Lally's 66 days and the range where habit replaces will── Habit removes the effort of starting, but the line marking where to stop and think remains, and it vanishes unless written down

Every morning I reread the draft that generative AI has handed back, in the same order, starting from the same place. I sit down and look first at the numbers, then at where those numbers came from. Can habit take the place of will? Habit takes over the effort of starting. But the line between when to run it and when to stop stays with will, and that line disappears unless it is written down.

01Habit takes on the starting; it does not take on the timing

Wood, Quinn and Kashy, in a 2002 study, classified as habits those behaviours people listed that were performed just about every day and usually in the same location. The share came to 35 per cent in one study and 43 per cent in another. A substantial part of what people do each day repeats in the same place. Between a third and a half of it is tied to a place.

Lally and colleagues measured, in 2010, how long it took before an action began without conscious effort. The average was 66 days. So the capacity to keep going is genuinely available. Given time, the resistance at the moment of starting falls away. That part moves.

Something else is not available. The line marking when to run that action, and where to stop and think. An action can become automatic while the line never does. What I can fix in advance about my morning is the cue for starting, and nothing more. When the numbers in a draft fail to match the source, whether I stop there is a judgement made fresh each morning.

02Lally's 66 days is an average; the spread ran from 18 to a predicted 254 days

Habit takes on the starting. So how long does it take before it does?

In Lally's study, 96 people spent 12 weeks performing one action daily in the same setting. Each participant kept recording how far that action had come to begin without conscious effort. At some point the record stops climbing. The time until the climb stopped averaged 66 days.

The spread was wide. The fastest participant reached the plateau in 18 days. At the slow end sits 254 days, a figure that was never measured; it comes from extending the curve past the 12 weeks that were observed.

An average of 66 days cannot serve as a target number of days. It is a single point at the centre of a spread, and nothing guarantees that any given person falls at that centre. Someone who needs 18 days and someone who needs more than half a year appear in the same study.

03In the operating room and at the desk alike, it is the procedure that persists, not the will

The time to automaticity varies enormously. Even so, there are working environments where a fixed procedure has taken hold.

Haynes and colleagues examined 7,688 operations across hospitals in eight cities in 2009. After a 19-item surgical safety checklist was introduced, inpatient deaths fell from 1.5 per cent to 0.8 per cent, and serious complications from 11 per cent to 7 per cent. The study, however, compares the period before introduction with the period after. Its design cannot establish that the checklist caused the fall in deaths.

Why such a procedure is easy to start can be explained without appealing to strength of will. Checking things in a fixed order does not draw on in-the-moment judgement at the point of starting. What decides the moment of starting is not the mood of the day but the setting: entering the operating room, sitting down at a desk. My own morning rereading survives on the same footing. Sitting down is the cue, and I do not measure my appetite for the work each morning.

04The idea that will gets used up lost its support in a 23-lab replication

A procedure starts without drawing on will. So does will itself actually run down?

One explanation held the field for a long time. Will is a resource that depletes with use, and once depleted a person gives way. Procedures, the argument went, are there to substitute for it.

Hagger and colleagues set out in 2016 to replicate that effect under a pre-registered protocol. Twenty-three laboratories took part, with 2,141 participants in total. The effect size came out at d=0.04, and the 95 per cent confidence interval spanned zero.

That result does not establish that the depletion of will never happens. What it establishes is that the effect did not replicate under this protocol. Still, the argument that procedures are needed because will runs out cannot be used at present. The experiment holding it up did not reappear in the same form.

05People who set "if Y happens, then X" in advance reached their goals more often

Depletion of will does not explain persistence. So what was producing the capacity to keep going?

Gollwitzer and Sheeran gathered 94 independent tests in 2006. People who had settled in advance on a form of "when this setting arrives, I perform this action" reached their goals more often. The effect size was d=.65, which falls in the range described as medium to large.

What worked was not strength of will. It was tying one setting to one action in a single sentence, in advance. Once tied, the arrival of that setting no longer requires deciding again what to do.

d=.65 is an average across 94 tests. That is not a guarantee of working in any one situation in front of you. Even so, it shows where to put the condition. What survived testing was the method that placed the condition on the side of the setting. The method that placed it on the side of will did not survive.

06That a procedure persists and that it works are two separate checks

Tying a setting to an action helps people arrive. So how does anyone see whether the tied procedure actually worked? The examination divides into three.

The first concerns what habit takes over. What becomes automatic across 66 days is the action of starting. Whether a person halts when the numbers fail to match does not become automatic. So a record showing "kept it up every morning this month" tells me only that I kept going. Whether I looked properly lies outside that record.

The second concerns cases where a procedure is introduced and the results do not move. Urbach and colleagues examined more than 200,000 operations in Ontario, Canada, in 2014. After the surgical safety checklist was adopted, neither mortality nor complications improved significantly. That runs opposite to the eight-city study. The same procedure can be introduced, and whether it works varies with place and conditions. So what measures the success of an adoption is not the number of days it survived. It is the outcome.

The third concerns a pull specific to reading AI drafts. Parasuraman and Manzey concluded in 2010 that automation complacency is found in naive and expert participants alike and cannot be overcome with simple practice. Drafts returned by generative AI arrive with prose that looks like grounds. The more of that prose is attached, the harder it becomes to see which parts I have not verified. Article 14(4) of the EU AI Act asks people overseeing high-risk AI to remain aware of the tendency to over-rely on its output. The duty in Article 14 reaches only settings classed as high risk, though. It does not govern every daily use of generative AI.

07As AI drafts multiply, setting the stopping points in advance is worth more

Persistence and effect are examined separately. What does that separation buy someone who reads AI drafts every morning?

The Parasuraman and Manzey conclusion carries a heavy implication for anyone using these tools. Simple practice alone was shown not to dissolve over-reliance. Expert participants carried the same pull. Accumulating repetitions is not, on its own, a countermeasure.

What follows from that is a narrow set of moves. Habit can absorb the order in which things are checked, and no further. Where to halt and think has to be decided in advance and written down. Left unwritten, the decision to halt is made fresh every time. In my own record, the mornings on which it dropped were the busy ones.

Applied to my own morning, two things needed adding.

After I wrote the second one, the content of my record changed. The old record was a count of days completed. What I write now is the number of times I halted, and what failed to match on each occasion. The count of days completed keeps rising. Nothing can be learned from it, though. The count of halts is counting the occasions on which the procedure actually did something.

Habit takes over the effort of starting, in place of will. The measured figures support that much. Resistance at the point of beginning thins out over an average of 66 days.

Something is left over. The line marking when to run the procedure and where to halt and think. That line sits with will, and its edges blur inside daily repetition. Unwritten, it does not survive.

So what I keep is not a count of days completed. It is the stopping conditions, together with a record of the days on which one of them actually stopped me. A count of days completed reassures me and hands tomorrow morning nothing.