Should students still learn SPSS when AI can analyse data?

Give a modern AI tool a spreadsheet and a sentence such as "compare anxiety scores before and after the intervention", and it may return a test, a chart, a p-value and a polished paragraph within seconds. SPSS appears to demand much more ceremony. You have to define variables, inspect the data, choose a procedure, decide which options to request and then work through an output window full of tables.

It is reasonable for students to ask whether learning SPSS is becoming the statistical equivalent of learning to use a fax machine. The answer depends on what "learning SPSS" means. Memorising the location of every menu is becoming less valuable. Understanding what the software is doing, why a test is appropriate and how to recognise a misleading result is becoming more valuable, not less.

The calculation has rarely been the hardest part of statistics. SPSS has been calculating test statistics faster and more accurately than students for decades. AI changes the interface: instead of clicking through menus or writing commands, you can describe what you want in ordinary language. It does not remove the need to know whether what you asked for makes sense.

There is also no longer a clean divide between "traditional software" and "AI". IBM's current SPSS Statistics v32 includes an AI Output Assistant that provides plain-language insights from selected results. The likely future is not SPSS on one side and AI on the other. It is statistical software, code and conversational assistance increasingly working together.

The real question is whether you can audit the answer

An AI tool can often produce an analysis. A student still has to answer three different questions:

• Was the right analysis requested? This depends on the research question, design, variable types and assumptions.
• Was it carried out correctly? This includes data cleaning, coding, exclusions, missing values, transformations and the exact options used.
• Does the interpretation follow from the result? A statistically significant association is not automatically important, causal or useful.

SPSS does not answer those questions for you either. It will calculate a mean for a numeric identification number, include 999 as a real questionnaire score if you have not defined it as missing, and produce an impressive regression table from a badly designed model. AI can make the same mistakes more fluently because it can add a convincing explanation around them.

The most dangerous analysis is often not one that crashes. It is one that runs, produces a neat table and answers the wrong question.

That is why the useful comparison is not "slow SPSS versus clever AI". It is between three interfaces for doing statistical work: clicking through SPSS, using SPSS syntax and asking an AI system to generate or execute an analysis. Each has strengths, and each can conceal different kinds of error.

Three ways to analyse the same question

Imagine that 120 students complete the same anxiety scale immediately before and after a six-week study-support programme. Each row contains one student, with separate variables for the pre-programme and post-programme score. The immediate question is whether average scores changed among those students.

Clicking through SPSS

A student might open Variable View, check that both score variables are numeric and treated as scale variables, inspect frequencies and boxplots, then use the paired-samples t-test dialogue. The interface makes the available choices visible and produces standard output without requiring programming knowledge.

This route is useful for learning because it slows the process down. You can see that the procedure asks for a pair of variables rather than two unrelated groups. You can request confidence intervals and inspect descriptive statistics. However, it is also easy to press buttons until output appears, particularly when a worksheet gives a menu path without explaining the design.

Using SPSS syntax

The same analysis can be expressed as saved commands:

T-TEST PAIRS=anxiety_pre WITH anxiety_post (PAIRED) /CRITERIA=CI(.95) /MISSING=ANALYSIS.

Syntax creates a record of what was run. It can be commented, corrected and rerun when the data change. A lecturer, supervisor or future version of you can inspect it without trying to reconstruct a chain of clicks. IBM's documentation confirms that the Paste button in SPSS dialogue boxes generates syntax from the selected options, so students do not have to choose between using menus and writing every command from memory.

Asking an AI tool

A code-executing AI tool may read the spreadsheet, generate Python or R code, run a paired t-test, create a plot and draft an interpretation. A text-only model may instead suggest code or instructions without genuinely executing anything. The difference matters: students should establish whether the numbers came from an actual calculation on the uploaded file or from a language model attempting to reason them out in text.

The AI route can be much faster, but only if the prompt contains the design information the software needs. "Compare the before and after scores" does not explicitly say that the observations are paired, that the same people supplied both scores, which values indicate missing responses, whether a high score means more anxiety, or whether the scale has been scored correctly. An AI may infer some of this, but an inference about the design is not a substitute for knowing the design.

Even a flawless paired t-test would not show that the programme caused any change. Without a comparison group, other explanations remain possible, including the passage of time, changes elsewhere in students' lives, regression to the mean and the effects of completing the measure twice. This limitation belongs to the study, not to SPSS or AI.

