say about X?" "Who wrote the paper on Y?" Nothing constrains the answer, and it will produce a confident, well-formed, possibly fictional one.
Almost every complaint about it being "wrong" comes from the second category. Almost every genuinely impressive result comes from the first.
What it does well
Summarising long documents you paste in. Reliable, because the source is present.
Rewriting — shorter, simpler, more formal, different audience. Its strongest single skill.
Explaining a concept at whatever level you ask for. "Explain this like I know nothing about finance" works well.
Translating, including into and out of Arabic, French, Spanish and German. Check specialised terminology.
Structuring — turning messy notes into an outline, or a wall of text into a table.
First drafts of anything you will edit anyway. Emails, descriptions, outlines, scripts.
Finding the flaw — "what is wrong with this argument", "what did I miss in this plan". Genuinely useful and underused.
Format conversion — prose into bullets, a list into a table, notes into an agenda.
What to avoid
Anything where being wrong is expensive — medical, legal, financial or tax decisions, unless a qualified human checks the output.
Recent events, unless the web search feature is actually on. Training stops at a fixed point.
Exact arithmetic on anything consequential. It approximates.
Real citations. It is notorious for producing plausible references to papers that do not exist. Every one needs checking.
Anything confidential. Client data, credentials, unpublished work.
The tone is identical whether it knows or is inventing — because to the model, confidence is a stylistic feature, not a signal of certainty. There is no wording that makes it admit uncertainty reliably. Every number, name, date and citation needs verifying before you act on it.
Five habits that change the output
1. Paste the material. "Summarise this: [text]" beats "what does that article say" by an enormous margin. This one habit accounts for most of the improvement available.
2. Say what you want back. Length, format, audience, tone. "Explain inflation" gives you an encyclopedia entry; "explain inflation in four sentences to a 15-year-old" gives you something usable.
3. Correct rather than restart. "Too long, halve it." "The third point is wrong." Faster than rewriting the prompt, and it keeps what already worked.
4. Ask for the reasoning on anything with steps. It improves accuracy and lets you see where it went wrong.
5. Start a new chat when the topic changes. Long conversations drift, because earlier material falls outside what it can consider at once.
More on this in how to write better AI prompts.
Free or paid
Free is enough for occasional use, learning and short tasks.
Paid is worth it if you use it daily for real work — stronger models, higher limits, file uploads and web search, which is the feature that most changes reliability on factual questions.
Subscribe for a month before committing to a year. This field moves fast enough that long commitments rarely pay.
The realistic summary
It is not a search engine, not an oracle, and not a replacement for knowing your subject. It is a very capable text-transformation tool that happens to answer questions — and answering questions is its weakest mode.
Used for what it is good at, it saves real hours. Used as a source of truth, it will eventually hand you something confident and wrong, and you will not be able to tell from the output which one you got.
See also: what a large language model actually does.
Frequently asked questions
What is ChatGPT actually good at?
Transforming text you give it — summarising, rewriting, translating, restructuring, explaining and drafting. It is far weaker at recalling facts from memory, which is where the confident errors come from.
Can I trust what it tells me?
Trust the shape, verify the specifics. Structure, phrasing and explanation are usually sound. Numbers, names, dates, citations and quotes need checking every time, because it produces them with identical confidence whether they are real or invented.
Is the paid version worth it?
If you use it daily for real work, usually yes — better models, higher limits, file handling and web search. For occasional use the free tier is generally enough.
Why does it make things up?
Because it is built to produce likely text, not true text. When it lacks a fact, nothing inside it stops, so it generates whatever most resembles a correct answer. This is called hallucination and it cannot be fully prompted away.
Should I paste confidential information into it?
Treat anything you type as potentially readable. Settings differ between personal and business accounts. Client data, credentials and anything under an NDA do not belong in a chat box.
Found an error? Email us and we will fix it and note the change at the bottom of this article. Hello@daily-atlas.com
