ICH Q9's Rule: Effort Proportionate to the Risk── Only when the effort is apportioned to the size of the risk, and the line and the reason are written down, does leaving something lighter become design

Within a single day there are documents read to the last corner and documents read only for their main points. Now that generative AI supplies the drafts, which one gets the lighter reading is still settled on the spot and left out of the record. Is deciding in advance where to ease off laziness, or is it design? Vary the effort by the size of the risk, and write down the line and the reason, and it becomes design.

01Spread the effort evenly and the riskiest places end up thin

ICH Q9(R1) requires the level of effort, formality and documentation in quality risk management to be commensurate with the level of risk. All three are finite. Give more to one place and another gets less.

The riskiest places carry the larger loss if something is missed, and they also have more to check. So spread evenly, the effort reaching them falls short. Which document was read lightly on a given day is usually settled on the spot and never written down. Unwritten, nobody can state the reason six months later.

The problem is not the act of leaving something lighter. The dividing line between design and laziness is whether the place was chosen in advance and the choice recorded.

02Q9(R1) treats formality as a continuum, not a switch

The principle of commensurate effort continues into something more specific. Section 5.1 of Q9(R1) does not treat formality as a choice between formal and informal. It treats it as a range running from lower to higher.

Three things settle where in that range a process should sit. How much uncertainty is present. How important the subject is. How complex the matter is. Where all three are small, the lower end of the range may be chosen.

One further sentence is explicit. Resource constraints must not be used to justify choosing lower formality. Within this guideline, "we were short-staffed, so we kept it light" is not an argument that carries. The argument that carries is that the risk was small. The two explanations can end in the same decision while resting on different ground: one on the organisation’s circumstances, the other on the nature of the thing being assessed.

03In clinical trials too, processes must be proportionate to the risk

Formality is chosen on three grounds. A requirement of the same shape appears elsewhere in medicines.

ICH E6(R3) requires trial processes to be proportionate to the risks inherent in the trial and to the importance of the information collected. It also requires that factors critical to the quality of the trial be identified prospectively. The word doing the work there is prospectively.

Reasons manufactured afterwards and reasons settled in advance look much alike once written. The difference is whether the writer already knows the outcome. A review process that reads every page at the same density is not proportionate to the risk. With a written reason it is design; without one it is merely even.

04A procedure that discards information can raise accuracy

Being told to choose the lighter places in advance raises a further doubt. Does doing less not cost accuracy? Psychology reports cases where it does not.

Gigerenzer and Brighton wrote in 2009 that less information, less computation and less time can in fact improve accuracy. Many decision procedures work by ignoring information, and that ignoring contributes not only to efficiency but to correctness.

Simon had reached the same territory from another angle in 1956. For a creature without the capacity to find the optimal path, he argued, choosing means following a path that satisfies all of its needs at some specified level. A design that keeps searching for the optimum is sometimes not implementable at all.

The claim comes with conditions. It holds where cues are few and uncertainty is large. Where the environment is highly regular, the same procedure passes an oversight straight through. Nothing in the paper says that doing less is safe everywhere.

05At ICML 2026 the line nobody set was set by each reviewer

Doing less does not necessarily increase error. Even so, who draws the line is settled separately. ICML 2026 left a record of that.

The conference handled more than 24,000 papers with 17,000 reviewers. Across them, a policy prohibiting LLM use and a policy permitting limited use were assigned at random.

Acceptance decisions, scores and stated confidence barely moved with the policy. Changing the written rule barely changed the results.

What moved was elsewhere. Of the reviewers under the prohibition, 22.5% reported using an LLM anyway. Under the permissive policy, 36.5% reported at least one explicitly disallowed use. The policy existed as a document while the line was being drawn in each person's own hands. Peer review is a different activity from reviewing promotional material, and the figures do not transfer. What transfers is the fact that policy and practice diverged.

06Uncertainty, importance and complexity together settle the apportioning

Policies were assigned and each reviewer still drew their own line. After that result, what remains to be settled comes down to three things.

The first is the act of easing off itself. Choosing where to go lighter is a judgement, and it is design if the reason can be written down. If it cannot, the same act is laziness. The awkward part is that the two are indistinguishable by the act alone.

The second is choosing where to go heavier. Of the three grounds, uncertainty is the one to read first: effort belongs where uncertainty is large, which is not necessarily where the work is unfamiliar. In a field handled many times over, uncertain questions remain. Familiarity supplies the feeling that uncertainty has fallen without reducing the uncertainty itself. Of the three grounds Q9(R1) names, the only one familiarity actually lowers is how complex the matter feels.

The third is what happens when nothing is written. The 22.5% and the 36.5% at ICML showed that putting a policy in writing does not by itself align the lines people draw. Write only the coarse rule — prohibited or partly permitted — without writing where it is acceptable to go lighter, and the line gets drawn in each person's own hands. Lines drawn separately cannot be reconciled afterwards, because no record survives of who went lighter and where.

07As AI drafts multiply, the line's position has to be written down

Effort is apportioned on three grounds. That arrangement assumes a person wrote the draft. As generative AI produces more of the drafting, the line marking what a person reads moves.

Article 14 of the EU AI Act requires those using high-risk AI to remain aware of any tendency to over-rely on the output, and to be able to decide not to use the system. Over-reliance is written into law as a requirement because it happens on its own. The tendency to spread attention thinly over output that reads well is what the article presupposes in demanding that awareness, and it reaches AI drafts as well.

One thing appears in no guideline at all. Q9(R1) states that resource constraints must not justify lower formality. May time saved by AI justify it? Saved time is not a shortage of resources, so the sentence does not reach it directly. No public statement on the question has been found.

Since no such statement has been found, it has to be settled in-house for now. Write down what was settled, and when a public line eventually appears, the distance between the two can be measured. Write nothing, and there is nothing to measure against.

Deciding in advance where to ease off is not laziness. Apportion by the size of the risk, write down the line and the reason, and fix when it will be redrawn, and it becomes design.

Skip without writing, and the same act is laziness. Looking only at the act, there is no way to tell from outside which one it was. The records that survive are the only thing that shows the difference.