03 · What You Need to Know
Understanding Selective Prompt Reporting and Its Consequences
What does cherry-picking mean in scientific research?
Cherry-picking refers to selectively emphasizing evidence that supports a preferred conclusion while disregarding or concealing relevant evidence that complicates it.
In scientific research, the concern is not simply that researchers make choices. All studies involve decisions about measurement, analysis, and reporting.
The concern arises when those decisions are guided by observed outcomes and subsequently presented without adequate disclosure.
For example, a researcher may test several analytical approaches but report only the one producing a statistically significant result.
Prompt cherry-picking applies a similar concern to generative AI research, where changing instructions can change the observations or performance scores being reported.
How does prompt cherry-picking occur?
Prompt cherry-picking can occur during development, evaluation, interpretation, or reporting.
A researcher might test several prompts and publish only the highest-performing formulation.
Another might repeatedly revise prompts until a model produces the desired conclusion, then present the selected output as representative of the system's behavior.
In comparative studies, researchers might optimize one model extensively while using an untested prompt for another.
The resulting findings may reflect selective configuration choices rather than the broader capability the study claims to measure.
These practices become particularly problematic when the selection process is hidden from readers.
What are common forms of prompt cherry-picking?
| Practice |
Illustrative behavior |
Potential distortion |
| Best-prompt reporting |
Reporting only the highest-scoring prompt among many tested |
Inflated apparent performance |
| Selective model optimization |
Refining prompts extensively for one model but not another |
Unfair model comparison |
| Outcome-driven rewriting |
Changing instructions until AI produces a preferred interpretation |
Confirmation of researcher expectations |
| Selective response reporting |
Showing only successful generations from repeated executions |
Misleading impression of response reliability |
| Selective task reporting |
Highlighting categories where one prompt performs well |
Incomplete account of performance |
| Undisclosed prompt switching |
Changing prompts during formal testing without reporting it |
Loss of procedural comparability |
| Post hoc prompt justification |
Describing an outcome-selected prompt as originally planned |
Misrepresentation of the research process |
These practices differ in severity. Not every instance necessarily constitutes research misconduct, but each can weaken the validity or transparency of the evaluation.
Is selecting the best prompt automatically cherry-picking?
No. Selecting a prompt based on development performance is a normal part of system optimization.
For example, researchers may compare five prompts using a pilot dataset and select the one achieving the highest accuracy under predefined criteria.
They may then evaluate the selected prompt on independent data.
That procedure is methodologically different from testing many prompts on the final evaluation dataset and reporting only the highest observed score as though it were independent evidence.
The distinction depends on how the prompt was selected, which data informed selection, and whether those decisions were reported accurately.
The principles of prompt engineering bias are relevant to this distinction.
Why can selecting the highest score exaggerate performance?
Observed performance varies across prompts and evaluation samples.
Some prompts may achieve unusually favorable scores on a particular dataset because of sampling variation or features specific to those items.
When researchers select the highest score from many attempts, the selected estimate may be more optimistic than the prompt's performance on new data.
This resembles the selection bias described in methodological research on model optimization and performance evaluation.
The problem does not require deliberate manipulation. It can arise simply because the same data are used both to choose a configuration and to estimate its generalizable performance.
How can selection optimism be illustrated?
Consider a researcher who evaluates five prompts using the same development dataset.
| Prompt |
Development accuracy |
Independent test accuracy |
| A |
79% |
80% |
| B |
83% |
82% |
| C |
88% |
81% |
| D |
84% |
83% |
| E |
82% |
82% |
Prompt C achieves the highest development accuracy at 88%, but its independent test accuracy is 81%.
This hypothetical result illustrates why the highest development score should not automatically be treated as the best estimate of future performance.
The difference could reflect overfitting, sampling variation, distribution differences, or other factors.
Researchers should investigate these possibilities rather than attributing every decline to one cause.
Can prompt cherry-picking affect qualitative research?
Yes. The concern extends beyond numerical AI benchmarks.
Suppose a researcher uses AI to interpret interview excerpts about students' experiences with online learning.
Several prompts produce different interpretations. Some emphasize flexibility and accessibility, while others identify isolation, workload, and unequal access to resources.
If the researcher reports only the prompt producing a preferred positive interpretation while concealing relevant alternatives, the resulting analysis may become selectively framed.
However, qualitative analysis legitimately involves iterative interpretation. The concern is not iteration itself but the absence of a defensible analytical rationale and transparent reporting of material decisions.
Researchers should distinguish theory-informed refinement from outcome-driven selection designed to confirm a preferred narrative.
Can prompt cherry-picking occur in AI-generated research summaries?
Yes. Researchers may repeatedly ask an AI system to summarize the same evidence until it produces a conclusion that supports their hypothesis.
For example, an initial summary may describe mixed findings about an educational intervention.
After several increasingly leading prompts, the model may emphasize positive outcomes while minimizing limitations.
Reporting only the final favorable summary could misrepresent the underlying evidence.
The appropriate response is not to assume that the first AI output was correct.
Researchers should evaluate summaries against the original sources and document material analytical decisions rather than selecting outputs solely because they support a preferred conclusion.
How does prompt cherry-picking differ from prompt sensitivity?
Prompt sensitivity
The extent to which AI responses or performance change when prompt formulations vary.
