Data and Using AI for Analysis: What to Do and What to Keep in Mind
A practical guide for researchers and Institutional Research (IR) professionals navigating the AI-assisted analysis landscape
As universities increasingly provide access to large language models (LLMs) for faculty, students, and staff for learning, research, administrative work, and teaching, it becomes important to understand what AI models can help with and what we should keep in mind while using them.
As AI tools become standard infrastructure at universities, the IR and data professionals who work with multiple expansive datasets must be at the forefront of exploring these emerging technologies while also utilizing them effectively. The integration of AI into data analysis workflows has moved from novelty to near-expectation in a remarkably short time. LLMs can now summarize and interpret datasets, generate code, flag anomalies, and draft narratives around findings with varying levels of competence, i.e., tasks that once consumed hours of human effort can now be performed by AI tools with reasonable reliability. But this capability comes with a set of risks that are easy to underestimate precisely because the outputs look authoritative. For professionals whose credibility rests on the integrity of their analysis, understanding both what AI does well and where it systematically fails and underperforms is a core competency. This skillset will become even more important as we move ahead and LLMs and related technology develop further along with other emerging technologies.
What LLMs Do Well in Analytical Workflows
Accelerating code generation and debugging: AI tools can be useful for translating analytical intent into working code. Whether you are writing SQL joins across multiple tables, building a data cleaning pipeline in Python, or drafting an R script for a statistical model, AI can produce a working first draft fast and give you a starting point. But the key phrase is “first draft.” The code is a starting point and not a final deliverable. Professionals who treat generated code as production-ready without review are not saving time. They are simply increasing or deferring the risk. Especially in our field where data has to be reported with highest levels of accuracy.
Augmenting exploratory data analysis: In the early stages of a project, understanding the shape of a dataset, identifying missing patterns, and generating summary statistics are some steps that AI tools can help with and compress the cycle considerably. Asking a model to describe a dataset's properties or to flag inconsistencies in the data is a reasonable use. These are tasks where AI excels and where the cost of an error is relatively low because the analyst is still in a discovery mode. Thus, utilizing an LLM can considerably help speed up the starting discovery work for a dataset.
Navigating and reformatting provided sources: Many AI tools allow you to upload documents, datasets, or reference materials directly, enabling more targeted analysis grounded in sources you have specified rather than the model's general training. This is useful when you need to extract structured information from unstructured inputs, for example, pulling consistent fields from a set of documents, reorganizing survey codebooks, or summarizing policy documents into comparable formats for a committee. The ability to ask a model to reformat or synthesize across provided materials can compress what would otherwise be a time-consuming manual aggregation task. As with other AI-assisted steps, the analyst should verify that the output accurately reflects the source.
Drafting interpretive narrative: AI is effective at translating numerical findings into plain language. When results are complex and the audience is non-technical, using AI to generate an initial interpretive paragraph, one that the analyst then revises against the actual output, can improve communication quality. Please note that this is distinct from asking AI to generate findings. The analyst must always control the substance while they utilize AI to assist with the expression, drafting for clarity and editing the language.
Literature synthesis and conceptual framing: For researchers situating their work in existing scholarship, AI tools can help map conceptual territory, identify theoretical frameworks, and help draft literature review scaffolding. This will work best when the analyst has deep domain knowledge so they can catch errors, which are more common than one would think, particularly on narrow or specialized topics. However, researchers should remember that although LLMs can summarize large quantities of text, they shouldn't see those summaries as replacing the need to actually read primary source material. But they can be very helpful tools in your work.
Where LLMs Lag, and Why It Matters
AI does not understand your data (in the human sense): A language model has no access to your actual dataset during a conversation unless you provide it directly. When you paste summary statistics or describe your variables, the model is reasoning from that description, not from the data itself. Unless provided access to the underlying data or code, the model cannot detect a coding error in your categorical variable, notice that your "continuous" measure is actually truncated, or flag that your response rate suggests systematic nonresponse bias. These are things a skilled analyst catches by looking at the raw data. LLMs cannot substitute for that. This is an important limitation to keep in mind.
Context collapse across datasets and domains: Models trained on broad corpora have absorbed conventions from many different analytical traditions. When you ask about model fit, the model does not know whether you are working in structural equation modeling, item response theory, or machine learning and the standards differ substantially across those traditions. Without careful prompting that specifies your domain and methodology, you are likely to receive advice that is generic, misapplied, or drawn from a different field's conventions.
