The cost of checking can exceed doing the work oneself

What it is

A literature review with citations is requested. Reading it, opening each citation, fixing the mismatches, takes longer than pulling three real reviews from a search and rewriting. Generation made production cheap and confirmation expensive. Verification cost inversion means the labour of checking a generated artefact can exceed completing the same task from a blank page.

It is not the algorithm of “a wrong line inside a right paragraph must be checked sentence by sentence.” It is whether confirming the whole delivery, as a unit of labour, still pays.

Why it happens

When people write, evidence is gathered along the way: pages they opened, numbers they wrote down, a memory of why this sentence is here. A generated artefact cuts that chain and leaves the finished surface. The checker must reverse-engineer: where this sentence might have come from, whether this number looks real, whether this structure is covering a hole. Reverse engineering has none of the writer’s situational memory; every step is a new search.

Generation also enlarges the surface to check. People writing themselves rarely invent twelve citations that look citable; a generator will. Checking volume follows generator yield, not the amount the user would have been willing to write.

Studying it

Same task (a review paragraph, a filled table, a fact-bearing email): compare “from a blank” with “generate then check to equal reliability.” Record total time, number of searches, residual errors. Independent variables: length of the generated artefact, fact density, whether openable sources are provided. Dependent variables: time ratio of the two paths, residual error after checking, whether people abandon checking and adopt anyway.

The key ratio is t(check-to-reliable) / t(write-to-reliable). Above 1 is inversion. Reporting only “generation was faster” books unchecked delivery as an advantage of generation.

Where it stops holding

Restyling, expanding bullets the user already checked, translating confirmed text: checking is almost comparison with the original, and inversion is unlikely. Tasks the user would never do and cannot do (a first draft in another language) have no self-write baseline; “more expensive than writing it myself” is meaningless and should be compared with not using the system. When a ready checker exists (a compiler, unit tests, a database query), confirmation is cheap and inversion is rare. This entry compares labour for the whole task. It does not treat whether the user will check, or can.

Applying it

• At the generation entrance, say that if the output will be used as fact, checking time must be budgeted. Do not let “a draft in seconds” imply the task is done.
• For fact-dense requests, default to a short output or an outline plus a source list, not a long piece that must be reverse-engineered entire.
• Provide comparative checking rather than “please verify yourself”: generated sentence beside source, click-to-locate.
• Check: have target users do a real task; measure time to “checked enough that I would hand this over” against time to finish from a blank. If the former is longer and residual errors remain, generation is a net liability on that task.

Related

• Same group: L3.03.1 Fluent wording is not the same as being correct · L3.03.3 High-stakes settings must not leave checking entirely to the user · L3.03.4 When errors sit inside correct content, checking is sentence by sentence and costs about as much as rewriting · L3.03.5 The less familiar the domain, the harder checking is — and that is exactly when people ask the system · L3.03.6 Certainty of wording has no relation to reliability of content · L3.03.7 Errors in numbers, dates, and names are the hardest to notice and the most damaging · L3.03.8 Putting the checking duty in a disclaimer does not reduce how far errors actually spread
• Nearby: L3.02 Source Attribution · L3.08 Source Attribution of Generated Results
• Search terms: verification cost inversion · checking cost · effort of validation