Prompt cherry-picking
Selective choice or reporting of prompt outcomes that creates a misleading impression of performance or evidence.
Prompt sensitivity is a characteristic that researchers may investigate scientifically.
Cherry-picking concerns how researchers select, interpret, and report results from that variation.
A model may be highly sensitive to prompt wording without the study being biased, provided the variation is examined transparently.
Conversely, even modest variation may become misleading when researchers selectively report favorable outcomes.
Does prompt cherry-picking require an intention to deceive?
No. Selective reporting can arise through ordinary research decisions, including a desire to improve a method or simplify a manuscript.
Researchers may mistakenly believe that reporting the best prompt is sufficient because the selected configuration is the one they intend to use.
However, intention matters when evaluating research-integrity allegations.
A methodological weakness does not automatically establish fabrication, falsification, or other misconduct under applicable policies.
Researchers should distinguish the validity of a performance claim from judgments about the researcher's conduct.
How does cherry-picking differ from prompt optimization?
Prompt optimization seeks to improve a model's performance on a defined task.
Prompt cherry-picking concerns selective presentation or interpretation of the resulting evidence.
A researcher can optimize prompts extensively without misleading readers if the development procedure is documented and final performance is evaluated independently.
Conversely, even a small number of prompt trials can create selective reporting concerns if only favorable outcomes are disclosed.
The number of prompts tested is relevant, but it does not independently determine whether cherry-picking occurred.
What is the role of prespecification?
Prespecification can reduce undisclosed flexibility by establishing prompt conditions or selection rules before confirmatory evaluation.
For example, a study may specify that the highest-performing prompt on a designated development dataset will be selected for final testing.
This allows optimization while preserving an independent evaluation stage.
Researchers may also prespecify multiple prompt variants when prompt robustness is itself the research question.
The principles of prespecifying AI research prompts are therefore relevant.
However, prespecification does not automatically guarantee that the selected prompt is valid or unbiased.
How can researchers prevent test-set contamination?
Researchers should separate prompt development materials from the dataset used for final performance estimation.
They should also consider whether benchmark answers, related documents, or near-duplicate items are available through other channels.
For example, an AI model with retrieval access may locate published answers to benchmark questions.
Such access may compromise an evaluation intended to measure independent problem solving.
Researchers should document plausible contamination pathways and apply appropriate safeguards.
Independent testing reduces some risks but does not guarantee the absence of all contamination.
Should researchers report all tested prompts?
When prompt selection materially affects the research findings, preserving the candidate prompts and their relevant outcomes can strengthen transparency.
However, this does not necessarily require reproducing every informal exploratory interaction in the main manuscript.
A development log, supplementary table, or repository may provide a more practical record.
Researchers should report enough information for readers to understand the selection procedure and assess whether the final performance estimate was independent.
The appropriate level of disclosure depends on the research question, the scale of development, and applicable confidentiality restrictions.
Should failed prompts be included in the research record?
Relevant unsuccessful prompts can provide important evidence about the development process and limitations of the AI system.
For example, repeated failures to classify mixed-methods abstracts may reveal a systematic difficulty that is concealed by reporting only the final optimized prompt.
However, researchers should distinguish failed development attempts from failures observed during the final evaluation.
These categories have different implications for performance estimation.
The reporting of failed AI research prompts should reflect their methodological significance.
Can cherry-picking affect human-versus-AI comparisons?
Yes. Researchers may test many AI prompts and select the configuration producing the largest advantage over human participants.
If the human baseline remains fixed while AI performance is repeatedly optimized using the final comparison items, the reported difference may be exaggerated.
However, comparing an optimized AI application with a realistic human workflow can be legitimate.
The study should document the optimization procedure and avoid presenting a development-selected advantage as an independent estimate of general superiority.
The principles of fair human-AI performance comparison remain applicable.
Can researchers calculate how much cherry-picking inflated the result?
Sometimes, but the estimate depends on the available records and research design.
Researchers may compare the selected development score with performance on an independent test set.
They may also examine the distribution of results across all tested prompts.
However, the difference between development and test performance is not necessarily attributable entirely to cherry-picking.
Dataset differences, model changes, and sampling variability may contribute.
A precise causal estimate of selection effects may require a controlled simulation or experimental design.
What should researchers do if they discover cherry-picking after analysis?
Researchers should reconstruct the relevant prompt-development history where possible and determine whether the reported results were selected using the final evaluation data.
If so, the existing results may need to be reclassified as exploratory or development findings.
A fresh independent evaluation may be necessary to support confirmatory claims.
Researchers should also correct any misleading description of the selection procedure.
If the study has already been published, the appropriate response depends on the severity of the issue and the policies of the journal or institution.
The priority is accurate reporting and preservation of the scientific record.
How should prompt selection be described in a research paper?
The methods section should identify the number or scope of candidate prompts, the development materials, selection criteria, and whether final evaluation data were used during optimization.
For example:
Illustrative Methods Description
"Twelve candidate prompts were developed using a separate set of educational research abstracts. Prompt selection was based on predefined classification accuracy and output-compliance criteria. The selected prompt was finalized before evaluation on the held-out dataset. Development scores were not included in the confirmatory performance estimate. The candidate prompts, selection procedure, and relevant results are documented in the supplementary materials."
This description allows readers to distinguish legitimate prompt optimization from undisclosed outcome-driven selection.
