Yes, if three levels are clarified separately: the promise made to employees matches the technology, every service that sees the comments works under a contract and with a legal basis, and the works council is involved. There is no fixed minimum group size for anonymous comments, and no rule that governs exactly this kind of analysis. According to the Commission’s guidelines, the ban on emotion recognition in the workplace does not cover plain text analysis.
At a glance
- Pseudonymous is not anonymous: if answers can be traced back to a person, you are processing personal data, and in 2025 a supervisory authority answered an anonymity promise that was technically untrue with a warning.
- There is no official minimum group size; platforms state their own defaults, and the draft of the European Data Protection Board mentions the number of people only as a factor.
- Deleting names from comments is not enough: the German Data Protection Conference warns that context allows conclusions to be drawn, especially with AI.
- The Federal Labour Court rejected co-determination for the survey because participation was voluntary, but affirmed it for the IT system that captured the answers, under section 87(1) no. 6 of the Works Constitution Act (BetrVG).
- Of 12 directory tools with AI analysis of survey or team feedback, a third publicly name a data processing agreement and none names a document for the works council.
Take an example. The annual survey is closed, 800 employees have answered, and the spreadsheet holds around 2,400 free-text comments. Someone on the team suggests pasting them into an AI tool and having it list the themes. That saves weeks. But it raises four questions that rarely get asked beforehand: was the survey really anonymous? Who else processes the texts now? What does the works council (Betriebsrat, the elected employee body) say? And does this fall under the EU AI Act?
This article walks through those four questions from the HR team’s point of view. It covers the analysis of free text from a survey of employees, not the assessment of individuals and not applicants; there are separate articles on those, which I link to. The legal statements are orientation based on statutory text, court decisions and papers from the supervisory authorities, as of 15 November 2026, and do not replace advice in an individual case. Where I give my own reading that I cannot back up, I say so.
I advise HR teams on AI and have a commercial interest in that, which I disclose. The final section covers when you do not need me for the analysis.
Can an HR team analyse free-text survey comments with AI?
Yes, but only if three levels are clarified separately: the survey itself, the processing by the AI tool and the involvement of the works council. I have not found a rule that governs exactly this analysis. The list of processing operations that require a data protection impact assessment (DPIA) published by the German Data Protection Conference (DSK) has no entry on employee surveys, and none of the employment examples the European Commission has published on the EU AI Act covers the analysis of surveys. So you derive the answer from general rules.
Two of these rules are in the DSK guidance on AI and data protection of 6 May 2024. Every processing step in which an AI application processes personal data needs a legal basis (section 1.5). And company data protection officers should always be involved, while involving works and staff councils should be considered (section 1.12).
The three levels are separate because they fail in different ways. A survey can be properly anonymous and the AI service still works without a contract. The contract can be in order and the anonymity promise still false. And both can be right while the works council first hears of the tool from the intranet. The list sorts the three levels, and the following sections take them in turn.
Three levels, three questions before the analysis
- The survey: what did you promise employees, and is it true?Anonymity, voluntariness, information on recipients (Art. 5, 13 GDPR).
- The processing: who sees the comments, on what basis and under what contract?Legal basis, data processing agreement, impact assessment, third countries (Art. 6, 28, 35, 44 GDPR).
- Co-determination: who in the company has a say?Sections 94, 87, 80, 90 BetrVG, plus classification under the EU AI Act.
What does anonymous mean for free text, and when is it only pseudonymous?
Anonymous means that nobody can attribute the answers to a person, not even with additional knowledge that is reasonably likely to be used. Anything short of that is pseudonymous, and pseudonymous data remains personal data. Art. 4 no. 5 GDPR describes pseudonymisation as processing after which the data can no longer be attributed to a person without additional information, provided that information is kept separately. Recital 26 says that such data should be regarded as information on an identifiable person. The data protection principles, by contrast, do not apply to anonymous information.