What clicking through SPSS still teaches well

SPSS remains a useful teaching tool because its Data View and Variable View make the structure of a dataset unusually visible. Students can see cases in rows, variables in columns, value labels, declared missing values and measurement levels. The dialogue boxes expose many of the decisions that a natural-language request can hide.

For a beginner, this reduces the extra difficulty of learning programming at the same time as statistics. A psychology student can concentrate on the difference between an independent and paired t-test without first learning objects, packages and function arguments in R. A nursing student can examine frequencies and cross-tabulations without debugging a misplaced bracket.

There are limits to this advantage. Following a menu path can become a ritual rather than an analysis. Students sometimes remember that "Analyze, Compare Means, Independent-Samples T Test" produced the desired table in a workshop but cannot explain why their own dissertation has repeated observations and needs a different approach. Menu locations also change between versions, while the underlying statistical ideas change much less.

A practical rule improves the click-through workflow considerably:

1. Use the dialogue box to explore the available options.
2. Choose Paste rather than OK.
3. Read the generated syntax and add a comment explaining why the analysis was chosen.
4. Run the syntax and save it with the data dictionary and output.

This approach preserves the accessibility of the graphical interface while creating an audit trail. It also begins to teach students how software represents an analysis, which makes later use of AI-generated code much easier to check.

Why syntax matters more in an AI-assisted world

Syntax is sometimes presented as the advanced, intimidating alternative to clicking. In practice, its main value is mundane: it records the work. If a dissertation dataset is updated from 183 to 201 participants, saved syntax can rerun the cleaning, scoring, descriptive analysis and models consistently. A manual workflow may require dozens of remembered actions.

Syntax is particularly valuable for transformations. Suppose a wellbeing scale contains four positively worded items and two reverse-worded items. A final score is only meaningful if those two items are reversed correctly before the items are combined. A syntax file can show the recoding rule, the treatment of missing items and the formula used for the total. A screenshot of the final reliability table cannot.

It also allows the analysis to be challenged. A supervisor can notice that a filter remained active, that listwise deletion removed an unexpected number of cases, or that the reference category in a regression was not the one described in the report. Reproducibility does not guarantee correctness, but an error that can be seen and rerun is easier to correct than one buried in a sequence of undocumented clicks.

AI makes syntax less intimidating because students can ask for an explanation of each line or request a first draft. The important change is that generated syntax must be treated as a proposal, not as authority. Run it on a copy of the data, inspect warnings, compare the variables with the data dictionary and confirm that the resulting sample size is plausible.

AI is genuinely useful for statistical work

A sensible warning about AI should not pretend that it has no value. Used carefully, it can remove a great deal of mechanical work and provide immediate explanations that were previously difficult to obtain outside a statistics class or supervision meeting.

Useful tasks include:

• turning a clearly specified analysis plan into draft SPSS, R or Python syntax;
• explaining an unfamiliar option or warning message in ordinary language;
• suggesting descriptive tables and graphs appropriate to known variable types;
• checking whether a written interpretation has confused a p-value, confidence interval or effect size;
• producing code to recode variables after the student has supplied the exact coding rules;
• generating a data dictionary template or a checklist of assumptions to investigate;
• translating a procedure from SPSS syntax into R or Python so the result can be independently reproduced;
• asking questions about anomalies, such as an unexpectedly small analysis sample or an implausible coefficient.

AI can also be a patient tutor. A student can ask why a repeated-measures design violates the independence assumption of an ordinary independent-samples test, request three examples, and then test their understanding on a new scenario. That is different from asking the tool to choose a test and submitting whatever it says.

The best prompts preserve the student's responsibility. "Tell me which test to use" invites a verdict. "List the plausible tests, state the conditions under which each would be appropriate, identify information that is still missing and do not generate code yet" invites scrutiny.

Plausible code is not the same as a correct analysis

Recent research helps explain why fluency can be deceptive. The StatLLM project, published in 2026, assembled 207 statistical tasks across 65 datasets and included human evaluation of code generated by three language models. The tasks ranged from descriptive work and hypothesis tests to regression, survival analysis and non-parametric methods.

