Can Generative AI Invent Data While Creating a Research Visualization?

What Happens When AI Adds Data That Your Research Never Collected?

You ask generative AI to create a graph showing how students' academic performance changed during an AI-supported learning intervention. Your dataset contains pretest and posttest scores, and you want a clear visualization of the observed difference.

The resulting graph looks convincing. It includes a baseline measurement, several intermediate observations, and a smooth improvement toward the final assessment.

There is one problem. Your study never collected the intermediate measurements.

The model has created a plausible trajectory between two actual observations. If you publish the graph without checking it, readers may believe your research documented a gradual improvement that was never measured.

How can researchers distinguish legitimate statistical estimation from invented data, and when does an apparently helpful visualization become a misrepresentation of scientific evidence?

Yes, Generative AI Can Introduce Values That Were Never Observed

How Invented Values Enter AI-Generated Research Figures

What Does It Mean for AI to Invent Research Data?

Invented data are numerical values or observations introduced without an authentic empirical source or a properly identified analytical basis.

For example, a researcher may supply two observed group means, but an AI model generates additional group means to create a more complete comparison.

The added values may resemble plausible research results, yet they do not correspond to measurements collected during the study.

It is important to distinguish the generation of unsupported values from legitimate computation. Calculating a mean from recorded observations is not inventing data. Neither is generating simulated observations for a clearly identified methodological demonstration.

The problem arises when generated values are presented, intentionally or inadvertently, as though they were actual research findings.

Why Might Generative AI Invent Values While Making a Graph?

Generative models can respond to incomplete instructions by producing plausible continuations based on learned patterns.

When a request describes a desired visualization without supplying all necessary values, the model may fill informational gaps rather than explicitly identifying the missing data.

For example, a prompt requesting a graph of student engagement across an entire semester may lead to an invented weekly series if only beginning- and end-of-semester measurements are available.

This behavior should not be confused with a universal property of every AI-enabled plotting system. A properly constrained computational workflow can refuse to plot unavailable observations or display missing values explicitly.

The risk depends on how the system interprets the request, which tools it uses, and whether the plotted values remain connected to a verified dataset.

What Is the Difference Between Invented, Simulated, Imputed, and Predicted Data?

These categories may involve values that were not directly observed, but their methodological meanings differ.

A value's numerical plausibility does not determine its evidential status. Researchers must establish how it was produced and what it can legitimately represent.

Can AI Invent Entire Experimental Groups?

Yes. If a prompt requests a comparison involving groups that are not represented in the supplied dataset, a generative model may produce illustrative values for the missing categories.

For example, a researcher may provide results for conventional instruction and AI-supported instruction, then request a chart comparing conventional, blended, and AI-supported learning.

Unless the system is constrained to the actual data, it may generate a blended-learning group to complete the requested figure.

The resulting chart could imply that the study included an experimental condition that was never implemented.

Researchers should verify that every displayed group corresponds to an actual study condition or is explicitly identified as hypothetical.

Can AI Invent Additional Measurement Occasions?

Yes. This is especially concerning in longitudinal research.

Suppose a study measured student motivation at the beginning and end of a course. An AI-generated graph may include intermediate weekly values to create a smooth trend.

Such a graph implies that the researcher observed the process of change over time.

However, two observed endpoints do not establish the trajectory between them. Motivation may have increased gradually, fluctuated, or changed only near the final assessment.

Researchers should not treat interpolation or generated trajectories as additional observations.

Where interpolation is scientifically justified, the estimated portions should be distinguishable from measured values.

Can AI Invent Error Bars and Confidence Intervals?

Yes. A model may add uncertainty indicators to make a figure appear complete, even when the necessary information has not been provided.

For example, a researcher supplies two means but no sample sizes, standard deviations, or model-based uncertainty estimates.

The AI-generated figure may nevertheless include error bars resembling 95% confidence intervals.

Those intervals cannot be justified solely from the two means.

Researchers should verify the analytical basis of every uncertainty indicator, including the estimation procedure, sample structure, and relevant assumptions.

Invented error bars are not merely decorative errors. They may mislead readers about the precision of the findings.

Can AI Invent Statistical Significance?