Whether a person is identifiable depends on all the means reasonably likely to be used (recital 26). In its Opinion 05/2014 (WP 216), the Article 29 Working Party writes that pseudonymisation is not a method of anonymisation, that it merely reduces the linkability of a dataset and is therefore a useful security measure. It names three risks: singling out a record, linking it with other data and inferring attributes. A sound solution has to defend against all three.
Free text adds something. On 4 September 2025, in case C-413/23 P (EDPS v SRB), the Court of Justice of the EU (CJEU) held that personal opinions, as an expression of a person’s thinking, are closely linked to that person. The General Court had misjudged the nature of personal opinions when it required an examination of the content, purpose or effect of the comments. Identifiability, the Court added, is to be assessed at the time of collection and from the controller’s point of view. Two limits: the judgment concerns Regulation (EU) 2018/1725 for EU institutions and not the GDPR itself, and I have read the Court’s press release, not the full text. It still works as a pointer: a comment someone writes in their own words is rarely a neutral data point.
The 41st annual report 2025 of the state data protection commissioner of Baden-Württemberg (section 5.5.6) shows what this looks like in supervisory practice. A company had told its employees that the survey was anonymous and that no answers would be passed on if fewer than five people had answered a question. In fact, it was recorded who had taken part, and the results were not anonymous for the company. The service provider processed as a processor on instructions. The authority issued a warning under Art. 58(2)(b) GDPR. A misleading anonymity promise, it held, is incompatible with good faith (Art. 5(1)(a) GDPR) because it can induce employees to give information they would not otherwise have given.
My reading: the lesson from the case is not that five is enough. The five was the company’s own commitment, not a requirement from the authority. The lesson is that promise and technology must match. If your provider technically knows the attribution, write “confidential, only the provider can attribute answers” and not “anonymous”. That is more honest, and it spares you the dispute the promise would otherwise cause later.
How many answers does a group need so that nobody can be identified?
There is no official number. Opinion WP 216 describes k-anonymity: the higher k, the stronger the guarantee, and a k that is too small is a problem. As an illustration it contrasts k = 2 with more than 10, and it names no minimum. The draft Guidelines 02/2026 on anonymisation of the European Data Protection Board (EDPB) list the number of people in a dataset as one of several factors, likewise without a value. The “five” in the Baden-Württemberg report was the promise of the company that was audited.
Numbers therefore come from the platforms. The documentation of Microsoft Viva Glint gives five responses as the default for scale scores, adjustable to three or four, and ten for comments; confidential surveys never go below three. For comments, what is counted is the responses to the survey, not the number of comments. Groups that are too small are also hidden together with the next larger one so that values cannot be guessed by subtraction. Culture Amp likewise describes that the minimum group for comments counts the people who responded, but on the help page I retrieved it gives no default value, and Qualtrics controls comments through a separate setting, also without a default number. This is documentation, not a recommendation.
Group size alone is not enough. Example 17 in the EDPB draft describes a team survey about the line manager in which the manager sees only averages. If every team member took part and none gave more than 3, the manager can infer that all of them are critical. The Board writes: although aggregated, the information still relates to each member of the team. WP 216 also names this weakness: if all k people are in the same group, the attribute is easy to infer once group membership is known.
Then there is the combination of attributes. Sweeney showed in 2002 that, in an analysis of 1990 US census data, 87 percent of the population (216 of 248 million) were likely to be unique on the basis of postcode, sex and date of birth alone. That is US data and says nothing about your workforce; as a pattern, though, department, location, year of joining and age band are harmless alone and a fingerprint together. For opinion surveys, the EDPB draft accordingly requires that the answers contain no demographic data that can be traced back to a person.
A caveat on the source: the Guidelines 02/2026 were adopted for public consultation on 7 July 2026, and the consultation ran until 30 October 2026. I have not found a final version. Read them as a direction, not as the applicable interpretation. In practice, this means for me: set the threshold for comments higher than for numerical values, count people and not comments, and do not analyse groups that can be narrowed to a few people through several filters. Which number suits your company is for your data protection officer to decide with you.
Take a guess
From how many responses does a platform show comments?
