AI and research data
-
AI tools in a research context
From transcription to analysis: An overview of AI tools and their effective use in research.
-
Data analysis with AI-tools
AI can support data analysis, but it cannot replace your part. What is possible, what isn’t, and what you need to bear in mind.
-
Toos for Data analysis
Which software is suitable for qualitative data analysis? An overview of specialised tools and their AI-powered features.
-
Data protection during data analysis
Personal data and AI tools: What is permitted, what isn’t, and how to stay on the safe side.
Transcription
Anyone who, as part of their research, collects oral data – for example, through interviews, focus groups or observations – or uses audio and video recordings for content analysis (e.g. media reports) must transcribe these after the data has been collected. The aim is to make oral information readable so that it can be subject to analysis. This does not mean to summarise it, but rather the reproduce the exact wording as faithfully as possible.
AI-powered transcription tools can significantly speed up the transcription process. Tools such as Whisper or u:ai automatically recognise spoken language and produce a transcribed text in a short period of time. Specialised research software such as MAXQDA or ATLAS.ti also offer integrated AI transcription functions that can be linked directly to the subsequent data analysis.
Please note, however, that AI-generated transcripts are rarely error-free and must always be corrected manually. Furthermore, data protection regulations apply as soon as personal data is entered into an external tool. See the section on data protection for more information.
Data analysis
Analysing research data is one of the most challenging stages in the scientific process. AI tools can assist in various ways. However, before you choose an AI tool for data analysis, it is worth considering the nature of your data, as Large Language Models are not well suited to every type of data.
- When dealing with qualitative data – that is, data in text form such as interview transcripts, open-ended responses or observation notes – AI tools can provide valuable support. Large Language Models have been trained primarily to process, understand and generate language. They can assist in developing categorisation systems and support reflection on one’s own interpretations. Qualitative analyses generally result in categorisation systems, typologies or summaries, supported by quotations.
- When it comes to quantitative data – that is, figures such as scale scores from questionnaires, results from experiments or demographic information – Large Language Models such as ChatGPT are less suitable. As they operate using tokens rather than exact numerical values, systematic errors can arise during calculations or counting tasks. However, LLMs can provide indirect support here by generating executable code in R or Python that actually performs the calculation correctly.
Analysing data using AI: possibilities and limitations
However, you cannot simply delegate the data analysis to an AI tool. This would be problematic for two reasons: it would mean you would not be contributing your own scientific work, and AI tools systematically make errors during the analysis. AI can, however, provide support alongside you during the analysis:
When used effectively, AI tools can provide support throughout the evaluation process:
-
Finding the right method
AI can provide suggestions for selecting suitable analytical methods, explain the requirements of a method, and compare the advantages and disadvantages of different methods.
-
Analysing data
AI tools can provide step-by-step instructions on how to apply a method correctly, generate executable code in R or Python for statistical analyses, and act as a methodological fact-checker to help critically assess one’s own interpretations.
-
Visualising data
AI can generate markup code for online editors such as Draw.io or Mermaid.live, as well as source code for visualisation libraries such as Vega-Lite.
The use of AI in data analysis carries risks that you should be aware of:
- Bias: Hidden biases in the model or the data can compromise the objectivity of your results.
- Lack of reproducibility: AI responses vary and cannot be replicated exactly. This makes it difficult to reliably repeat analyses.
- Hallucinations and calculation errors: AI can invent methods, sources or key figures, and sometimes produces systematically incorrect results in its calculations.
- The ‘black box’ problem: The decision-making process of an AI model usually remains unclear. This limits transparency and traceability.
You should therefore always scrutinise AI-generated results.
Even though AI tools can save you a great deal of work when it comes to analysis, the academic responsibility remains with you. In practical terms, this means that if you have your interview data analysed by an AI tool and the result is incorrect, it is your responsibility to identify and correct this. AI-generated results may contain biased, incomplete or simply incorrect information that nevertheless sounds convincing. You should therefore always check the results yourself – at least on a random basis – by analysing a portion of your data independently and comparing it with the AI output.
You should also keep a record of which AI tools you have used, when and for what purpose – for example, in a research journal. This makes your research process transparent and traceable, which is particularly important for seminar papers and dissertations.
Data analysis tools
For simpler analysis tasks, Large Language Models such as ChatGPT or u:ai can be a good place to start. They are suitable, for example, for helping to develop initial categorisation systems, group text passages by theme, or critically examining one’s own interpretations.
However, they are not sufficient for a systematic and scientifically rigorous qualitative analysis. To this end, the University of Vienna offers affordable licences for specialised tools that combine AI-powered functions with scientific workflows.
- MAXQDA (campus licence: 20.00 € per year) enables the analysis of texts, PDFs, images, audio and video data, as well as survey data. Data is tagged with custom-developed codes and categorised; coded passages can be retrieved across the entire dataset and tabulated or presented as a graph. The integrated AI function, AI Assist, automatically summarises documents or coded segments, suggests codes, codes automatically and answers questions via chat regarding your own data.