In the associated evaluation, the models' average code-quality score was 94.1 per cent of the available points. That sounds excellent. The average fell to 61.2 per cent for executability and 51.7 per cent for output quality. These are scores across evaluation criteria rather than a simple percentage of tasks passed, but the pattern is revealing: code could look well structured and statistically plausible while failing to run reliably or produce the correct result.

A separate 2024 benchmark, StatQA, contained 11,623 examples designed to test method selection, relevant-column selection and the applicability of statistical procedures. The best reported GPT-4o result was 64.83 per cent. The researchers found that language models' main errors concerned whether a method was applicable, whereas human participants more often confused one statistical task with another.

Neither figure should be treated as the permanent accuracy of every AI tool in 2026. Models improve, the benchmarks used particular systems and tasks, and specialist tools may perform differently. The findings nevertheless identify a durable risk: generating credible statistical language and code is easier than deciding whether a method fits the design and data.

This is exactly where students need statistical education. AI can increasingly handle computation. The user still needs enough knowledge to notice that a model has treated paired observations as independent, selected the wrong outcome column, ignored clustering, read a category code as a quantity or interpreted association as causation.

Variable types are not administrative details

Many weak analyses begin before a test is selected. SPSS distinguishes the stored data type from the variable's measurement level. Its current documentation uses nominal, ordinal and scale measurement levels. These labels are not decoration: they describe what the values mean and which summaries or procedures may be sensible.

A number is not necessarily a quantity

A participant identifier such as 10457 may be stored as a number, but averaging identifiers is meaningless. The values 1, 2 and 3 might represent biology, law and history students; the distance between 1 and 2 has no quantitative meaning. These are nominal variables even though the cells contain digits.

AI sometimes has an advantage here because variable names and labels provide clues. A column called "student_id" is unlikely to be a genuine scale. It still cannot be relied upon to infer an undocumented coding scheme accurately, especially when names such as "group", "status" or "score2" reveal little.

Order does not always imply equal spacing

Degree classifications, satisfaction categories and stages of disease have an order, but the gap between adjacent categories is not necessarily equal. They are ordinarily treated as ordinal. A single five-point agreement item also has ordered responses, although researchers sometimes treat a total or mean formed from several well-designed items as approximately continuous. That decision requires methodological justification; changing the SPSS Measure column to "Scale" does not create interval-level measurement.

Binary outcomes change the model

Pass or fail, employed or unemployed, and condition present or absent are categorical outcomes. A linear regression may produce numbers for a 0 and 1 outcome, but logistic regression is generally designed for modelling a binary response. The output then uses log-odds and odds ratios, which require a different interpretation from an ordinary regression coefficient.

Dates, durations and units need attention

A date is not an ordinary category and a duration is not necessarily stored in the unit a student assumes. A response-time column recorded in milliseconds can be misread as seconds. Age at enrolment may have been calculated at different dates. AI can detect formats and suspicious ranges, but it cannot recover a unit or definition that was never documented.

Missing values can change the result dramatically

Survey exports often use values such as 9, 99 or 999 for "not applicable" or "prefer not to say". Unless they are declared as missing or recoded, software may treat them as genuine observations. IBM's SPSS documentation provides for user-defined missing values, but the user has to specify them.

An AI tool may notice that 999 is an outlier and ask whether it represents missing data. It may also delete it automatically, even though 999 could be a valid value in another variable. The data dictionary, questionnaire and collection process are the authority, not the model's guess.

Test selection starts before any software is opened

A test should follow from the research question and design, not from whichever variable happens to be highlighted in SPSS. Before opening a menu or uploading a file, write down:

1. The question. Are you describing a sample, comparing groups, examining change, estimating an association or predicting an outcome?
2. The outcome variable. Is it continuous, ordinal, binary, a count, a time-to-event measure or something else?
3. The explanatory variable or grouping factor. Is it categorical or continuous, and how many groups or levels are involved?
4. The relationship between observations. Are groups independent, are the same people measured more than once, or are students clustered within classes, wards or universities?
5. The intended population. What sampling method was used, and to whom can the findings reasonably be generalised?
6. The likely distribution and sample size. Are there severe skew, outliers, sparse categories or very small groups?
7. Potential confounders. Is a simple comparison enough, or does the question require an adjusted model?

