In the LLM article I explained why language models can be confidently wrong: they generate plausible text rather than retrieving verified facts. This article is the practical follow up, a working system for using AI heavily without ever being the person who sent the client a made up figure.
Know the four failure modes
Almost every AI error I have seen in the wild is one of these:
- Stale knowledge. The model's training has a cutoff, so recent prices, regulations, people in roles, and product versions may be out of date
- Hallucination. A fabricated detail delivered fluently: a citation that does not exist, a statistic with no source, a feature a product does not have
- Misreading your intent. The model answered the question it thought you asked. Technically fine output, wrong job
- Plausible arithmetic. Sums and percentages that look right at a glance and are not. Always the most embarrassing one to ship
Naming these matters because each has a different fix, and none of them is "stop using AI".
The habits that catch them
Feed it the source material. The single best protection. If the answer should come from your contract, your data, or your email thread, paste it in. Grounded answers fail far less than remembered ones.
Make it search and cite. For anything current, factual, or important, have Claude search the web and show its sources, then click the ones that matter. A claim with a checkable citation is a different species from a claim without one.
Ask for its uncertainty. "What in this answer are you least sure about, and what would you check?" is absurdly effective. Models are surprisingly good at flagging their own weak points when invited to.
Re run the numbers. Any figure that leaves your hands gets recalculated, by a spreadsheet, by asking Claude to compute it step by step and show working, or by a second pass in a fresh chat. Thirty seconds, every time.
Use a second pass for high stakes work. Paste the finished draft into a new conversation with "act as a sceptical reviewer, find the errors and unsupported claims." Fresh context, no attachment to the draft, ruthless results.
Scale scrutiny to stakes
Verifying everything equally is how people burn out on the tool and quit. Instead, match the checking to the blast radius:
- Low stakes, brainstorms, internal drafts, summaries for your own use: read it, use it, move on
- Medium stakes, client emails, quotes, reports: check names, dates, numbers, and commitments, the nouns that hurt
- High stakes, legal, medical, financial, safety, anything public: full verification, cited sources, second pass review, and a qualified human sign off where the domain demands one
This is exactly how you treat human work, by the way. You do not fact check a colleague's lunch order, you do fact check the tender submission. AI just makes the colleague faster and the discipline more important.
The mindset
The goal is not blind trust, and it is not paranoia. It is the pilot's relationship with instruments: rely on them constantly, cross check them routinely, and never hand over your accountability. You remain the signature on the work. Everything on this page is designed to let you delegate the labour without delegating the responsibility.
Next: the other side of safety, what you should and should not be pasting into these tools in the first place.