- ATLAS.ti (campus licence: €25.00 € per year) offers similar functions for texts, PDFs, images, audio and video files. Coded segments are visualised as networks or diagrams. The AI functions enable automatic coding based on your own codes or specified research objectives, as well as the automatic summarisation of documents and the ability to ask questions about your own data via chat.
Both tools can be ordered via the AcadCloud portal.
It doesn’t always have to be AI
For quantitative analysis, there is specialised software that has been used reliably for many years: Excel for simple calculations and frequency analyses, SPSS as graphical statistics software with an integrated scripting mode, and R and Python for complex and reproducible statistical analyses.
Data protection when using AI: What you need to know
Regardless of which tools you use for transcription and analysis: as soon as you work with personal or confidential data during your studies, data protection regulations apply, which you must be aware of and comply with. This applies, for example, to
- personal data from interviews, surveys or classes
- sensitive information that allows conclusions to be drawn about specific individuals, as well as
- confidential research data or internal documents.
In principle, the processing of personal data is prohibited under the General Data Protection Regulation (GDPR). There are, however, two exceptions:
- Consent of the data subject: If you conduct interviews, for instance, you must obtain the consent of the interviewees before collecting any data, by means of an informed consent form.
- Legal requirement: In certain cases, data processing is required by law. The use of the Moodle learning platform, for example, is covered by the University of Vienna’s privacy policy, which is why you, as a student, do not need to give separate consent.
In the context of student research, the collection of personal data is, in most cases, only permitted with the express consent of the data subjects. Furthermore, the data must not be disclosed to third parties.
This is precisely where the use of AI tools becomes problematic: as soon as you enter personal data into an AI tool, that data leaves your sphere of control. The tool’s provider gains access to the content, which is legally regarded as a disclosure to third parties. Without the explicit consent of the data subjects, this constitutes a breach of the GDPR. Furthermore, many tools – particularly free ones – process data on servers outside Europe. This means that data protection in accordance with EU standards is not guaranteed. In addition, the data entered is often reused to train new models, which amounts to a further disclosure to third parties without you or the data subjects being informed.
Data pseudonymisation
One practical step you can take to ensure data protection in your research is to pseudonymise the data you intend to analyse. This means replacing all details that could identify an individual with neutral codes. This allows you to process and analyse the data without it being possible to identify specific individuals. It is best to draw up a pseudonymisation plan:
- Assign codes: Replace all names and identifying details with neutral codes, for example B1, B2, B3 for respondents. Do not use initials or abbreviations that could reveal the identity of individuals.
- Create a code list: Draw up a separate table showing which code corresponds to which person. This table must be stored separately from the research data.
- Check contextual information: Check whether it is still possible to identify individuals from the context of individual statements, for example through a combination of occupation, age or place of residence. Remove or generalise such details where necessary.
- Apply codes consistently: Use the codes consistently throughout all documents: in the transcript, in the analysis and in the final thesis. Never use real names in external communications.
Important: Pseudonymised data is still considered personal data. The GDPR therefore also applies to pseudonymised data.
-
Checklist for data processing in AI tools
As a general rule, you should only process anonymised data in AI tools wherever possible. If anonymisation is not possible or if it cannot be ruled out that individuals can be identified – for example, in the case of voice recordings or qualitative interviews:
- Obtain consent for data collection.
- Specify in the informed consent form which AI tools you are using and for what purpose, for example, for transcribing interviews or analysing data.
- Check the tool’s server location: Is the data processed within the EU? If this is not the case, you must not, as a general rule, enter personal data into this tool, unless the tool’s terms of use expressly guarantee that data will not be stored or reused for training purposes.
- Check the terms of use: Does the tool guarantee that data will not be reused for training purposes?
- Clarify in advance how you will proceed if a person withdraws their consent.
Even if you do not enter any personal data, the issue of confidentiality remains: confidential information and unpublished research findings should, as a general rule, not be entered into an AI tool. An exception applies if you run a tool exclusively locally on your computer or if the tool’s terms of use guarantee that data entered will not be stored or reused for training purposes. Please also bear in mind the guidelines for your course: anything not explicitly permitted is prohibited.
Please also familiarise yourself with the rules regarding the use of AI in your course. Your lecturer decides which tools may be used for data collection, transcription and analysis. Anything not explicitly permitted is considered prohibited. If in doubt, please clarify this directly with your lecturer.
AI is no substitute for original scientific work
Even though AI tools can save you a great deal of work when it comes to analysing data, the academic responsibility remains with you. In practical terms, this means that if you have your interview data analysed by an AI tool and the result is incorrect, you are responsible for identifying and correcting this. AI-generated results may contain biased, incomplete or simply incorrect information that nevertheless sounds convincing. You should therefore always check the results yourself – at least on a random basis – by analysing a portion of your data independently and comparing it with the AI output.
You should also keep a record of which AI tools you have used, when and for what purpose, for example in a research journal. This makes your research process transparent and traceable, which is particularly important for seminar papers and dissertations.