One example from a vendor’s documentation. There is no official number, and the value is no benchmark for your survey.
From how many responses to the survey does the Viva Glint platform show the comments of a group under its default setting?
Source: Microsoft Learn, Manage Viva Glint confidentiality thresholds (page dated 6 May 2026, retrieved on 15 November 2026).
What is in free text, and how much of it can an AI find again?
More than the question asked for: names of managers, details about health, complaints about conduct, hints of wrongdoing. And deleting names is not enough. The DSK writes in its guidance (section 3.1) that it is regularly not sufficient to remove names and addresses from an input, because a link to a person can be established from the context. For AI applications meant to draw cross-references even from unstructured data, the danger is particularly high. The Hamburg data protection commissioner (HmbBfDI) says the same in his checklist on LLM chatbots.
Special categories often appear in the text unasked. Art. 9(1) GDPR prohibits the processing of, among other things, data on health, trade union membership, political opinions and philosophical beliefs. There are exceptions, for example with explicit consent (Art. 9(2)(a)) or under section 26(3) of the Federal Data Protection Act (BDSG) for obligations under employment law. You cannot obtain explicit consent in advance for something a person only writes while typing; that is my reading. Whether an exception holds is for your data protection officer to clarify.
Automatic redaction helps only to a limited degree. A tool from the US Department of Defense for open-ended survey comments (spaCy, more than 20,000 training examples) found, in 2024, 91 percent of dates, 82 percent of places and 80 to 81 percent of ranks, but only 31 percent of medical information. That was an English-language military case and cannot be transferred to German comments. As a warning it is enough: redaction without a human spot check is not anonymisation.
And the AI itself can become the attacker. Staab and colleagues showed in 2024 that language models can infer attributes such as place of residence, income and sex from Reddit texts with up to 85 percent top-1 and 95 percent top-3 accuracy, at a fraction of the cost and time of humans; text anonymisation, by contrast, is currently largely ineffective. In 2026, Lermen and colleagues reached up to 68 percent recall at 90 percent precision when linking pseudonymous online profiles to people, while the best methods without a language model were close to zero. Both are preprints using public online texts that are longer than typical survey comments. Whether this holds for short comments is not established. But anyone who relies on “the AI won’t figure that out anyway” has no source for it.
Tap to open
Four things hidden in free text
Orientation based on the supervisory authorities’ papers, not legal advice. What to do in each case is for your data protection officer to decide with you.
Source: DSK guidance on AI (6 May 2024); GDPR Art. 9; Whistleblower Protection Act (HinSchG) sections 2, 12; General Equal Treatment Act (AGG) sections 12, 13. As of 15 November 2026.
What legal basis and what contracts does the analysis need?
A legal basis for the survey and for every further step, a contract under Art. 28 GDPR with every service that sees the comments, and, where the risk is high, a DPIA. Start with the basis, because it is where surveys fail most often.
The state data protection commissioner of Baden-Württemberg considers Art. 6(1)(b) GDPR (performance of a contract) relevant to employee surveys only if the answers are necessary to carry out the employment relationship. For questions about general satisfaction that is not the case. As an alternative he names voluntary participation with consent, without pressure. In the case examined, consent was not voluntary because participation was monitored, and not informed because the notices under Art. 13 GDPR were missing, among other things on the data passed to the service provider in advance. Section 26(2) BDSG requires an assessment of the dependence that exists in the employment relationship when judging voluntariness.
How far section 26 BDSG itself carries is disputed. Under the CJEU judgment of 30 March 2023 (C-34/21), employee data protection rules that do not meet Art. 88 GDPR are in principle left unapplied; the Baden-Württemberg authority concludes that section 26(1) sentence 1 BDSG is no longer applicable. That is a supervisory authority’s classification, not a decision on section 26 BDSG, because the judgment concerned a Hessian state provision.
