Judging AI output: a quality rubric that fits your day

Why judgement is the real skill

Every tool you've learned produces text that reads well. That is exactly the problem: fluency is not accuracy, completeness or fit, but it feels like all three. A confident, well-structured answer triggers a "looks right" response in the reader, and that response is doing no checking at all. The whole of this lesson is about replacing "looks right" with a fast, deliberate judgement you can apply to anything an AI hands you.

You don't need a heavy process. You need four questions and a habit of asking them before output leaves your hands.

The four-dimension rubric

Score any AI output against these four, in this order. The order matters: accuracy first, because a beautifully written wrong answer is worse than a rough right one.

• Accuracy: is it true? Are the facts, figures, names and claims correct? This is where hallucination hides: the tool states things with total confidence whether or not they're real. Accuracy is the dimension fluency disguises most, and the one that does the most damage when it's wrong.
• Completeness: is anything missing? Did it answer the whole question, or drop a part? Did it quietly skip the awkward bit, the caveat, the exception? A tidy answer can be tidy because it left things out.
• Tone: does it sound right? Formal where it should be, warm where it should be, not overblown or robotic. Tone is easy to fix but easy to miss until a reader reacts to it.
• Fit for audience: is it right for this reader? A brilliant answer aimed at the wrong person fails. A board wants brevity and confidence; a frustrated customer wants acknowledgement; a new colleague wants the jargon explained. Fit is the dimension that a generic prompt gets wrong most.

Most weak output fails on one specific dimension, not all four. Naming which one turns a vague "this isn't quite right" into a precise fix, and often tells you exactly what to add to the prompt.

Here's a draft reply to a customer. Before I use it, assess it against four things and flag the weakest: (1) accuracy: any claim I should verify; (2) completeness: anything the customer asked that it doesn't address; (3) tone: right for someone who's annoyed but polite; (4) fit: right for a customer, not internal. Then give me one improved version. [paste draft]

The spot-check habit

You can't verify every word, and you don't need to. The trick is to check the things that matter and skim the rest. A fast spot-check:

• Check what's load-bearing. Any fact, figure, name, date or claim the reader will act on: verify those, at the source, every time. The best-man speech doesn't need checking; the board figure does.
• Sample the rest. Read a couple of supporting points closely. If they hold up, the middle is probably fine; if one is off, check harder: errors travel in packs.
• Watch for the confident specific. A precise statistic or a named source is exactly where a tool hallucinates most convincingly. Specificity earns more scrutiny, not less.
• Read once as the audience. One pass imagining you're the actual reader catches tone and fit problems that a writer's-eye read misses.

From the summary below, list separately every specific factual claim (figures, dates, names, sources) that I'd need to verify before relying on this. Don't re-summarise; just give me the check-list of claims. [paste summary]

When 'good enough plus an edit' wins

Here's the counter-intuitive part. People burn real time chasing the perfect prompt to get flawless output in one shot, rewriting the prompt five times to avoid a two-minute edit. That's usually a bad trade.

For most everyday work, the fastest route is: get a solid draft with a decent prompt, then finish it by hand. You bring the last 10% (the specific fact, the in-house phrasing, the judgement call the tool can't make) and you keep control of the final product. Chasing a perfect one-shot prompt has diminishing returns and hands more of the thinking to the tool.

Draft a short internal update on the feedback project for the ops team: three short paragraphs, plain UK English. Leave a clearly marked gap '[FIGURE]' anywhere a specific number is needed rather than guessing one, and flag anything you're unsure about. I'll finish it.

The exception is when the output goes out unedited: an automation that sends without review, an assistant a customer talks to directly. There, you can't add the last 10% afterwards, so the prompt and the guardrails have to carry more, and the checking has to happen up front. The rule underneath both cases is the same: know where the human edit is, and if there isn't one, raise your standards before the output ships.

Try it now

Common mistakes

• Reading for fluency, not truth. A polished answer feels finished. Fluency is the tool's default, not evidence of accuracy: score the content, not the prose.
• Checking everything or nothing. Verifying every word is too slow, so people give up and check nothing. Spot-check the load-bearing claims and sample the rest.
• Chasing the perfect prompt. Rewriting a prompt ten times to dodge a quick edit is slower and worse. Accept a strong draft and bring the last 10% yourself.
• Over-trusting because it self-assessed well. Asking a tool to rate its own output feels like a check, but a model can be confidently wrong about its own accuracy. Its self-score is a prompt for your attention, not a verdict. The human check, especially on facts, still has to be yours.

Keeping current

The rubric is durable: accuracy, completeness, tone, fit will judge AI output for as long as we use it. What changes is how convincing the fluency gets, which makes the discipline more important over time, not less. For how the tools themselves frame accuracy and their known limits, the vendors' own guidance is the honest source, such as OpenAI's help centre and Anthropic's Claude release notes. Accurate as of 13 July 2026.