AI bot FAQ

Chat bots predict likely text. When they don't know, a likely-sounding answer still comes out. It is worst for exact numbers, quotes, citations, recent events, and anything about your own accounts that the bot hasn't actually opened.

Ask for the source link and open it. Search the page for the exact sentence. If the bot can't name where a number came from, treat it as unverified. For actions ("I posted it", "it's deployed"), ask for the URL, ID or screenshot and check it yourself.

It helps a little, it is not a guarantee. Better: tell it to write "unknown" when it has no source, and to separate what it checked from what it assumes.

Any of the big ones (ChatGPT, Claude, Gemini) writes well. Pick the one whose default tone you edit least. Test the same prompt in two free tiers before paying.

Use one that searches the web and shows sources (Perplexity, or ChatGPT/Claude/Gemini with search on). Then open the sources. A bot with no search only knows its training data.

Coding tools that see your files (Cursor, GitHub Copilot, Claude Code, Codex) beat pasting snippets into chat, because they can read the rest of the project and run tests.

An agent with connectors and an approval step, like Grok Bot. Turn on approvals for anything sent, posted, bought or deleted.

Start new chats for new topics (long chats re-send history), don't use the biggest model or deep research for simple questions, batch small questions into one message, and list and switch off scheduled jobs you don't need.

Claude: Settings > Usage. GitHub Copilot: your billing/usage page. Cursor and Grok Bot: the usage page or ask the bot to check your plan usage. ChatGPT and Gemini show a notice when you hit a limit.

Usually no. Anthropic says payments are generally non-refundable; Google gives no refunds for partial billing periods except where law requires. Usage spent on a wrong answer or a failed run is spent.

Upload only what you'd be fine with the company storing. Remove account numbers, IDs and other people's private data first. Check the bot's data controls (for example the training opt-out in settings).

No. Never paste passwords, card numbers, or one-time codes into a chat. Use the app's official connect/sign-in flow, which uses permission tokens you can revoke.

Use the official connector, grant the smallest scope it offers, test with one harmless task, and revoke access from your Google account's third-party access page when you stop using it.

Chats have a context limit; old messages fall out. Put standing facts in custom instructions / profile preferences / project instructions, or in a project file, instead of repeating them.

Standing text added to every chat: who you are, how you want answers. ChatGPT calls it Custom Instructions, Claude has profile preferences and project instructions, Gemini has Gems.

Generation is random by design. If you need consistency, give an exact format, an example output, and fixed rules.

As long as needed to state the task, the audience, the format and the check. One clear paragraph usually beats one vague line.

Only if that bot has browsing on and the site lets it in. Many sites block bots. If it can't load the page, it may guess. Ask it to quote the page to prove it read it.

No. It mostly buys more usage and access to bigger models. Bigger models make fewer mistakes, not zero.

A chat bot answers. An agent takes steps: opens sites, runs code, sends things. Agents need approvals and checking because their mistakes are actions, not just words.

Yes, several can (ChatGPT scheduled tasks on paid plans, Grok Bot routines). Each run uses your allowance even when nothing happens, so audit them weekly.

Paste 2-3 samples of your writing and say "match this tone, sentence length and vocabulary". Save that as a project instruction.

It has no intent, but it can confidently state false things and claim work is done when it isn't. Treat status reports as claims until you see proof.

Cancel from where you subscribed (the website billing page, or the App Store / Google Play if you bought in-app). Anthropic says cancel at least 24 hours before renewal.