For the service that processes the texts: where a body uses an external AI application as a cloud solution, the DSK says a processing relationship often exists, and an agreement under Art. 28(3) GDPR (a data processing agreement) must be concluded with the provider. Processing may only take place on documented instructions. The Hamburg data protection commissioner goes further for chatbots: if the provider reserves the right in its terms to use the data for its own purposes, no personal data may be transmitted, not even that of one’s own employees, and a legal basis will as a rule not be found. The paper dates from 2023 and by its wording applies to chatbots. The check is the same as the one described in the article on the HR chatbot for employee questions.
The impact assessment is where you answer the questions of this article in writing. Art. 35(1) GDPR requires it where processing is likely to result in a high risk, above all with new technologies. According to the DSK guidance it is “often” required for AI applications. The DSK must-do list names AI for evaluating personal aspects (no. 11) and extensive processing of employee behavioural data (no. 8); it has no entry on surveys. Whether your analysis falls under these follows from Art. 35(1). Art. 35(9) requires seeking the views of data subjects or their representatives where appropriate, and for employees that is the works council. My advice: do the impact assessment with your data protection officer as soon as there is doubt. What belongs in that check is also set out in the overview on GDPR and AI in HR.
If the service is located outside the EU, Art. 44 and 45 GDPR come in. For certified US companies, the adequacy decision on the EU-US Data Privacy Framework applies (Implementing Decision (EU) 2023/1795 of 10 July 2023). On 3 September 2025 (T-553/23), the General Court (EuG) dismissed the action for annulment. According to the press release, an appeal to the Court of Justice was possible; whether one was filed or decided I have not checked. Check the status before you rely on the decision. Under Art. 13(1)(e) and (f) GDPR you must inform employees about recipients and the intended transfer to a third country.
Self-check · interactive
Is your analysis prepared?
Five questions for a first assessment. Orientation, not legal advice. The decision is made by your data protection officer, works council and legal counsel.
Does an external service process the comments, such as a survey provider or an AI interface?
A cloud function inside the survey tool counts too.
When does the works council co-determine?
For the survey itself only in some circumstances, for the AI tool very probably. The Federal Labour Court (BAG) rejected co-determination for two surveys because participation was strictly voluntary, but affirmed it for the IT system that captured the answers, under section 87(1) no. 6 of the Works Constitution Act (BetrVG). There is no decision on AI analysis of free text.
In its decision of 21 November 2017 (1 ABR 47/16), the BAG rejected co-determination under section 94(1) BetrVG for an employee survey because it was not a personnel questionnaire: participation was designed to be “strictly voluntary”. Whether the survey was sufficiently anonymised, the panel expressly left open. What carried the decision was voluntariness, not anonymity. Section 87(1) no. 7 (health protection) did not apply either, because the survey was objectively not a risk assessment.
The decision of 11 December 2018 (1 ABR 13/17) concerned the IT system of a survey. A technical device is intended for monitoring under section 87(1) no. 6 BetrVG if it is objectively suitable for collecting and recording behaviour or performance information about employees; the employer’s intention is irrelevant. The survey system fell under that. A mere change of the questions did not change the device, so no new co-determination right arose there. On section 87(1) no. 1 and section 94, the court noted that the survey had been voluntary and anonymised for participants. How far no. 6 reaches is also shown by the decision on an employer’s Facebook page (1 ABR 7/15): the function for visitor posts is subject to co-determination.
My reading of this: anonymity does not protect you from co-determination on the tool. A system that captures and analyses comments about teams and managers can be objectively suitable for collecting behaviour and performance information. That is a derivation and not a decision on AI. If you want to be safe, treat the survey tool and the AI analysis as relevant to co-determination and settle purpose, access, deletion and analysis thresholds in an agreement; what that can look like is described on the page on the works council agreement on AI.
For AI, the Works Constitution Act has its own provisions. Where the works council has to assess the introduction or use of AI, calling in an expert is deemed necessary (section 80(3) BetrVG). The employer must inform the works council about the planned use of AI in good time and with the documents (section 90(1) no. 3) and discuss the measures in time for proposals and concerns to be taken into account in the planning (section 90(2)). That applies before the tool is bought.
