The best chatbot for a quick answer may be the wrong one for a folder of work documents. A tool can give you tidy source links but struggle with a long PDF. Another can accept the PDF yet make it hard to check where an answer came from. And a generous upload limit says little about what happens to the file after you send it.

Start with the work you’ll actually do. If you mainly ask questions about recent events, put source quality first. If you review contracts, reports, or spreadsheets, test file handling before paying. If you expect to build a useful archive of conversations, check how you can export it. Privacy runs through all three choices: the account, plan, and settings matter as much as the chatbot’s name.

Illustrative image: how to choose an AI chatbot

Decide what you need to control

A privacy setting can answer several different questions. Will your chats be used to train models? Will they stay in your history? Can the service remember details across conversations? How long might a temporary chat be retained? Turning off one feature doesn’t necessarily turn off the others.

For ordinary personal use, look for a model-training control and a way to delete or export saved chats. For confidential work, first check your employer’s policy and the protections of the specific work account. A personal subscription is not a substitute for an approved business workspace. Even with training switched off, avoid uploading passwords, customer records, or a full document when a short, redacted excerpt will do. If email is the task, organizing it without sharing sensitive information calls for tighter limits than drafting a generic message.

Then separate three kinds of “sources”:

  • A web citation points to an outside page the chatbot used or found.
  • A file reference points back to material you uploaded or connected.
  • A related link may help you check an answer without being the source used to produce it.

None is proof on its own. Open the link, find the specific claim, and check the date and context. For file work, ask for page numbers or quoted passages, then inspect the original. This matters especially when you’re using AI to check your understanding of a difficult reading: a plausible explanation can still miss what the author said.

Finally, treat upload limits as several limits, not one headline number. Check file size, number of files per prompt, daily or rolling upload quotas, supported formats, and how much text the chatbot can process at once. A file that uploads successfully may still be too long for a reliable answer to a question about a detail buried near the end.

The options, grouped by what you need

All five services below offer ways to start without choosing a paid individual plan. Paid plans are worth considering only after a real task exposes a limit you need to remove. Features and allowances change, so check the plan and account you intend to use before subscribing.

For one general-purpose workspace: ChatGPT

ChatGPT is a sensible starting point if you want writing help, file analysis, and web search in the same place. It can search the web and provide source links, but search isn’t the same as every answer being researched. Ask it to search when freshness matters, and inspect the links behind consequential claims.

For a personal account, Data controls let you turn off “Improve the model for everyone” while keeping chats in your history. Temporary chats stay out of history and aren’t used to improve OpenAI models while they remain temporary, though they may be retained for up to 30 days for safety. That’s useful for a one-off question, not a reason to upload material you aren’t permitted to share.

The catch is that a large maximum file size doesn’t mean unlimited file use. ChatGPT’s document limit is 512 MB per file, with a separate two-million-token cap for text and documents; spreadsheets have a lower size limit, and free users have a daily upload cap. If repeat document work is your reason to pay, test both your file type and how often you hit that cap.

For long-document work: Claude

Claude is worth trying when your main job is reading, drafting from, and asking follow-up questions about documents. Its chat uploads allow up to 20 files per chat and 500 MB per file, but projects have a different, smaller per-file limit. PDF handling has another distinction: Claude can analyze visual elements in PDFs of up to 100 pages; for longer PDFs within its supported range, it processes text rather than those visuals. A scanned chart deserves a direct test before you rely on the answer.

Claude also offers web search with citations when search is used. That makes it a stronger fit for someone who alternates between their own documents and outside research, but don’t mistake a web citation for a page reference to an uploaded file.

On consumer accounts, you can turn off the setting that allows new chats to be used for future model training. Incognito chats aren’t saved to chat history or used for training, but they have a retention period and can’t be reopened after you close them. The practical catch: save work you need before ending an incognito session, or use a regular chat with the privacy settings you’ve checked.

For Google-centered work: Gemini

Gemini makes the most sense if your work already lives in Google’s tools and you want an easy route from an answer to a document. You can export an individual response to Google Docs, while Google Takeout offers a separate route for downloading account activity. Those solve different problems: one saves a usable draft; the other helps you keep a record.

Its upload rules allow up to 10 files in one prompt, subject to availability, and up to 100 MB for each supported nonvideo file. Rolling usage and processing limits still apply. For a long report, test whether Gemini can answer questions from the beginning, middle, and end rather than judging it by a summary alone.

Gemini’s privacy trade-off needs attention before you connect personal material. With Keep Activity on, chats and uploads are saved in Activity and may be used to improve Google services, including AI models. Turning it off changes what is saved and used for model training, but some connected-app features become unavailable; future chats are still kept for 72 hours for specified service and safety purposes. Gemini sometimes shows sources or related links, but not for every response. Check whether a link supports the precise sentence you plan to use.

For questions led by web sources: Perplexity

Choose Perplexity if most sessions begin with “What do the sources say?” rather than “Help me develop this draft.” Its search product is built around answers with visible citations, which makes it convenient for finding pages to inspect. The catch is that citation-heavy formatting can make a weak source look reassuring. Check the original publication, not just the answer.

The free tier suits occasional searches; its paid individual tiers offer more access to research tools, advanced models, and file-upload sessions. The plan comparison describes some individual upload allowances as limited or variable rather than giving a simple number to budget around. If files are central to your work, check the allowance shown in your account and trial the task before paying. Perplexity provides a self-service account data export, useful if you want to keep your research trail.

For Microsoft-centered work: Copilot

Copilot is the option to examine if your main task happens alongside Microsoft’s apps, but first identify which Copilot you mean. The consumer Copilot app and Copilot used through a work or school account have different data protections and administrative controls; don’t apply a personal-account description to an employer’s workspace.

For the updated consumer Copilot app, Microsoft says prompts, responses, and file contents aren’t used to train foundation models. A signed-in personal-account user can export Copilot app activity history as a CSV file through Microsoft’s privacy dashboard. That’s useful for keeping a record, though a CSV is less convenient than a finished document if you want to reuse a polished answer. Test the exact file and app workflow you intend to use rather than assuming every Copilot surface behaves alike.

Run a short trial before you commit

You can learn more in an hour with your own tasks than from a long feature checklist. Use material you’re allowed to share, and give two or three candidates the same tests:

  1. Check a sourced answer. Ask a question whose answer you can verify. Require links, open two of them, and see whether they support the exact claims made.
  2. Check a document answer. Upload a representative file and ask about a specific detail near its end. Check the cited passage or page yourself. Repeat with a table or image if those matter to you.
  3. Check the controls. Find the training, history, memory, and deletion settings. Note what changes when you turn a setting off.
  4. Take your work out. Save one useful answer in your preferred format, then locate the full account export. Don’t assume those are the same feature.

Pick the tool that passes your hardest routine task, not the one with the longest list of features. For many people, that means one general-purpose chatbot and a second, source-led tool for research. If neither handles a sensitive document within the permissions you have, the practical choice is to leave that document out of a consumer chatbot.