Pre-Registration on OSF: AI-Assisted Drafting That Doesn't Box You In

Preregistration is useful because it records the study plan before outcomes can influence analytical choices. AI can turn settled decisions into a structured draft, but it cannot decide which outcomes, exclusions, covariates, or contingencies are scientifically defensible.

OSF registrations create a time-stamped, read-only version of a project. The value comes from the specificity of the plan and the transparent reporting of later deviations—not from completing every field quickly.

Two ways a plan can fail

Under-specification leaves the analyst too many undisclosed choices. “We will use appropriate statistical tests” does not identify the estimand, model, covariates, missing-data strategy, or decision criteria.

False precision is different. A model may invent a complete-looking analysis pipeline despite missing information about clustering, repeated measures, outcome distribution, or the unit of analysis. Detail is only useful when it represents a decision the research team has actually made.

The OSF Template Fields Where AI Is Safe

AI is useful for converting source material into draft prose:

• summarize the approved protocol background;
• translate a fixed PICO into explicit hypotheses;
• restate registered eligibility criteria;
• format an independently verified sample-size justification.

Require a source location for each statement. If the protocol does not contain the decision, the draft should display [DECISION REQUIRED] rather than fill the gap.

Protect the analysis plan

Start with the estimand and design, not a menu of hypothesis tests. Specify the unit of analysis, outcome definition and time point, model family, effect measure, covariates, multiplicity, missing-data handling, exclusions, and sensitivity analyses.

Contingencies are legitimate when they are explicit and scientifically motivated. Write them as decision trees with observable criteria. Avoid automatic rules such as choosing a parametric test whenever Shapiro–Wilk gives p > 0.05; such a threshold does not by itself establish model adequacy and encourages data-dependent test selection.

For example:

The primary analysis will estimate the adjusted between-group difference using [model], consistent with the prespecified estimand. If model diagnostics indicate [defined problem], we will apply [specified robust approach]. The original analysis will remain reported as a sensitivity analysis.

The bracketed decisions must come from the study team and statistician.

A safer drafting prompt

Using only the approved protocol and statistical analysis plan, draft the OSF fields. For every analytical choice, cite its source location. Mark absent decisions as [DECISION REQUIRED]. Separate confirmatory, secondary, sensitivity, and exploratory analyses. Do not invent thresholds, covariates, exclusions, or fallback tests.

Then run a second pass that asks only for ambiguity: list every point at which two analysts could make different choices while still claiming to follow the text. Resolve those branches before registration.

Freeze only what the team can defend

Registration makes the submitted plan read-only, so review the final preview rather than relying on the editable project. Confirm authorship, embargo settings, linked files, hypotheses, outcome timing, and analysis text. Sensitive or identifiable data do not belong in public registration materials.

An embargo can control when the registration becomes public; it does not turn an unfinished analysis plan into a defensible one. Complete the scientific decisions first, then choose the access setting that fits the study and applicable policy.

The Over-Specification Audit Before Filing

Before registration, verify the draft line by line against the protocol, confirm that primary and exploratory analyses are visibly separated, and archive the code or pseudo-code needed to implement the plan. Later changes may be justified; they should be dated, explained, and reported as deviations rather than quietly folded into the confirmatory analysis.

Self-Peer-Review with AI shows how to compare the eventual manuscript with the registered plan. Statistical Assumption Checks covers the diagnostic layer without turning it into post-hoc test shopping.

AI for Academic’s Research Mentor can help structure the upstream question and PICO. Carry only approved decisions from that work into OSF, and leave unresolved choices visibly unresolved until the study team settles them.