Frequently asked questions on research ethics

The checklist asks the applicant for a summary of the research (a clear and concise description, including rationale, objective, design and methods). This description must be very clear about what the researcher proposes to do in this project, for which the researcher needs ethical approval. Describe very well the objective, design, and method relevant to your project.

If, on the other hand, you include sentences about the larger context that your project is a part of (for example, a bigger project of which this project is a small part, or a longer-term aim of which this project is only the beginning), make very clear that you refer to the larger context, and not to the project you seek approval for here. It must be very clear for the reviewer of your application what the boundary is between your project (for which you seek approval now) and the larger context (for which you do not).

We advise against uploading your research plan into AI tools that use input data to train their models, as doing so compromises your intellectual property claims.

The ethics review is intended to help you think through potential ethical risks and mitigation strategies specific to your project; generic or superficial answers, whether written by AI or copied from other ethics questionnaires, do not fulfil this objective.

You may choose to use AI tools, specifically generative AI, for example, to improve the clarity of your writing or to explore possible ethical concerns of your research.  If you do so, be aware that AI-generated responses may contain inaccuracies or be too generic. These tools cannot fully comprehend the specific context or ethical dimensions of your project. Your answers must be based on your own understanding of the project and must reflect your critical engagement with its ethical dimensions.

Writing about the mitigation of risks and writing this with AI is problematic because general models are not able to contextualise. You need to be able to implement the mitigation strategies you suggest.

You are fully responsible for the accuracy and originality of all content in your submission, and you are expected to act on ethics advice received.

Regarding interviews, sometimes researchers are uncertain whether their relatively simple interview with an expert or client falls under “research with human participants”. Roughly, one could take the following guide / rule of thumb (but when in doubt you can of course still ask an ethics adviser):

1. Starting point: you want to research a new topic. You don’t know enough about it, so you read literature (research WITHOUT human subjects).
2. What if you get stuck in literature and ask a colleague “can you please show me the most important literature and quickly summarize it for me?: that is not "research with human subjects" but "literature research with some help from a human colleague". The research data is the literature; the colleague helps you to read it.
3. What if you find that the literature is incomplete and you search the web for other sources; you call the Rathenau Institute whether they can summarize and send to you their latest reports on the topic. The research data is the reports; the human (in this case: employee of Rathenau) is only someone giving the material to you: still not "research with human subjects"
4. What if you think the literature is incomplete; you decide to interview a few professionals about their knowledge, insights, needs, current practices, etcetera, and you use their answers to map the field complementary to what you find in literature: then it is "ethnographic research with human subjects".
5. Grey area: what if you are interviewing your client / having a vivid and long discussion with your supervisors, who may work outside the UT? What if you talk to your client  in a design work to find out their requirements for your thesis? Then you have to distinguish their role as client/supervisor, telling you what you should be doing (not research, within the general parameters of supporting you in your studies), from their role as informant, being interviewed by you, with recordings, or taking quotes to use later, to gather insights about the domain (probably research).

Obviously, the latter two may  still be an ethically quite trivial kind of research with human subjects, in which case the corresponding section of the checklist should be very easy and quick to fill out.

“Research population” covers all the individuals and organizations acting as sources for your data collection and other research and design work, including participants, respondents, subjects in experiments, informants, co-designers, interviewees, and people to be observed.  
 
The number of participants: must be reasonably decided; e.g. through sample size calculation, expected point of saturation, etc. 
 
Inclusion: characteristics that participants must have in order to be included (e.g., profession, age, membership of an organization, capacity to speak a certain language, etc) 
 
Exclusion: Where relevant, you can formulate additional exclusion criteria. This concerns individuals who meet the inclusion criteria but who will be excluded from the study for other reasons (e.g., because they may be particularly impacted by the risks to which participants will be exposed, or lack a necessary competency in a certain language) 

Recruitment: How will the invitation be distributed? The exact method can carry several implications, e.g. for representativity and diversity of the population, and the level of pressure that people may feel to respond. 

