Quality Control and Verification

The concept

This module has covered research, drafting, and analysis — three places AI genuinely speeds up real work. This lesson gives you the habit that makes all three safe to use: a short, repeatable verification checklist applied to any AI output before it goes anywhere that matters (a client, a publication, a decision, your boss).

The checklist has four questions, each catching a different kind of error. First: does every number, name, date, and quote trace to a real source? This is the most important check and the one most people skip, because confident prose feels verified even when it isn't — a well-written sentence containing a wrong statistic reads exactly like a well-written sentence containing a correct one. Second: does the logic actually follow? AI output can sound coherent sentence-to-sentence while the overall argument doesn't hold together — a conclusion that doesn't actually follow from the evidence presented, or a recommendation that contradicts an earlier paragraph. Third: does the tone match the audience? An AI-drafted message can be factually fine but wrong in register — too casual for a board update, too stiff for a customer reply — and this is a judgment call only you can make, since the AI doesn't know your specific audience's expectations as well as you do. Fourth, and the most useful gut-check: would you be comfortable if someone fact-checked this publicly? If the honest answer is "I'd need to double-check that first," that's the output telling you it isn't done yet.

A useful supplementary technique is cross-checking one AI output with a second prompt or a different AI model — asking "does this hold up, what would you push back on" or running the same question through a different tool and comparing answers. This catches a meaningful share of errors, especially logical ones, because a second independent pass is more likely to notice something the first pass smoothed over. But it is explicitly a supplement to human verification, not a replacement for it — two AI models can share the same blind spot (for instance, both confidently misremembering the same kind of statistic), and neither one has access to your specific source documents unless you give them to it. The final check always has to be a human looking at a real source.

The good news about this checklist: it gets faster with practice, not slower. The first few times, running all four questions on every output feels like it's adding real time. After a few weeks it becomes closer to a quick reflex — the kind of "wait, let me check that number" pause that any good professional already does before sending something important, just applied consistently to AI-assisted work too.

Worked example

A learner asks an AI assistant for a quick market-sizing stat to open a presentation: "What's the size of the U.S. meal-kit delivery market?" The AI responds with a specific figure and a confident, well-written sentence citing a market research estimate. It reads as fully verified — specific number, named source, no hedging language.

Running the checklist: the learner checks the number against the actual market research report the AI named. The figure is in the right ballpark — but it's from a report two years older than the AI implied, describing a market that has since grown substantially. The number "sounds right" in isolation (it's not wildly off) but is wrong in a way that matters for a presentation about current market size — "right ballpark, wrong year" is exactly the kind of error that confident prose hides well. The learner replaces it with the correct current-year figure from the report, cited properly, before using the slide.

This is the core lesson of quality control: "sounds right" and "is right" are not the same test, and only checking the original source catches the difference.

Exercise

Pick one AI output you generated earlier in this course — a research synthesis, a drafted email, an analysis summary. Apply the four-question checklist to it: trace every number, name, date, and quote to a real source; check whether the logic holds together end to end; check whether the tone fits the intended audience; and ask yourself honestly whether you'd be comfortable if it were fact-checked publicly.

Manually verify at least two specific factual claims against independent sources (not just by asking the AI again). Note whether anything needed correction, and if it did, what kind of error it was — a wrong number, a logic gap, a tone mismatch, or an unverifiable claim. Keep this short checklist somewhere you'll see it again; the goal is to make running it a reflex on real work after this course ends.

Recap

• Run four checks on any AI output before it matters: sourced facts, sound logic, matching tone, and the "comfortable if fact-checked publicly" gut-check.
• "Sounds right" and "is right" are different tests — confident, well-written prose can carry a subtly wrong fact undetected.
• Use a second AI pass as a useful supplement to catch logic gaps, but always finish with a human checking a real source — that step can't be skipped.

Resources

• Anthropic: "Reducing hallucinations" — https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
• Poynter: "A guide to fact-checking" — https://www.poynter.org/fact-checking/
• International Fact-Checking Network (IFCN) Code of Principles — https://www.ifcncodeofprinciples.poynter.org/