When ChatGPT is the wrong tool

ChatGPT is a general assistant: one text box that can draft, explain, plan, read a file, search, and write code. That range is the point. It also means that for any single job, there is probably a tool built specifically for it, and that tool will feel better at the edges.

This page is not a benchmark. We do not publish scores we cannot reproduce, and model versions change faster than any table survives. What follows is a job-based map: where a chat assistant is awkward, what kind of tool fits better, and how to test the swap without wasting a month.

chatpt is a typo, not a competitor

There is no ChatPT assistant to compare against. The comparison is between ChatGPT and the tools you might use for a specific job. If you want the naming question answered first, read what chatpt means.

The six jobs where chat is the awkward choice

1. Reading a very long document, repeatedly

Chat is built around a conversation. A 300-page report is not a conversation; it is a reference you return to.

What fits better: a tool with a document workspace, persistent files, and citations back to the page. Several assistants now offer a project-style container; document tools and PDF readers add their own AI features. The feature to look for is not “can it read a PDF” but “can it show me the page a claim came from”.

Test it: upload the same 80-page document to two tools and ask for the same five questions, including one whose answer is on page 61. Check the citation, not the prose.

2. Facts that must be current

A chat model answers from training plus whatever it can look up in the moment. If the answer depends on today’s price, today’s news, or a page that changed this morning, you need a tool that always searches and always shows sources.

What fits better: an answer engine that cites every claim, or a search tool with an AI summary. ChatGPT has a search mode; the question is whether it actually ran for your query.

Test it: ask a question whose answer changed this week. If every claim has a link you can open, the tool is doing the job. If the links are decorative, it is not.

3. Code that lives in a repository

Pasting one function into a chat works. Pasting a repository does not. The context that matters — imports, types, tests, the file that calls the function — is exactly what gets left out.

What fits better: an assistant inside the editor, with the repo indexed and the diff in view. These tools see the same code you do.

Test it: give both the same bug and the failing test. Compare the patch, not the explanation. A patch that breaks two other tests is worse than no patch.

4. Data you cannot send to a vendor

Sometimes the answer is not a better cloud tool. It is a model that runs on your machine, or a vendor with a contract your legal team accepts.

What fits better: an open-weight model running locally, or an enterprise deployment with the paperwork you need. Local models are slower and smaller; for narrow tasks that is often fine.

Test it: run your three most common prompts on a laptop-class model. If quality is enough for those three, you have a private option for the sensitive 20% of your work.

5. Image and video work with a specific look

Chat assistants generate images, but they are designed to be general. If you need a consistent character, a brand style, or a video sequence, a dedicated tool gives you controls a chat box does not.

What fits better: an image or video tool with reference images, seeds, and style controls. Use the chat assistant for the brief and the iteration notes; use the specialist for the render.

Test it: generate the same subject five times. Consistency is the thing to measure, not the best single image.

6. Anything that must work offline

No cloud assistant works offline. If the job happens on a plane, in a basement, or on a network you do not trust, the answer is a local model or a non-AI tool.

Test it: turn off Wi-Fi and try the task. If the tool fails, it was never an option.

These six are tool-fit problems. For the costs that follow you after you pick a tool — confident wrongness, data leakage, flattened voice, invisible labour — read the downsides guide.

The names you will hear

You do not need a scorecard to notice the pattern. The main alternatives fall into four groups:

• General assistants from other labs — Claude (Anthropic), Gemini (Google), Copilot (Microsoft). Same category as ChatGPT, different defaults, different app integrations.
• Answer engines — Perplexity and similar tools that always search and always show sources. Better when the question is factual and current.
• Tools built into the work — editor assistants for code, document tools for long PDFs, image tools for visual work. They see the context a chat window has to be told.
• Local, open-weight models — Llama, Qwen, Mistral and others you run yourself. The option when the data cannot leave the machine.

We do not publish model rankings. The right question is not “which is smartest” but “which one removes the wall I keep hitting”, and that is a question about your work, not a leaderboard.

The switching-cost trap

People change tools for the wrong reason: a single bad answer, a demo that looked good, or a recommendation from someone with a different job.

The cost of switching is not the subscription. It is the prompts you have tuned, the files you have organised, the habits you have built, and the second place you now have to check.

A rule that works: do not switch for a feature. Switch for a repeated wall. If the same kind of task fails three times in a month, and a specialist tool removes that specific wall, switch for that task only. Keep the generalist for everything else.

Test a replacement in twenty minutes

1. Pick five prompts you actually ran in the last month.
2. Pick one document you actually use.
3. Run both tools with the same inputs. No prompt tuning on the second one; that biases the test.
4. Score only three things: did it finish the task, did it cite what it claimed, and how long did you spend fixing it.
5. Read the outputs the next morning. Fresh eyes catch the confident wrongness.

If the new tool does not win on at least three of the five, keep what you have.

What ChatGPT is still the best at

• Being one place. One subscription, one thread history, one set of custom instructions, one app on every device.
• General drafting and rewriting. The most common knowledge-work task, done well enough.
• Explaining at a chosen level. “Explain this to a smart 14-year-old” remains a killer feature.
• Switching between modes. A question, a file, an image, a search, and a draft in the same conversation.

The specialist tools win at the edges. The generalist wins at the centre, which is where most people actually work.

Related

• Free, or paid? Decide by what you actually do — the same decision method, applied to plans.
• The office workflows worth automating first — where a generalist is enough.
• A prompt is a brief, not a wish — the skill that transfers between tools.
• Open the workbench — get the brief right before you compare tools.