Only then do familiar test names become useful. Two independent groups with a roughly continuous outcome may suggest an independent-samples t-test, often with the unequal-variance or Welch version considered. Measurements on the same people before and after an intervention suggest a paired test. Three or more independent groups may lead to ANOVA or an alternative designed for unequal variances or non-parametric comparison. Two categorical variables may call for a chi-square test, with Fisher's exact approach considered when cells are sparse.

Those are starting points rather than a test-selection vending machine. A Mann–Whitney test, for example, is not simply "the non-parametric t-test" in every setting, and a Spearman correlation does not solve every problem involving an ordinal variable. The precise hypothesis and shape of the distributions affect what the result means.

Assumptions are part of the analysis

Students are often taught assumptions as a checklist to complete after choosing a test. In reality, some assumptions come from the study design and cannot be repaired by running another SPSS table.

Independence

Independent observations must genuinely be independent. If each student contributes several rows, patients are treated within the same ward, or pupils are sampled within classes, ordinary methods may underestimate uncertainty because observations within a cluster can resemble one another. A normality test cannot diagnose this. The analyst needs to know how the data were collected.

Normality

The relevant normality assumption is frequently misunderstood. For a paired t-test, attention falls on the distribution of the within-person differences, not on whether each raw variable is perfectly normal. In linear regression, the assumption concerns model errors or residuals rather than requiring every predictor to be normally distributed.

A single significance test of normality should not become an automatic gatekeeper. With a large sample, a minor and unimportant departure may produce a small p-value; with a small sample, serious non-normality may be difficult to detect. Plots, outliers, sample size, the robustness of the intended method and the purpose of the analysis all contribute to the decision.

Equal variances

The conventional equal-variance independent t-test assumes the population variances are the same. IBM's own documentation lists independence, normality and equal population variance among its assumptions. If group variances differ, an unequal-variance procedure may be more appropriate than automatically abandoning all parametric analysis.

Linearity and influential observations

Pearson correlation and ordinary linear regression concern linear relationships. A strong curved relationship can produce a weak linear correlation. One extreme observation can also dominate a coefficient. A scatterplot and residual diagnostics may reveal more than the final p-value.

Expected counts

A chi-square table can contain categories with very few observations. The issue is not simply the total sample size but the expected counts across cells. Combining categories solely to obtain significance is not defensible; any regrouping should make substantive sense and preferably be planned before examining the result.

Missingness

Complete-case or listwise analysis may silently remove anyone missing one variable in a model. The remaining sample can be much smaller and systematically different from the original participants. Students should report how many cases entered each analysis and investigate patterns of missing data rather than accepting the default without comment.

SPSS output does not interpret itself

A standard SPSS output file can be dozens of pages long. The skill is not copying every table into an assignment but identifying the pieces needed to answer the question:

• the number of observations included and excluded;
• descriptive statistics for the relevant groups or time points;
• the estimated difference, association or model coefficient;
• an effect size in a form appropriate to the analysis;
• a confidence interval showing the range of values compatible with the data and model;
• the test statistic, degrees of freedom and p-value where relevant;
• diagnostics and assumption checks;
• model-fit information and sensitivity analyses where the method requires them.

The p-value is particularly prone to confident misinterpretation. The American Statistical Association's statement stresses that a p-value does not tell you the probability that the studied hypothesis is true.

"A p-value, or statistical significance, does not measure the size of an effect or the importance of a result."

A result of p = .03 is not a 97 per cent probability that the alternative hypothesis is correct. A result above .05 is not proof that there is no difference. The estimate might be genuinely close to zero, or the study might be too imprecise to distinguish a meaningful effect from no effect. The effect estimate and confidence interval are needed to tell those possibilities apart.

Interpretation also depends on coding. A negative regression coefficient may indicate a lower outcome for the category coded 1 than for the reference category coded 0, but reversing the reference group reverses the sign. An odds ratio above 1 can sound alarming until the outcome and reference category are checked. An R-squared value describes variation accounted for under a model; it does not establish causation or prove that the model predicts well in new data.

AI can explain all of these statistics clearly, but it can also write an incorrect explanation with excellent grammar. Students should trace every sentence back to a specific output value and ask whether the wording goes beyond what the design supports.

An analysis can be technically correct and still mislead

Suppose a university compares the exam marks of 600 students who did and did not use an optional tutoring service. An AI tool runs an independent-samples t-test correctly and finds that tutoring users had lower average marks. The code executes, the p-value is accurate and the output agrees with SPSS.