Join in
Three questions on the works council and the survey
Test your knowledge of the decisions. Orientation, not legal advice.
Question 1 of 3
Why did the BAG in 2017 reject co-determination under section 94 BetrVG for the employee survey?
Source: BAG, 21 November 2017, 1 ABR 47/16; BAG, 11 December 2018, 1 ABR 13/17; BetrVG section 80(3). As of 15 November 2026.
Does the analysis fall under the EU AI Act?
The ban on emotion recognition in the workplace does not cover plain text analysis. Whether the analysis is a high-risk system under Annex III no. 4, no source I have found answers. The obligations for high-risk systems under Annex III apply, as amended by the Digital Omnibus, from 2 December 2027.
Art. 5(1)(f) of the EU AI Act prohibits AI systems that infer the emotions of a person in the workplace or in education, except for medical or safety reasons. Under Art. 3 no. 39, an emotion recognition system is one that infers emotions or intentions on the basis of biometric data. The Commission’s guidelines on prohibited practices (C(2025) 5052 of 29 July 2025) say in paragraphs 250 and 251 that a system that infers emotions from written text does not fall under the ban because it is not based on biometric data; paragraph 265 repeats this. It is different with webcam, speech recognition and keystrokes. The guidelines are not binding; the final interpretation lies with the CJEU (para. 5). Violations of the ban cost up to 35 million euros or 7 percent of worldwide annual turnover (Art. 99(3)).
The second question is high risk. Annex III no. 4 names AI for recruitment or selection and AI that makes decisions on working conditions, allocates tasks on the basis of behaviour or personal traits, or monitors and evaluates the performance and behaviour of employees. An analysis of free-text survey comments is not listed there. Under Art. 6(3), an Annex III system can be exempt, for example for a narrow procedural task or a preparatory activity, but never where profiling is involved. On 19 May 2026 the Commission published a draft of guidelines on classification, updated on 23 July 2026; I have not found a final version.
My interpretation, not the Commission’s: a topic overview across large groups, in which no person and no small team can be identified, is far from monitoring and evaluating employees. As soon as comments are attributed to individuals or small teams and evaluated, it moves closer, and with it the questions from the article on AI in performance reviews. That is one more reason not to cut analysis groups too small.
If your tool counts as a high-risk system, you as deployer have, under Art. 26, among others these duties: you inform the employee representatives and the affected employees before use (para. 7), ensure human oversight by competent people (para. 2), keep logs for at least six months (para. 6) and use the provider’s information for the data protection impact assessment (para. 9). Violations of deployer obligations cost up to 15 million euros or 3 percent of annual turnover (Art. 99(4)). Regardless of classification, Art. 4 requires measures for the AI literacy of your staff. The overall overview for HR is on the page on the EU AI Act for HR departments.
Status and deadlines · click through
What applies when, or applied when
The dates come from the sources named. The Commission’s guidelines are not binding, and the EDPB text is a consultation draft.
Feb 2, 2025
Prohibitions under Art. 5 apply
The ban on emotion recognition in the workplace has applied since this day (Art. 113 AI Act in the consolidated version).
How good is AI at analysing free text, and where does it get it wrong?
Good enough for a first overview of themes, not good enough to leave the analysis to it unchecked. This is evidenced only through substitute data: none of the studies I found uses employee surveys, and some are preprints.
Von der Heyde and colleagues compared language models with human experts in 2025 when coding German free-text answers on reasons for taking part in a survey. Performance differed strongly between models, only a fine-tuned model reached satisfactory results, and without fine-tuning unequal hit rates led to distorted distributions of the categories. Soria (2026, preprint) examined 3,200 responses with eight models: agreement between model and human did not reach agreement between two humans, the best model reproduced the full human label set in fewer than 60 percent of responses, and the models assigned on average 67 percent more categories per response. For some demographic groups, agreement was lower.
