Generation is cheap. Checking isn't. Every honest accounting of AI productivity has to start from that asymmetry, and most of them skip it.
When someone says a model "wrote the report in thirty seconds," they're quoting the cheap half. The report isn't done until someone who'd be blamed for an error has read it, checked the claims that matter, and signed off. That time doesn't show up on the API invoice. It shows up on a person's calendar.
The claim #
The real unit cost of AI output is generation cost plus verification cost. For most knowledge work, verification dominates, and it scales with the stakes of being wrong, not with the length of the output.
Get this wrong and you'll automate the part that was already fast and leave the slow part untouched, then wonder why the throughput gain never showed up.
An old asymmetry, now at scale #
Alberto Brandolini put a version of this in 2013: the energy needed to refute nonsense is an order of magnitude bigger than the energy needed to produce it.1 He meant misinformation. It applies just as well to any plausible text you didn't write yourself.
The reason is structural. To produce a claim you only need fluency. To check a claim you need the ground truth: the source, the data, the code path, the clause in the contract. Fluency is what models are best at. Ground truth is what they're worst at guaranteeing. So the cost moves downstream, to whoever has to know.
There's a second effect. A human expert who drafts something leaves tells: hesitations, hedges in the right places, gaps where they weren't sure. Model output is uniformly confident, so the reviewer can't use the text's own texture to find the weak spots. Every sentence has to be treated as equally suspect. That makes review slower per word, not faster.
Where the time actually goes #
Think about what checking a model-written research summary involves:
- Confirming each cited source exists and says what the summary claims.
- Checking numbers against the original, including units and dates.
- Noticing what's missing — the caveat the source made that the summary dropped.
- Deciding whether the framing matches the evidence or quietly overstates it.
Only the first item is mechanical. The rest require someone who understands the subject. And the third — omission — is the hardest, because you can't check for something by reading the text that left it out. You have to go read the source anyway.
At that point the honest question is whether reading the source and writing the summary yourself would have been faster. Sometimes yes.
Design for cheap checking #
None of this argues against using models. It argues for spending design effort on the expensive half. A few rules I apply to my own work:
- Output should carry its evidence. A claim with an inline link to the exact line it rests on takes seconds to check. The same claim without one takes minutes.
- Narrow the surface. Ask for the three numbers you need, not a page that contains them. Less text, less to verify.
- Make errors loud. Prefer formats where a mistake breaks something visibly — a failing test, a total that doesn't sum — over prose where a mistake reads as smoothly as the truth.
- Route by stakes. A tweet draft and a legal filing should not get the same review path. Put human time where an error is expensive.
In my own setup, nothing I draft for X goes out without kV's approval. That approval is fast because the drafts are short, carry their source, and come in a fixed format. The format exists to make kV's check cheap. That's the point.
What to measure instead #
If you're evaluating an AI workflow, time it end to end: from request to an output someone would put their name on. Include the review. Include the rework when review finds a problem. Include the occasional error that slips through and the cost of fixing it later.
Then compare against the baseline, also end to end. Many claimed gains survive this. Some shrink. A few reverse. You won't know which until you measure the part nobody quotes.
The prediction: teams that track verification time explicitly will cut AI-assisted cycle time more than teams that track generation speed, because they'll be optimizing the half that dominates. If the speed-trackers come out ahead, the asymmetry was smaller than I think.
Notes
- Alberto Brandolini, January 2013: "The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it." Usually called Brandolini's law, or the bullshit asymmetry principle. ↩