Yes. Generative AI may add significance stars, p-values, or other annotations that resemble conventional scientific reporting.

For example, a graph may display an asterisk indicating a statistically significant difference without any corresponding statistical test.

Visual differences between group means do not independently establish statistical significance.

Researchers should verify that each statistical annotation corresponds to an actual analysis and that the test, assumptions, and multiple-comparison procedures are appropriate.

When inferential results are unavailable, the graph should not imply that they have been calculated.

Can AI Invent a Trend Without Inventing the Original Values?

Yes. A model may use genuine observations but generate an unsupported relationship between them.

For example, the researcher provides a pretest mean of 60 and a posttest mean of 80. AI then creates a smooth line increasing steadily from 60 to 80 across ten weeks.

The endpoints are authentic, but the intermediate trajectory is not observed.

This illustrates why a figure can contain some correct values while still communicating a fabricated temporal pattern.

Researchers should distinguish observed relationships from model-based interpolations or predictions.

Can AI Generate Values by Misunderstanding Missing Data?

Yes. Missing-data codes, empty spreadsheet cells, or absent categories may be interpreted incorrectly.

For example, a dataset may use 999 to indicate a missing assessment score. If the AI-generated code treats 999 as an actual score, the resulting mean or graph may be severely distorted.

Alternatively, the model may replace missing values with plausible numbers without documenting the substitution.

Researchers should inspect missing-value conventions and specify how missing observations must be handled.

Legitimate missing-data analysis requires methodological justification. It should not be improvised to make a visualization appear complete.

Does Statistical Imputation Count as Inventing Data?

Not when it is used as a legitimate, appropriately justified statistical procedure and represented accurately.

Imputation estimates missing values using information and assumptions about the data-generating process.

For example, multiple imputation can generate several completed datasets to reflect uncertainty about missing observations, with analysis results subsequently combined using appropriate procedures.

Rubin's work on multiple imputation provides a foundational statistical framework for this approach.

However, imputed values should not be represented as though they were directly observed. The missingness assumptions, imputation model, and analytical implications should be reported.

Simply asking a generative AI model to fill empty spreadsheet cells with reasonable-looking numbers is not equivalent to a validated missing-data procedure.

Can Simulated Data Be Used in Research Figures?

Yes. Simulation is an established research method, particularly in methodological studies, computational modeling, and educational demonstrations.

A simulation may generate data under explicitly defined parameters to investigate how a method behaves or to illustrate a statistical concept.

For example, a researcher may simulate datasets with different sample sizes to examine the performance of an estimator.

The resulting figures can be legitimate when the simulation design and assumptions are reported accurately.

Problems arise when simulated outcomes are presented as though they were collected from actual participants or experiments.

The distinction concerns provenance and representation, not whether a computer generated the values.

Can AI Forecast Values for a Research Visualization?

Potentially. Forecasting involves estimating future or otherwise unobserved values using a defined model.

For example, a time-series model may forecast enrollment trends based on historical records.

Such forecasts should be clearly distinguished from observed data, with appropriate uncertainty and methodological assumptions reported.

Researchers should not use generative AI to produce a smooth future trajectory without establishing the prediction method.

A forecast is an estimate conditional on assumptions, not a measurement of an event that has already occurred.

Can AI Invent Data While Generating Plotting Code?

Yes. Code generation does not eliminate the possibility of unsupported values.

For example, AI may insert a hard-coded array of example scores instead of reading the uploaded dataset.

The code may execute successfully and produce a visually correct graph, but the values would come from the generated script rather than the study.

Researchers should inspect whether the plotting procedure loads the intended source data and whether any hard-coded values represent verified statistics.

The distinction between creating figures from authentic raw data and generating illustrative numbers must remain explicit.

How Can Researchers Detect Invented Values?

Detection begins by establishing a traceable source for every plotted quantity.

For each observation, summary statistic, prediction, or interval, researchers should identify the corresponding source record or documented calculation.

If a value cannot be traced to either authentic data or a legitimate analytical procedure, it should not be accepted as an empirical result.

This provenance-based approach is more reliable than asking whether the values look reasonable.

What Is the Difference Between Invented Data and Incorrectly Plotted Data?

An incorrectly plotted value may originate from genuine data but be displayed inaccurately because of a calculation or mapping error.

