AI in social sciences: data analysis, education, and psychology

A sociologist analyzing 50,000 qualitative interviews spends months on thematic coding. A language model completes an initial pass in hours. The question is no longer "whether AI is suitable for social sciences," but "where the model's utility ends and the researcher's responsibility begins."

At Cashcrown, we work with institutions that use AI for survey data analysis, educational support systems, and preliminary psychological assessment. Below, we describe what works, what fails, and where human-oversight is mandatory, not optional.

Social sciences operate on three types of data: text (interviews, open-ended surveys, social media), quantitative data (scales, demographic indicators, test results), and behavioral data (observations, interaction logs). AI has different levels of utility for each.

Text analysis. An LLM with a good system prompt classifies open-ended responses into thematic categories faster than two coders combined. We use a multi-stage approach: the model proposes codes, the researcher approves or adjusts the scheme, and the model codes the rest of the corpus according to the approved scheme. Key point: the researcher defines the codes, not the model.

Quantitative data. Classical machine learning (regression, random forests, gradient models) has been used in social sciences for years to predict social risk, model voting behavior, or segmentation. AI hasn’t invented anything new here. It accelerates iteration and improves explainability of models (SHAP, LIME) where results previously came from a "black box."

Behavioral data. Analyzing logs from educational platforms, response patterns in tests, or therapy session transcripts is an area where AI acts as a classifier for signals difficult to capture manually. However, this data is sensitive and subject to stricter GDPR requirements than a survey with declared consent.

Adaptive educational systems are one of the most mature areas of AI application in social sciences. The logic is simple: the model analyzes a student’s response sequence, identifies gaps, and adjusts subsequent tasks to the measured level of mastery.

In practice, these systems work well for tasks with clearly defined correctness (math, grammar, foreign language vocabulary). They struggle with assessing open-ended reasoning, argumentation, or creativity. A machine error in this area is costly: a student labeled as "weak" in reasoning may stop trying.

The biggest risk in educational systems is reinforcing inequalities. A model trained primarily on data from well-equipped institutions may systematically underestimate students from different cultural or linguistic backgrounds. An audit for demographic drift is mandatory before deployment, not after.

Clinical and research psychology is an area where AI generates both the highest expectations and the greatest concerns. Models processing data from questionnaires (e.g., PHQ-9, GAD-7), interview transcripts, or speech patterns can flag disorder risks with effectiveness comparable to a clinician’s preliminary screening. This has real value in contexts where specialist access is limited.

Several limitations that cannot be overlooked:

Models hallucinate. In a medical or psychological context, misclassification carries a different weight than in a corporate knowledge management app. A system that incorrectly labels someone as "low risk" may delay intervention.

Training data from clinical studies is historically unrepresentative. Overrepresentation of certain demographic groups in training sets translates to poorer model performance for others.

The AI Act in Poland (as part of EU regulations) classifies AI systems used for psychological assessment and health diagnostics as high-risk systems (Annex III). This means mandatory registration, documentation, and human oversight for every decision affecting an individual.

AI in social sciences does not replace methodological rigor. It enhances capabilities but also amplifies weaknesses.

Lack of causal model. Correlation in survey data between variable X and Y is not implication. The model cannot distinguish causal from confounding relationships. Interpretation lies with the researcher.

Inherited biases. The internet corpus used to train most LLMs reflects representation inequalities: geographic, linguistic, demographic. Analyzing social phenomena through such a model reproduces these inequalities unless active mitigations are applied.

Reproducibility. Generative models with default randomness yield different results for the same input. The reproducibility standard in social sciences requires a deterministic environment: fixed seed, model versioning, prompt logging. Without this, results are unrepeatable and thus unfalsifiable.

Responsibility for conclusions. Regardless of whether a thesis comes from the researcher or the model, the researcher signs off on the article. No academic publisher accepts AI as a co-author responsible for claims. This means verifying every key conclusion before including it in the manuscript.

The pattern for designing analytical systems for research institutions includes three human intervention points in the loop: approving the coding scheme before mass analysis, verifying a sample of results before generalization, and ethical review before publishing sensitive personal data.

GDPR imposes special requirements on sensitive data, which includes data on mental health, ethnic affiliation, or political views. Social research often touches precisely these categories.

Four questions to ask before starting any AI research project:

Can the data be used to identify a person, even indirectly? Declarative anonymization is insufficient for high-dimensional data.

Will the model make decisions affecting individuals (risk assessment, selection, recommendations)? If so, a Data Protection Impact Assessment (DPIA) and a legal basis for processing are required.

Do participants understand that their data will go to an AI model? Informed consent for algorithmic analysis is not the same as consent to participate in a survey.

Will the model’s results be presented as conclusions without indicating uncertainty? Omitting model uncertainty in a report is a methodological and potentially ethical error.

More on responsible system architecture in AI and responsible innovation and on the role of humans in the loop.

Not methodologically. AI can perform an initial thematic coding pass, significantly reducing work time, but the researcher must define and approve the coding scheme. The quality of the resulting analysis depends on prompt quality and critical evaluation of the model’s proposals. Treating the model’s output as a ready-made analysis without validation is a methodological error.

Anonymized data (not pseudonymized) with aggregate assessments, without the possibility of linking to a specific individual. Identifiable data concerning mental health requires a DPIA, a legal basis under Article 9 of GDPR, and typically approval from an ethics committee. AI systems making individual decisions (diagnostic, therapeutic) are classified as high-risk under the AI Act.

Three steps: audit the training set before deployment (which groups are overrepresented), validate the model’s results on demographic subsamples (whether errors are evenly distributed), and monitor results post-deployment for systematic deviations. This is not a one-time step but a continuous process.

The AI Act distinguishes system risk, not intent of use. AI systems used for psychological assessment, educational selection, or social risk evaluation are classified as high-risk. They require registration in the EU database, conformity assessment, technical documentation, and post-deployment oversight. Systems that only assist in literature searches or hypothesis generation without making decisions about individuals have lighter requirements.

Major publishers (Nature, Science, ICMJE) do not accept AI as a co-author. The standard is a declaration in the Methods section: which tool, for what stage, and how results were verified. Prompt logs and model outputs should be part of the research documentation available to reviewers. The researcher is responsible for every claim in the work, regardless of its source.

More on the limits of language models in research: LLM as a hypothesis generator and AI as an autonomous scientist. We discuss the future of research work separately.