That fits a finding on summaries. Zhu and colleagues (2025, preprint) had 18 models summarise the opinions of 3,000 participants and found persistent weaknesses, especially in the under-representation of minority positions. In surveys, these are often exactly the voices you want to hear. To the contrary: in Gilardi and colleagues (2023, preprint), ChatGPT was more accurate than crowd workers on four of five tasks with 2,382 tweets, with a 2023 model and on other texts.
My conclusion, without a study behind it: have the AI give you themes and frequencies, but read a sample for each theme yourself and look separately and explicitly for the rare, sharp comments. A percentage from the AI summary is a hint and not a result. Whether anonymity raises honesty is, by the way, more open than one assumes: the meta-analysis by Richman and colleagues (1999, 61 studies) found the distortion small when people answered alone and anonymously, and Lelkes and colleagues (2012, students) found that complete anonymity tends to worsen accuracy. Neither involves companies. The only conclusion left is: do not promise anything you cannot keep.
On adoption: according to Bitkom Research, 57 percent of companies with 20 or more employees use AI (603 companies, 2026), and according to the Federal Statistical Office 26 percent of companies with 10 or more employees (2025). Neither reports on HR; how many HR departments use AI for surveys, I do not know.
What do vendors of survey tools disclose?
Little of what you need for data protection and the works council. In the directory I maintain, I went through 103 active tools in the areas of core HR system, employee experience and analytics. Of these, 12 name in their profile an AI function that analyses survey or team feedback. I counted what vendors state on their public pages, checked between 25 September and 26 October 2026 and evaluated on 15 November 2026. I do not name names; this is about the state of play and not a ranking.
Of the 12, 4 publish a data processing agreement, and for 8 I found no statement. On hosting, for 5 the vendor page names the EU among the regions, for one only the USA and the United Kingdom, and for 6 the statement is missing. On the EU AI Act, 2 comment with their own statement or classification (one of them classes itself as not high risk), for 7 I checked and found nothing, and for 3 I did not check. 8 name a security certification under ISO/IEC 27001, and 3 offer a German-language interface. A page or document for the works council I found at none. Information not found does not mean it does not exist.
So ask the vendor in writing before you upload comments: is there a contract under Art. 28 GDPR, and who are the sub-processors? Where is the data, and who can access it from outside the EU? Are comments used for training? Can the minimum size for comments be set? How does the vendor classify the AI function under the EU AI Act, and what documents does the works council get? The answer “we have nothing on that” is itself information.
Own survey
What survey tools with AI analysis disclose
12 of 103 active tools (core HR system, employee experience, analytics) that name an AI function for survey or team feedback in their profile. What is counted is what vendors state publicly. Information not found does not mean it does not exist.
The two values add up to 12. This means a publicly available data processing agreement.
Source: own survey by Nick Schaefer, public vendor pages, checked between 25.09.2026 and 26.10.2026, evaluated on 15.11.2026. As of 15 November 2026.
How do you proceed, and what do you do with reports hidden in free text?
First the promise and the works council, then contract and impact assessment, then the technology, and last the analysis with a sample. The checklist below sums this up in ten points. The state is saved in your browser, and the button copies it as text.
One point belongs before the survey and not after it: what happens if a comment contains more than an opinion? The Whistleblower Protection Act (HinSchG) obliges employers with, as a rule, at least 50 employees to set up an internal reporting channel (section 12). The office confirms receipt after seven days at the latest and gives feedback within three months (section 17), it preserves the confidentiality of those involved (section 8), and it should also handle anonymous reports, without there being a duty to do so (section 16(1)). The material scope is limited (section 2), and not every complaint about management falls under it. Whether a free-text comment in a survey is a “report” within the meaning of the statute does not follow from the statute; that is a legal question you clarify in advance.
Then there is the General Equal Treatment Act (AGG). Employees have the right to complain to the competent bodies of the establishment if they feel disadvantaged on a ground under section 1; the complaint must be examined and the result communicated (section 13(1)). The employer must take the necessary measures to protect against disadvantage, including preventive ones (section 12(1)). An anonymous complaint is not regulated in the statutory text. My conclusion: do not promise in the survey a confidentiality you cannot keep when such a report arrives, and tell employees where to file reports and complaints properly.
