The plan, before any data
A study can find whatever it wants once it has the data. Here’s how we stop ourselves from doing that: everything the study will test, how we’ll test it, and what would count as finding nothing, written down before anyone takes part and dated by an outside public record. We can still change it, but only in the open.
This is a draft. We revised it once after asking an AI model to review its methods, and again on November 12, 2026 after checking its wording, scales and order against survey practice, and we say so here. The AI model is not a substitute for a person who does this for a living, and one will read it before it is filed. It goes to OSF Registries once the questions have been tried and the open decisions below are settled. Until then nothing here is a registration, and we will not call it one.
01
What it is
An observational study. Nobody is assigned to anything and nothing is changed in anyone’s work. For about a week, people report the moments that changed their mood or behavior, as they happen, in their own words, by text or voice note.
The question is not how bad work feels, but where the frustration goes: onto the tool, onto other people, or back onto the self, and whether that is patterned by apps, dashboards, metrics or automation being part of the moment. Results are reported as patterns within people, never as how common something is in the population, and never as one moment causing another.
02
The hypotheses
H1, the main one
For the same person, a moment involving an automated or algorithmic system is more likely to land on people, someone else or themselves, than one of their moments that did not. Someone else and themselves are also reported on their own, as two secondary outcomes, because a system that makes people blame themselves more and snap at others less would cancel out in the combined measure.
Against it: where the frustration goes has nothing to do with whether a system was involved. The difference has to be an odds ratio of at least 1.5, the odds of landing on people at least half again as high, and the 95% interval around it has to stay above 1, which means no difference at all is ruled out. If the estimate clears 1.5 but the interval reaches 1, the result is reported as unclear. Smaller than 1.5, an effect of this kind cannot be told apart from the ordinary hour-to-hour things that could also explain it (confounds) in a study this size, so we report it as no meaningful difference, however unlikely it is to be chance (its p-value).
H2, second
For the same person, those moments feel stronger: at least half a point higher on the 0 to 10 scale, about a quarter of a typical spread.
Exploratory
Whether H1 is stronger for people with more automation in their work. A test like this needs far more people than 60 to 90, so it is labeled exploratory everywhere it appears and cannot pass or fail the study. Comparisons between groups of people with more and less automation are described, never tested as the answer, because who signs up differs between those groups in ways we cannot control.
03
How each thing is measured
Whether a system was involved is read from what people wrote, never asked, because asking would put the idea in their heads. The same goes for what we show people: no example, prompt or recruitment page contains an app together with a reaction to it. Because a thin story could hide a system, every entry is also marked by how much was written, and the main analysis is repeated on entries with at least a sentence or two. Entries too short to code are reported as uncodable, with their share by person, and are never guessed. Two people code it, each sorting entries by the same written rules (the codebook), working separately, from transcripts, never audio, and without knowing anything else about the person. They both code the same fifth of entries, and they must agree well beyond what chance would give (Cohen’s κ of at least 0.70) before the measure is used at all. If they do not, the main analysis is reported as failed rather than patched.
Where it went is the person’s own answer at each check-in, to the question “where did most of that frustration end up?”: the tool, other people, myself, or nowhere, picking the closest. Choices are shown in an order that differs between people and stays the same for each person. For the main analysis it is whether it landed on people (other people or myself) or not, and each of the two is also reported alone. How strong is one number, 0 to 10, “how frustrated were you?”, with both ends described in every prompt in the same words: 0 is barely registered, 10 is as worked up as you get. When is one tap: just now, within the hour, earlier today, or before today. And each person’s exposure to automation is a score from the screener, used as a number, not cut into groups.
04
Trying the questions first
Before the wording is fixed, 10 to 15 people from our own networks answer the same questions for two or three days, with more choices than the study will have: nowhere split into “it faded” and “I kept it in”, split about evenly, somewhere else, and a box for none of these fit well. The four places are kept only if at least 90% of their answers fit them cleanly, the lowest the share could plausibly be (the 95% Wilson bound) is at least 80%, and there are at least 40 answers from at least 10 people. If it comes out looking split, the two kinds of nowhere are separated. Each pilot answer also says when it happened, which the study asks too. Because the study will take answers by text, the questions’ free-text parser is also run over the pilot’s stories and compared with what people picked. Before the wording is frozen, five to eight people who work in the jobs the study is for, such as driving, nursing or support, read each question aloud and say what it means to them. Their answers are design data: never analyzed, and none of them can join the study. Changed November 1, 2026: the draft had them use the real text-message flow, but a texting number needs carrier registration, which takes weeks, so they answer on a private page of their own instead. This tests the wording, not texting; a second short pilot by text, with different people, comes before the study opens.
05
Who takes part, and when it stops
At least 20 people in each of three bands of automation exposure, 60 to 90 in all, in one wave. Unpaid, adults, in English for now, recruited in public, at first through LinkedIn and the people we know. Recruitment closes at 90 people or six weeks from launch, whichever comes first, and does not reopen to chase a result. The analysis begins only after the last person’s week ends, and runs once.
An unpaid public study draws people who already feel something about work and technology. That is the central property of this sample, not a footnote, and it is why the main question compares people with themselves. People will drop out, probably the most frustrated first; we report who finished against who started, and we do not fill in missing answers.
06
The analysis
H1 is a multilevel logistic regression, a model of a yes-or-no outcome that compares each person’s entries with their own other entries: person_directed ~ system_involved_within + system_involved_person_mean + (1 | participant), where the first term is each entry’s system involvement minus that person’s own average and the second is the average, so the first term is the within-person effect alone, as the plan claims. We try letting the effect vary by person, and if that model will not fit, as is likely with five to seven entries each, we fall back to the simpler one. That fallback is declared here so it cannot be chosen after seeing the data. A conditional logistic model, which also uses only within-person differences, is run as a check.
Only people with at least one moment of each kind can inform the comparison. If fewer than 30 do, the main analysis is reported as too small to test (underpowered) and is described, not tested. Before the plan is filed, a simulation using realistic dropout and entries per person says how many people it takes for 30 to have both kinds, and what odds ratio the study could detect; if the honest answer is that it is likely underpowered, the plan says so up front. Three checks are declared in advance: the same model on moments reported within the hour, on one entry per day, and with time of day added. H2 is the same comparison on intensity. Moments reported as before today are left out of every test of the hypotheses and counted separately. Every subgroup, by job, channel or anything else, is exploratory and labeled so.
07
Ethics
An independent review board is to review the study before it opens. Consent comes before any data, and says plainly that the study is unpaid. Voice recordings get their own written permission, are transcribed on our own computers and never go to an outside company. A plan for anyone in distress, with crisis lines, is in place before the first check-in.
08
Still to decide
Who runs the models (they will not code), a survey methodologist who is a person and who reads this and the questions before it is filed, the screener’s questions, published so they can be checked, with the high, medium and low exposure groups renamed for what they measure (automation and metrics, not AI), how a voice note records the answers to where it went and how strong it was, which recruitment channels beyond the first, the review board and its reference, and what, if any, de-identified data is shared, which needs a check for re-identification first. Each blocks registration until it is settled, and each will be settled here, in public.
The limits are written out on the study’s page, and the consent people would be asked for is here. The questions word for word, where we think this plan is weakest, and a way to object are on the check page.