Invented data have no corresponding authentic observation or appropriately documented analytical basis.

For example, reversing two group labels creates an incorrect graph without necessarily inventing any values.

Adding an entirely fictional third group introduces unsupported results.

The related issue of AI graphs plotting incorrect values should therefore be distinguished from the introduction of new, unobserved data.

Both problems can mislead readers, but their causes and corrective actions may differ.

When Does AI-Invented Data Become Research Fabrication?

Under the U.S. federal research misconduct framework, fabrication involves making up data or results and recording or reporting them.

Presenting AI-invented observations as authentic empirical findings may therefore raise fabrication concerns.

However, an AI model generating hypothetical values in a clearly labeled teaching example is not the same as a researcher presenting fictional measurements as collected evidence.

Formal misconduct findings also require satisfaction of applicable criteria concerning the conduct, evidence, and mental state.

Researchers should distinguish an AI output containing unsupported values, an honest error that enters a manuscript, and conduct meeting the formal definition of research misconduct.

Regardless of the formal classification, unsupported values should be removed or corrected before the figure is used to support scientific claims.

Does Disclosure Make Invented Research Data Acceptable?

No. Disclosing that AI was used does not transform fictional values into authentic observations.

For example, a researcher cannot present invented student assessment scores as experimental results merely by adding an AI-use statement.

However, clearly labeled simulated or hypothetical values may be appropriate for methodological demonstrations when the research purpose supports their use.

The distinction is whether the figure accurately communicates the origin and evidential status of the values.

Disclosure supports transparency, but scientific legitimacy depends on the underlying method and representation.

How Can Researchers Prevent AI From Inventing Values?

Researchers should use verified datasets and require the visualization workflow to expose the exact data used for plotting.

Prompts should explicitly prohibit unsupported observations, invented statistics, and unrequested interpolation.

However, prompt instructions alone are not sufficient safeguards.

The plotting code, intermediate data tables, and final figure should be independently checked.

Where a required value is unavailable, the workflow should identify the limitation rather than generate a plausible replacement.

The process of verifying AI-generated research figures should establish the provenance of every displayed value.

When AI Invents Weekly Student Performance Scores Between Two Actual Assessments

Step 1: Identify the Actual Observations

The verified mean pretest score is 61.0, and the verified mean posttest score is 78.0.

No weekly assessment scores were collected.

Step 2: Examine the AI-Generated Figure

The AI tool generates a line graph containing the following values:

Step 3: Identify the Misrepresentation

The graph suggests that the study documented a gradual increase in performance across the intervention.

However, the actual data establish only the observed pretest and posttest summaries.

The intermediate trajectory was not measured.

Step 4: Correct the Visualization

The researcher removes the invented weekly observations and creates a figure displaying the two actual assessment occasions.

If participant-level paired data are available, a paired-point plot or change-score distribution may communicate the observed changes more informatively.

Step 5: Consider Whether Interpolation Would Be Appropriate

The researcher could create a clearly identified illustrative interpolation for a methodological purpose, but it should not be represented as observed weekly performance.

With only two measurement occasions, the actual pattern of change between assessments remains unknown.

Step 6: Revise the Interpretation

The researcher reports that mean performance was higher at posttest than at pretest.

The manuscript does not claim that performance increased steadily throughout the intervention because the study did not collect evidence supporting that trajectory.

The example illustrates how AI can create a plausible narrative of change that exceeds what the research actually observed.

Common Misconceptions About AI-Invented Data in Research Graphs

How to Prevent Invented Values From Entering Research Visualizations

Begin by defining which quantities the study actually observed and which, if any, were derived through legitimate analytical procedures.

Every plotted value should have an identifiable evidential or computational origin.

A Practical Prompt for Preventing AI-Invented Graph Data

The prompt establishes useful constraints, but it cannot guarantee that a generative model will follow them correctly.

Researchers should inspect the plotting data, verify calculations independently, and confirm that no unsupported values appear in the final figure.

Before Publishing an AI-Generated Visualization, Verify the Origin of Every Value

Frequently Asked Questions About AI-Invented Research Visualization Data

AI Can Complete a Graph Without Completing the Evidence

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.