It would still be reckless to conclude that tutoring reduced attainment. Students who were struggling may have been more likely to seek help. Prior attainment, subject, year of study, attendance and other factors could differ between groups. An adjusted regression may address some measured differences, but an observational comparison still does not automatically estimate the causal effect of tutoring.

This example shows why statistical judgement is not merely error checking. The wrong conclusion can follow from correct arithmetic. Neither SPSS nor AI knows the institutional context unless it is supplied, and no software can transform weak observational evidence into a randomised experiment.

The same problem appears in dissertation data. A correlation between social-media use and anxiety does not show which came first or whether another factor contributes to both. A statistically significant improvement among people who completed a programme may ignore those who dropped out. A model based on repeated weekly responses may overstate precision if it treats every response as independent.

Data cleaning is part of the analysis

Students often imagine that analysis begins once the dataset is "clean". In practice, cleaning decisions determine what is analysed. A credible workflow records how the following were handled:

• duplicate participant records;
• values outside the permitted range;
• inconsistent category labels and spelling;
• reverse-scored questionnaire items;
• units, dates and calculated durations;
• participants who were ineligible or withdrew;
• missing responses and partial scale completion;
• outliers and influential cases;
• active filters, split files and weighting;
• changes made after looking at the results.

AI can help write cleaning syntax and flag suspicious values. It should not quietly decide that an inconvenient observation is an error. An extreme value may be a data-entry mistake, a genuine rare case or evidence that the assumed model is unsuitable. The correct response depends on source records, collection rules and a defensible analysis plan.

Keep the raw file unchanged. Create a working copy, save syntax for every transformation and produce checks after major recodes. If a five-item scale should range from 5 to 25 and the computed scores range from 4 to 125, that check has found a problem before a sophisticated model makes it harder to see.

Privacy and assessment rules can limit what you upload

A dissertation spreadsheet may contain ages, diagnoses, interview identifiers, placement details, free-text responses or combinations of variables that make people identifiable. Removing names does not always make data anonymous. Before uploading participant-level data to an external AI service, students need to check the consent materials, ethics approval, data-management plan, university policy and the tool's terms.

UK Research and Innovation's policy states that sensitive or personal data belonging to others must not be entered into a generative AI tool without formal consent. The Information Commissioner's Office guidance on AI and data protection also emphasises lawful processing, security and data minimisation.

Practical alternatives include using an institutionally approved tool, analysing a synthetic or heavily anonymised teaching dataset, asking for code using only a variable dictionary, or running generated code locally without sending the data to the model. Health data, ethnicity, political opinions and several other categories receive particular legal protection and should not be treated casually because an upload button is convenient.

Assessment rules also vary. A module may permit AI to explain syntax but prohibit generated interpretation, or require any use to be declared. A student who could legitimately use AI in one assignment may breach the rules in another. Keep a record of prompts, generated code, changes and verification where the institution requires it, and never invent an analysis process that did not occur.

How SPSS support should connect method to evidence

Students using SPSS in psychology, health, business, education and the social sciences frequently have to connect a research question with data preparation, test selection, assumptions, output and a defensible written interpretation. Those stages are also where specialist SPSS assignment help belongs naturally: not as a mysterious route to a p-value, but as support with understanding what the procedure does, why it fits the design and how the evidence should be reported.

Good statistical support should make the work more transparent. It should identify the variables used, explain exclusions, show the syntax or reproducible steps, distinguish required checks from optional output and avoid claiming more than the study design allows. A page of unexplained SPSS tables is not meaningful evidence of understanding, however technically polished it looks.

What universities should teach now

A statistics module built entirely around memorising menu routes is increasingly difficult to defend. Students need enough familiarity with SPSS to inspect data, manage variables, run common procedures and interrogate output, but the curriculum should spend more time on the decisions that remain difficult to automate.

A useful modern sequence would include:

• statistical foundations, including study design, variable types, sampling, uncertainty and the meaning of common estimates;
• a graphical interface, so beginners can see data structures and procedure options without a programming barrier;
• syntax and reproducibility, including comments, saved transformations and rerunning an analysis on updated data;
• AI literacy, including prompt design, verification, privacy, applicability errors and disclosure;
• interpretation and communication, with emphasis on estimates, effect sizes, confidence intervals, limitations and honest causal language;
• comparison across tools, so students can reproduce a small analysis in SPSS and another environment rather than treating one package as an oracle.

