OmniPulse | AI, tech and markets

Research integrity preserves an honest relationship between questions, methods, evidence and reported conclusions. Selective reporting can make a literature appear more decisive than the underlying research.

About 3 min to read · Take extra time for practice · AI-assisted learning notes · Expanded October 17, 2026

First, the key terms.

Preregistration: A time-stamped statement of intended questions and methods before relevant results are examined.
Publication bias: A published record distorted by selective availability of results.
HARKing: Presenting a hypothesis developed after seeing results as though it had been specified beforehand.

The idea, explained

Fabrication invents evidence; falsification misrepresents research processes or data. Questionable practices also include presenting a post-hoc hypothesis as planned, selectively dropping unfavorable results or searching analyses without disclosing the search. The Chrysalis effect describes how selective changes between initial research and publication can make results look more successful. This concern does not mean every revised analysis is misconduct: justified exploratory work is valuable when transparently labeled. Citation count and polished presentation do not substitute for checking methods and evidence.

STEP 1

Preserve the original question

Record the initial protocol, outcome definitions and analysis plan. Explain legitimate changes and their timing. The purpose is to let readers distinguish tests of prior claims from discoveries generated by the same data.

STEP 2

Report the analytical search

Document exclusions, alternative specifications, missing results and uncertainty. Replication checks whether findings transfer to additional evidence. Reproducibility of code is useful but does not by itself establish that the design or interpretation is valid.

STEP 3

Check sources and participant obligations

Verify citations and quotations against real publications, including corrections. For human-participant research, consult the relevant ethics-review and consent requirements before collecting or reusing data. AI tools do not inherit permission to process identifiable participant material.

Follow a worked example.

A fictional researcher tries 20 outcomes and finds one small p-value. Reporting only that outcome as the original hypothesis conceals the search and exaggerates how surprising the result is under the intended interpretation.

Walk it through, one step at a time

Keep the full set of outcomes and explain which were planned. Do not invent a prior prediction to match the favorable result.
Label the unexpected pattern exploratory and consider an appropriate multiplicity analysis. A new independent test can assess the resulting hypothesis.
If an AI assistant supplies supporting references, open them and inspect the relevant passages. Fabricated or irrelevant citations cannot become evidence through repetition.

Where people get stuck

Retelling an exploratory result as a planned success can mislead even when no individual data point has been fabricated.

Now try it yourself.

Pause before opening each answer. Explain your reasoning in your own words.

1. Check the core idea

Is it wrong to change a hypothesis after exploring data?

Reveal the answer & reasoning

Not automatically. Exploration generates useful hypotheses. The problem is hiding that timing or presenting the same-data discovery as independent confirmation.

2. Apply what you learned

Does a highly cited article require no further scrutiny?