Reading Well Is Not Being Right

The problem with good prose

You asked for a summary of a forty-page policy. What came back is clear, well-organised, appropriately hedged, and pleasant to read.

So you accept it.

Somewhere in there, three claims are wrong. Not absurdly wrong. Slightly wrong in the way that matters: a threshold, a date, an exception that does not exist.

Nothing about how it reads distinguishes those three from the seven that are correct.

What you are actually judging

The trap is that fluency is easy to assess and correctness is not. So attention drifts to the easy signal.

Discernment means checking the things that are harder to see:

• Are the claims true? Against the source, not against plausibility
• Does the logic hold? Does the conclusion follow from the stated reasons
• Is it right for the audience? Reading level, assumed knowledge, tone
• What is missing? The hardest one, because absence has no shape

That last one deserves attention. An AI summary rarely tells you what it left out. If a policy has an exception that matters to your case and the summary omits it, nothing in the text signals the gap.

Critique becomes the next prompt

When the first output is wrong, the common conclusion is that the model cannot do this task.

Usually it means the request under-specified something, and your reaction just told you what.

"Too formal" becomes write this the way you would explain it to a colleague at the next desk.

"Missed the exceptions" becomes list every exception and qualifier separately, even minor ones.

This is one loop, not two phases. Each judgement you make is the next instruction, and two or three rounds of it beats one carefully crafted prompt more often than people expect.

Where to look first

You cannot check everything. Some places repay attention far more:

• Numbers, dates, and names. Cheap to check, easy to fabricate
• Anything specific enough to be wrong. Vague claims are usually safe and usually useless
• Claims you were hoping were true. Your own motivation is a reliable pointer at where you will not look hard
• Anything you will forward. The moment it leaves you, it is yours

That third one is worth being honest about. The claim that supports the argument you already wanted to make is the one you are least likely to verify.