To take away
Checklist: analysing free text with AI
Your progress stays saved in this browser.
When you do not need me for the analysis
If your data protection officer does the impact assessment with you, the works council is involved and someone in the company reads the comments as a sample, you do not need me for the analysis. Then this article is your checklist. This section describes my own service; I advise HR teams on AI and have a commercial interest in that.
I do not answer legal questions in a binding way. Whether a comment is a report within the meaning of the Whistleblower Protection Act, whether your analysis requires an impact assessment and whether the works council co-determines are matters for your legal counsel, your data protection officer and the works council.
It is different if you want to set up the analysis as a process before a large survey: promise, group cut, sample, rules for reports, questions to the vendor. I support your HR team with that. In the first conversation I will tell you if you do not need me.
Evidence1
Sources
- DSK: Guidance on artificial intelligence and data protection, version 1.0 of 6 May 2024 (in German)
- DSK: List of processing operations requiring a DPIA (must-do list), version 1.1 (in German)
- European Commission, AI Act Service Desk: Employment (classification examples)
- GDPR Art. 4 no. 1 and no. 5: definitions (Regulation (EU) 2016/679, in German)
- GDPR recital 26: scope, pseudonymisation and anonymous information (in German)
- Article 29 Working Party: Opinion 05/2014 on anonymisation techniques (WP 216)
- CJEU, judgment of 4 September 2025, C-413/23 P (EDPS/SRB), press release No 107/25
- State data protection commissioner of Baden-Württemberg: 41st data protection annual report 2025, section 5.5.6 (employee survey, in German)
- EDPB: Guidelines 02/2026 on Anonymisation, version 1.0 for public consultation (7 July 2026)
- Microsoft Learn: Manage Viva Glint confidentiality thresholds
- Culture Amp Support: Confidentiality protections in reporting
- Qualtrics Support: Enhanced Confidentiality for Filters and Breakouts (EX)
- Sweeney 2002: k-anonymity: a model for protecting privacy, International Journal on Uncertainty, Fuzziness and Knowledge-based Systems 10(5)
- HmbBfDI: Checklist on the use of LLM-based chatbots (13 November 2023, in German)
- GDPR Art. 9: processing of special categories of personal data (in German)
- BDSG section 26: data processing for employment-related purposes (in German)
- US Department of Defense (DPAC), FedCASIC 2024: Overview of PII Removal Tool for Survey Comments
- Staab, Vero, Balunović, Vechev (ICLR 2024): Beyond Memorization: Violating Privacy via Inference with Large Language Models
- Lermen and others (2026): Large-scale online deanonymization with LLMs
- CJEU, judgment of 30 March 2023, C-34/21 (Hauptpersonalrat der Lehrerinnen und Lehrer), press release No 54/23 (in German)
- State data protection commissioner of Baden-Württemberg: discussion paper on legal bases in data protection for the use of AI, version 2.0 (October 2024, in German)
- GDPR Art. 13: information to be provided where data are collected (in German)
- GDPR Art. 28: processor (in German)
- GDPR Art. 35: data protection impact assessment (in German)
- GDPR Art. 44 and 45: transfers to third countries (in German)
- General Court, judgment of 3 September 2025, T-553/23 (Latombe/Commission), press release No 106/25
- European Commission: EU-US data transfers
- BetrVG section 94: personnel questionnaires, assessment principles (in German)
- BetrVG section 87: co-determination rights (in German)
- BetrVG section 80: general duties, subsections 2 and 3 (in German)
- BetrVG section 90: rights to information and consultation (in German)
- BAG, decision of 21 November 2017, 1 ABR 47/16 (in German)
- BAG, decision of 11 December 2018, 1 ABR 13/17 (in German)
- BAG, decision of 13 December 2016, 1 ABR 7/15 (in German)
- EU AI Act (Regulation (EU) 2024/1689) Art. 5: prohibited practices
- EU AI Act Art. 3: definitions (nos. 34 and 39)
- European Commission: Guidelines on prohibited AI practices, C(2025) 5052 final of 29 July 2025
- EU AI Act Annex III: high-risk AI systems, no. 4 employment
- EU AI Act Art. 6: classification rules for high-risk AI systems
- European Commission: draft guidelines on the classification of high-risk AI systems
- EU AI Act Art. 26: obligations of deployers of high-risk AI systems
- EU AI Act Art. 99: penalties
- EU AI Act Art. 4: AI literacy (consolidated version)
- EU AI Act Art. 113: entry into force and application (consolidated version)
- European Commission: AI Omnibus enters into force
- von der Heyde, Haensch, Weiß, Daikeler (2025): AIn't Nothing But a Survey? Using LLMs for Coding German Open-Ended Survey Responses
- Soria (2026): High Agreement, Different Stories: How LLM Classifiers Reshape Demographic Patterns in Survey Data (preprint)
- Zhu, Yang, Bakker, Pentland, Pei (2025): Can AI Truly Represent Your Voice in Deliberations?
