03 · What You Need to Know
Building Reproducible Research Around AI Systems That May Change
What does reproducibility mean when studying AI?
Research reproducibility is often discussed as though it means obtaining exactly the same result whenever an experiment is repeated. In practice, scientific terminology distinguishes several related objectives.
The National Academies of Sciences, Engineering, and Medicine distinguishes computational reproducibility, which concerns obtaining consistent results using the same data and computational procedures, from replicability, which concerns obtaining consistent findings across studies using new data intended to address the same scientific question.
Other disciplines and standards use these terms differently. Researchers should therefore define the meaning adopted in their own study.
| Objective |
What another researcher attempts |
Principal requirement |
| Computational reproducibility |
Recalculate the published findings from the original research data |
Preserved outputs, scoring data, analysis code, and computational dependencies |
| Procedural reconstruction |
Reconstruct the original AI evaluation procedure |
Complete protocol, prompts, configurations, and testing records |
| Experimental replication |
Repeat the evaluation with newly generated responses |
Access to sufficiently comparable systems and testing conditions |
| Temporal robustness |
Determine whether the findings remain consistent after system changes |
Comparable evaluations across documented model versions or periods |
These objectives should not be treated as interchangeable. A researcher may successfully reproduce an analysis using archived responses while being unable to regenerate those responses from a discontinued commercial model.
That limitation does not erase the original observations. It changes which aspects of the study can be independently verified.
Why are commercial AI systems difficult to reproduce exactly?
Commercial AI services are often accessed through interfaces or application programming interfaces controlled by their providers.
Researchers may know the public model identifier but lack access to the model weights, training configuration, deployment infrastructure, hidden instructions, or internal update history.
A provider may replace a model, revise safety mechanisms, change tool integrations, or modify request processing. Some changes may occur without a new public-facing product name.
Even when the nominal model version remains stable, probabilistic generation and implementation-level factors can produce different outputs.
Consequently, researchers should distinguish changes in the system from ordinary variability under a fixed system condition.
The challenge is particularly relevant when studying commercial AI models that can change without notice.
What information should be recorded about the AI system?
A reproducibility record should identify the evaluated system as precisely as the available evidence permits.
Depending on the platform, useful information may include the provider, model identifier, application or API version, available generation settings, enabled tools, and any researcher-controlled system instructions.
The testing date is also important because an identical public identifier may not guarantee an unchanged deployment.
Researchers should record the exact available model version and preserve testing dates and times at a level appropriate to the research question.
When the provider does not expose a precise version or configuration, the researcher should document that limitation instead of inventing technical details.
Why should the complete prompt context be preserved?
The visible user prompt may represent only part of the input influencing an AI response.
System messages, developer instructions, few-shot examples, previous conversation turns, uploaded documents, retrieved information, and tool outputs may also affect generation.
For studies using an API, researchers can often preserve the complete request payload, subject to confidentiality and security restrictions.
For studies using a commercial chat interface, some higher-priority instructions may remain inaccessible.
Researchers should preserve the information they control and clearly distinguish it from provider-controlled conditions they cannot inspect.
When preserving prompts used in AI research, the record should maintain the order and structure of relevant messages rather than reducing a multi-turn interaction to its final user question.
Available system instructions should also be documented when they could influence the findings.
Should researchers archive the original AI outputs?
Yes, when permitted and feasible. Preserving the original outputs is one of the most effective ways to maintain the computational reproducibility of an AI evaluation.
If a model later becomes inaccessible, another researcher may still be able to inspect the generated responses, apply the scoring procedure, and verify the reported statistics.
For each evaluation item, the record may include the input identifier, complete output, model identifier, timestamp, trial number, scoring result, and relevant execution metadata.
Researchers should preserve unsuccessful generations, refusals, malformed outputs, and errors when these are part of the defined evaluation procedure.
Retaining only successful responses can distort the apparent performance of the system.
Public sharing is not always possible. Research records may contain confidential participant information, copyrighted material, proprietary prompts, or other restricted content. Appropriate access controls and institutional retention requirements should govern preservation.
Can a random seed guarantee reproducibility?
Not universally.
Some AI platforms expose a random seed or related reproducibility controls. Under certain implementations, these may make repeated generations more consistent.
However, a fixed seed does not necessarily guarantee identical outputs across different hardware, software versions, backend configurations, or provider deployments.
Researchers should document available seeds and settings while avoiding claims of exact determinism unless the relevant conditions have been established.
Even with a stable seed, a changed model may produce different results.
How should researchers handle nondeterministic responses?
For probabilistic systems, reproducing the same scientific finding does not necessarily require reproducing every generated sentence.
A study may instead evaluate whether repeated executions produce statistically compatible performance estimates.
Suppose a model is evaluated on research-method classification. Different runs may classify individual abstracts differently while producing similar aggregate accuracy estimates.
Whether this constitutes adequate replication depends on the study's outcome, uncertainty estimates, and prespecified criteria.
Researchers should use an appropriate repeated-prompt testing strategy when output variation is relevant.
They should also account for nondeterministic AI responses in the statistical analysis rather than treating every generated output as an independent evaluation item.
How can versioned research records improve reproducibility?
Version control helps distinguish the original evaluation procedure from later revisions.
Researchers can preserve dated versions of prompts, datasets, scoring rubrics, scripts, and analysis plans. Changes should be accompanied by an explanation of what was modified and why.
For example, revising a classification rubric after examining model outputs may affect the resulting accuracy estimate. The original and revised rubrics should remain identifiable.
Where appropriate, repositories can provide persistent identifiers for released research materials. Software environments may also be documented using dependency manifests, containers, or other reproducible computing practices.
These measures help reconstruct the analysis, but they cannot preserve a remote proprietary model unless the provider makes the necessary implementation available.
What if the original model is no longer accessible?
The study may still support computational reproduction using archived outputs and analysis materials.
However, independent regeneration of the original responses may no longer be possible.
Researchers should state this distinction explicitly.
For example, the methods or limitations section might explain that the original commercial deployment was unavailable at the time of replication, while the archived response dataset and analysis code remained accessible.
Using a newer model does not constitute an exact replication of the original system. It may instead provide evidence about whether the findings generalize across model versions.
The consequences of model research becoming outdated before publication should therefore be assessed separately from the reproducibility of the original analysis.
How should researchers design longitudinal AI evaluations?
When the research question concerns how AI performance changes over time, model updates become part of the phenomenon under investigation.
A longitudinal design may repeatedly administer a stable evaluation dataset to the same service or successive model versions.
Researchers should document the testing schedule, available model identifiers, prompt configurations, and any changes to the evaluation procedure.
Where possible, common evaluation items and scoring standards should be retained across testing periods.
However, repeated exposure to public benchmark items may introduce contamination concerns, and changes in tools or deployment conditions may complicate interpretation.
A longitudinal study should distinguish observed temporal differences from causal claims about particular provider updates unless the design supports those claims.
How much reproducibility documentation is enough?
The appropriate level of documentation depends on the study's claims and the technical accessibility of the system.
A study examining one fixed prompt may require a relatively simple record. A large evaluation involving multiple models, prompt variants, external tools, and repeated generations requires more extensive documentation.
The central question is whether an independent researcher can understand what was tested, reconstruct the accessible procedure, verify the reported analysis, and identify the limitations of repeating the experiment.
Researchers should not claim complete reproducibility when essential system components remain inaccessible.
At the same time, limited access to proprietary model internals does not automatically make a carefully documented behavioral study scientifically worthless.
