ChatGPT

Where does your data go?

Your conversations go to OpenAI's cloud (US-based). On Free, Go, and Plus, OpenAI trains its models on your chats by default — you must opt out in Settings → Data Controls → "Improve the model for everyone." Business and Enterprise plans don't train on your data; consumer plans do. Deletion requests are honored within roughly 30 days.

What happens when it's wrong?

The pattern across published benchmarks: the rate depends entirely on the task.

• Summarizing a document you hand it — very good. Vectara's leaderboard: ~9% hallucination rate for GPT-5.5-class models in 2026.
• Facts from memory, no web search — bad. OpenAI's own system card: 47% hallucination rate on SimpleQA without browsing. Roughly a coin flip between a right answer and a confident invention.
• Citations and niche specifics — the danger zone. A 2026 Stanford benchmark found GPT-5.1 fabricating 6.57% of legal citations — five times the rate of the year-older model. Newer is not automatically safer.
• Browsing helps but doesn't fix it. The app browses by default, dropping real-world claim errors to roughly 9.6%.

What's it actually bad at?

• Guessing instead of admitting ignorance. On hard factual questions, GPT-5 guessed rather than said "I don't know" the overwhelming majority of the time. It will rarely tell you it's unsure.
• Anything where being wrong is expensive — legal, medical, tax. Published error rates in those domains are the highest measured.
• Math and precise counting. A long-standing weakness across the model family.
• The free tier runs the budget engine. It's the smallest model in the current family — unlimited chats, but Plus gets a meaningfully stronger model.

What does it cost really?

The trap: jumping to Plus because of message limits is mostly obsolete — text is unlimited on Free now. What you're actually paying for at Plus is the stronger model and no ads. Pay for the model, not the volume.

What does it replace — and what do you still do?

Realistically kills 70–80% of first-draft work: emails, posts, outlines, summaries, meeting notes, brainstorming. You still own: final review, factual verification, and a human pass on anything customer-facing. It does not replace your judgment on numbers, contracts, or regulated content.

What the ecosystem offers

• Projects — folders that keep their memory. Group your chats, files, and instructions for one client or job, and you never re-explain yourself. Every plan has them, and you can share them with teammates.
• Custom GPTs — a version of ChatGPT you set up once with your rules and files, then reuse for the same kind of job over and over.
• Connectors — plug in your Slack, Outlook, Teams, Dropbox, or SharePoint and ChatGPT can work with what's already there. It can also read your company's own tools as research sources.
• Agents — set up once, then they do a recurring job on a schedule: weekly reports, meeting prep, finished documents and spreadsheets. Business plans can share agents across the team.
• The coding agent (Codex) — included in the business plans, if you have developers.
• Admin controls and data guarantees — the business tiers add admin dashboards and the no-training guarantee. The security details live in the sources below.
• The newest model (Sep 2026) — OpenAI's flagship reached the tools this month, but access is rationed: the $200 plan gets the most, business plans get a slice, and the $20 Plus plan got none at launch. Free users get neither flagship model. Practical takeaway: check which tier has which model before you pay.

If I stop paying, what do I lose?

• Export: yes, self-serve — Settings → Data Controls → Export. ZIP of your history, up to 7 days to arrive, download link expires in 24 hours.
• Gotcha: no self-serve export on Business/Enterprise workspaces — the plans most likely to hold important data are the ones where leaving is hardest.
• Cancellation: self-serve — downgrade to Free, keep your history.
• Account deletion: yes — data removed within ~30 days.
• The training caveat: everything typed before opting out may already be in training data. Opt out on day one, not month six.

Who is this NOT for?

Anyone feeding it customer records, patient data, or anything regulated — on consumer plans that data trains their models by default (use the Business tier or a private setup instead). Also not for tasks where an unverified wrong answer costs money: legal documents, tax filings, published citations.