Assessment can change as well. Requiring a data dictionary, annotated syntax and a short decision log makes it harder to conceal a thoughtless analysis behind polished prose. Students can be asked to critique an AI-generated method, diagnose errors in an output interpretation or explain why two apparently reasonable procedures answer different questions.

A practical AI-assisted SPSS workflow

The following approach uses AI without surrendering the analysis:

1. Write the research question in plain English. State the population, outcome, comparison or relationship and whether the aim is descriptive, associational, predictive or causal.
2. Create a data dictionary. Record each variable's meaning, coding, unit, measurement level, permitted range and missing-value codes.
3. Protect the raw data. Work from a copy and decide whether the data can lawfully and ethically be shared with any external tool.
4. Inspect before testing. Run frequencies, descriptives and appropriate plots. Check duplicates, impossible values, group sizes and missingness.
5. Choose candidate methods on paper. Explain how the design and variables narrow the choices before asking software to calculate anything.
6. Use SPSS dialogue boxes to explore options. Paste the resulting syntax rather than relying on undocumented clicks.
7. Run the main analysis and diagnostics. Record the analysis sample, assumptions, warnings and any deviations from the original plan.
8. Ask AI to challenge the choice. Request alternative explanations, possible assumption failures and checks you may have missed.
9. Verify generated code line by line. Confirm variable names, category coding, filters, missing-data handling and model options.
10. Reproduce the key result. Where practical, compare AI-executed output with SPSS or rerun generated syntax locally. Investigate discrepancies rather than choosing the result you prefer.
11. Interpret estimates before p-values. Report direction, magnitude, uncertainty and practical importance, then place significance tests in that context.
12. Save the record. Retain syntax, output, cleaned data, a decision log and any required AI-use declaration.

A prompt that makes the AI show its working

A vague prompt such as "analyse my dissertation data" encourages the tool to fill gaps by guessing. A more useful request might read:

I have one row per participant. The outcome is a 0–40 anxiety total, where higher values mean greater anxiety. The same 120 participants completed it before and after a programme. The value 999 means missing. First identify the design and list plausible analyses. State the assumptions and information you still need. Do not run a test or claim causation yet. Then provide SPSS syntax only after explaining why the selected procedure fits.

This prompt does not guarantee a correct answer. It makes hidden assumptions easier to see and separates method selection from computation. You can improve it further by stating how the scale was validated, how much missing item data is permitted, whether there is a control group and whether observations are clustered.

How to inspect an AI-generated analysis

Before using the result, ask:

• Did the tool analyse the intended file, sheet, rows and variables?
• Did it execute code, or merely describe what the output might be?
• How were missing values, duplicates, filters and excluded cases handled?
• Were category labels and reference groups interpreted correctly?
• Does the procedure match the outcome, design and dependence between observations?
• Were assumptions checked in a way relevant to the actual model?
• Are the sample size and descriptive statistics consistent with SPSS?
• Can the code be saved and rerun to obtain the same result?
• Does the written conclusion match the coefficient, confidence interval and p-value?
• Has association been described as causation, or non-significance as proof of no effect?
• Does the answer acknowledge study limitations rather than attaching them as a generic final sentence?

One of the strongest checks is independent reproduction. If SPSS syntax and AI-generated Python code use the same cases, variables, transformations and method, key estimates should agree apart from documented differences in defaults or numerical precision. When they do not, the disagreement is useful evidence that something in the workflows differs.

Should students still learn SPSS?

Students should still learn SPSS when it is used in their course, discipline or intended research, but the goal should no longer be mastery of menu geography. They should learn enough to understand the structure of a dataset, define variables properly, run and diagnose common analyses, save reproducible syntax and read the output without treating it as a verdict.

AI is likely to make routine statistical work faster. SPSS itself is already moving in that direction. This does not make statistical knowledge redundant; it shifts the valuable skill from producing output to evaluating it. A graduate who can only follow menu instructions is vulnerable when the procedure changes. A graduate who can only prompt an AI is vulnerable when the answer sounds right but rests on a false assumption.

The strongest student can move between all three approaches. They can use the SPSS interface to inspect and explore, syntax to document and reproduce, and AI to explain, draft and challenge. Most importantly, they can say what the data show, what they do not show and exactly how they reached that conclusion.