Hallucination in statistical interpretation: AI models frequently produce plausible-sounding but incorrect interpretations of statistical outputs and sometimes concepts too. A model may confidently misstate what a particular output of confidence intervals means, conflate statistical significance with practical significance, or describe a model fit index using the wrong threshold. The outputs are fluent and formatted in the register of expertise, which makes them harder to catch than an obviously wrong answer. Don’t be surprised when you get three different outputs when you run the same analysis three times depending on your prompts and how well they are written. Professionals who do not independently verify AI-generated statistical claims using primary sources, textbooks, or their own expertise are exposing themselves to serious risk of generating incorrect conclusions. One should keep in mind that models can misread, omit, or subtly mischaracterize content in uploaded documents, particularly when the source material is dense or inconsistently formatted. As good practitioners of data science skills, we must always strive for the highest levels of accuracy.
Spurious precision and false confidence: As the users of LLMs and our present cohort of LLM product offerings may have noticed, AI outputs tend toward the definitive (in the form of declarations). When asked to interpret a finding, a model will typically give you an answer rather than a range of plausible answers, even when uncertainty is warranted for the task at hand. In domains where effect sizes, confidence intervals, and replication uncertainty genuinely matter like clinical research, policy analysis, psychometrics, and financial modeling, this tendency of LLMs towards false confidence is something an analyst should always keep in mind. In many ways the analyst's job is to reintroduce appropriate epistemic humility that the AI has flattened out. Be a skeptical editor in this case, it will prove to be immensely helpful.
Operational Principles for AI-Assisted Analysis
Verify Data Governance, Security and Compliance Before Uploading: Never upload sensitive institutional data into public or non-enterprise AI tools as it may violate data security laws and the university data governance policies. When possible, work with de-identified or aggregate data.
Any work being done with datasets should be in compliance with IRB protocols as well as with IT security. Be equally cautious about what you include in prompts. Any identifiable details entered as text carry the same compliance risk as uploaded files. Data security is paramount. Always use tools that have been vetted by your technology and compliance teams, and direct any questions about approved tools to those teams before proceeding.
Treat AI as a junior analyst/intern, not a peer reviewer: A useful mental model for this is that: AI can do the work of a capable but untrained assistant who reads fast, writes clearly, and has limited to no domain expertise based on the kind of data it was trained upon. You would not let such an assistant sign off on your analysis. You would use their work as raw material that you review, correct, and take responsibility for. The same standard applies here in most cases.
Never outsource your cross-checking of facts and claims to AI: If a specific number, coefficient, or finding is going in a report, verify it against your own software output. Do not rely on AI to confirm that a result is correct. AI will often agree with a plausible-sounding claim even when it is wrong. A phenomenon sometimes called “sycophancy” [1] [2]. The source of truth for a statistical result is the output file from your analysis software, not a conversation thread. Always keep that in mind.
Prompt with specificity and domain context: The quality of AI output scales directly with the quality of the prompt. Generic questions produce generic answers. Effective prompting for analytical tasks includes: the software and version you are using, the specific method and model type, the population or sample, the dependent and independent variables with their measurement properties, and the specific decision or interpretation you need help with. A prompt that contains this information will return substantially more useful output than one that does not.
Always maintain a clear audit trail: This is one of the best practices any practitioner can develop while using AI based LLMs in their analysis. In any setting where your analysis may be reviewed, be it peer review, accreditation, litigation, audits, you need to be able to defend every analytical decision. AI-assisted work is no exception. Document which tasks AI supported in your workflow, what prompts you used for the steps, what you reviewed, and what you changed during the process. Think of this as professional accountability and equal parts caution. After all, at the end, after everything is said and done the analyst, not the AI, is responsible for the work. You own the analysis that you put out. The tools matter less than the analyst's responsibility for the final product. Increasingly, journals and publishing organizations are asking authors to acknowledge AI use and to explain which tools they used in the process and the extent of the use in the final outcome. As AI usage increases we will see more and more avenues asking for such disclosures to be made.
Conclusion
The professionals who will use AI most effectively are those who are already analytically rigorous— people who know their data, understand their methods, and can detect a wrong answer when they see one. For those analysts, AI can prove to be a genuine force multiplier.
For those who are not, AI can create a new category of risk: Using AI tools haphazardly can lead to outputs that are well-formatted and appear correct but contain errors, some small, some potentially huge. The tools used by professionals may have changed, but the burden of expertise has not.
References:
[1] Fanous, A., Goldberg, J., Agarwal, A., Lin, J., Zhou, A., Xu, S., ... & Koyejo, S. (2025, October). Syceval: Evaluating LLM Sycophancy. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (Vol. 8, No. 1, pp. 893-900).
[2] Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L., & Jurafsky, D. (2025). ELEPHANT: Measuring and understanding social sycophancy in LLMs. arXiv preprint arXiv:2505.13995.
Arun Sharma is Reporting and Systems Manager at Lynch School of Education and Human Development, Boston College.