- Gilardi, Alizadeh, Kubli (2023): ChatGPT Outperforms Crowd-Workers for Text-Annotation Tasks (preprint)
- Richman, Kiesler, Weisband, Drasgow (1999): A Meta-Analytic Study of Social Desirability Distortion, Journal of Applied Psychology 84(5)
- Lelkes, Krosnick, Marx, Judd, Park (2012): Complete anonymity compromises the accuracy of self-reports, Journal of Experimental Social Psychology 48(6)
- Bitkom Research (14 September 2026): for the first time, the majority of companies use AI (in German)
- Federal Statistical Office (2025): enterprises using artificial intelligence technologies by employment size class (in German)
- HinSchG section 12: duty to set up internal reporting offices (in German)
- HinSchG section 17: procedure for internal reports (in German)
- HinSchG section 8: confidentiality requirement (in German)
- HinSchG section 16: reporting channels for internal reporting offices (in German)
- HinSchG section 2: material scope (in German)
- AGG section 13: right to complain (in German)
- AGG section 12: measures and duties of the employer (in German)
Frequently asked questions
May I paste free-text survey comments into a public AI tool?
Not without checking. The Hamburg data protection commissioner writes in his checklist on LLM chatbots that no personal data may be transmitted if the provider reserves the right to use it for its own purposes, not even that of one’s own employees. Even without names, conclusions can be drawn from the context. Clarify the contract, the data protection officer and the works council beforehand.
How many answers does a group need for comments to stay anonymous?
There is no official number. The draft of the European Data Protection Board names the number of people as one factor, without a value. One large platform documents a default of ten responses for comments and five for scale scores. What also matters is whether groups can be narrowed to a few people through several filters.
Does the works council have to agree if we analyse survey comments with AI?
The Federal Labour Court rejected co-determination for the survey itself because participation was voluntary, but affirmed it for the IT system that captured the answers, under section 87(1) no. 6 BetrVG (1 ABR 13/17). There is no decision on AI analysis of free text. You should inform the works council in good time in any case (sections 80, 90 BetrVG).
Does sentiment analysis of comments fall under the ban on emotion recognition?
Not according to the Commission’s guidelines of 29 July 2025 if it infers from written text, because emotion recognition within the meaning of Art. 3 no. 39 of the EU AI Act requires biometric data. The guidelines are not binding. It is different when analysing face, voice or keystrokes. Whether the analysis is high risk, no source I have found answers.
Do I need a data protection impact assessment for the analysis?
That follows from Art. 35(1) GDPR. The DSK guidance considers it often necessary for AI applications, and the DSK must-do list has no separate entry on surveys. Where in doubt, do it with your data protection officer and seek the works council’s view (Art. 35(9)).
Clarify the AI analysis of surveys
A free first conversation in which we sort out what must be settled for your survey: promise, contract, works council and analysis. If your data protection officer and your works council can carry this with you on their own, I will tell you so there.
Clarify the AI analysis of surveys