General-purpose chatbots like ChatGPT are genuinely useful for parts of a law practice — and genuinely dangerous for others. The line is not "AI good" or "AI bad." It's about which tier you use, what you put in, and whether you verify what comes out.

What's reasonably safe

Non-confidential drafting scaffolds. Outlining an argument, brainstorming headings, or rewriting your own prose for tone — with no client confidences in the prompt.
Explaining a concept to yourself. Using it as a study aid, then confirming against primary authority.
Boilerplate you'll fully rewrite. A generic clause skeleton you replace with real, jurisdiction-specific language.

What crosses the line

Pasting client confidences into the free consumer tier. The consumer tier may use inputs to train. That's a Rule 1.6 confidentiality problem unless you have informed consent and the right contract. See our checklist on AI and client data.
Filing its citations without checking them. The 2023 sanction of two New York lawyers for citing fictional cases generated by ChatGPT is the cautionary tale. General chatbots hallucinate citations that look perfect and don't exist.
Treating output as authority. It predicts plausible text; it is not a research database.

The two fixes

First, control the data path: use a tier with a contractual no-training posture, not the free consumer one. Second, validate every citation against a real database before it leaves your office. A purpose-built legal tool does both by design.

How Draftiro is different

Draftiro runs on the paid Gemini tier — your data never trains AI models — validates every case citation against the public CourtListener corpus, flags statutes and rules as "unchecked," and appends an AI disclaimer to every assistant message. It's the same generative capability, inside the guardrails a law practice needs. Compare it to Clio Work or see the practice workspaces.