A non-exhaustive list of research types:

Interviewing: paper/online questionnaires, survey, face to face or online interview, focus group

Field research in natural setting: observations, contextual interview, automated data collection

Individual or group design work: e.g., codesign workshop

Participation in non-experiment activity: can be formative evaluation of prototypes, but more in general providing artificial tasks, includes triggering stimuli and tasks to elicit observable behavior and responses; measured with e.g. observations, interviews, and manual or automated data collection

Participation in formal experiment: asked to perform a set task, respond to predefined stimuli, measuring behaviour and outcomes

Some contexts pose special ethics considerations. For example, many researches take place with children in a school classroom, or during sports training, or…. In which case aspects of safety; accidental inclusion of non-consenting participants, and other aspects play a role.

Adverse effects may include, for example, the danger posed by equipment or human error, e.g. risk of injury; danger posed by the task; adverse effects on mental health and well being (e.g. due to unpleasant stimuli such as shocking visual materials or confronting interview questions); physical effects such as epilepsy or motion sickness; and other effects.  

 Screening: If certain participants are more than average vulnerable to adverse effects from the research, you can decide to exclude those participants through screening before the start of your research (e.g., you might want to exclude people with epilepsy when subjecting them to certain visual content). You have described this as part of your inclusion/exclusion criteria and can refer to these here. 

 Mitigations might include monitoring, defining criteria for discontinuing the research because of major discomfort (“stop protocol”), providing debriefing, or other facilities. 

How can you do this: You can inform participants about the research through an information letter/e-mail/brochure, information posted on a website which is brought to the participant’s attention, information delivered orally by the researcher or intermediary, or other means. The level of appropriate formality also depends on the possible severity/urgency of ethics issues.

Understanding is key: It is important that you show how you ensure that the participant actually received and processed the information. Make sure the information is not unnecessarily complex or hard to understand; take the expected reading ability of the audience into account; the amount of information that is delivered must not be out of proportion with the size and scope of the research, leading to an unreasonable burden of another kind.

Do not withhold information unless: Information may not be withheld from participants unless it could bias subsequent findings (e.g., if information about the research question/hypothesis or awareness of the details of the research procedure is likely to influence participants’ behaviour or responses to questions) or if doing so has been found to be in the public interest. Information concerning any actual or potential risks or burdens of a study should never be withheld.

Debriefing is used to inform participants about any incomplete or deceptive information they received earlier, and to inform them of possible follow-ups to the research.  

Debrief after deception/withholding information: If the information has been withheld from participants, the information should be provided to them immediately after their participation is completed. This debriefing is typically done by explaining in person or providing a standard document with an explanation. Such debriefing may have to involve an opportunity for participants to confirm/withdraw their initial consent.  

Debriefing about follow-up: In the interest of transparency, it is good to inform participants about what will happen after their participation is completed. How will you inform participants about what will happen after their participation is concluded? Some options (of which you can use more than one) are: provide the participant with the researcher’s contact details so they can ask for more information if they would like to know more; provide oral or written information about what the researchers will do with the collected data; participants and others who are interested may provide their contact details to the researcher and will then receive a summary of the research results. If you use none of these, please explain why. 

The freedom to participate or not and to withdraw might be reduced in various circumstances. Generally, this reduction is not desirable; sometimes it can be considered acceptable but it always requires a clear explanation (in this request, but also to the participant where possible).  
 
Limiting freedom to participate: Some underlying reasons that the participant does not have freedom of choice to participate (but not in every circumstance acceptable): they are not aware that research is going on (so cannot choose NOT to participate); there is no reasonable alternative to participation (almost always not acceptable);  
 
Limiting freedom to withdraw: Some underlying reasons limiting the freedom to withdraw after consenting to participate (but not in every circumstance acceptable): Data has been anonymized and the link to the participant’s identity has been broken; data has been publicly released in a way that cannot reasonably be undone; data is part of a data set for more people and the data for this person cannot be withdrawn without withdrawing a larger group of data of other people (placing an undue cost on research); 

When you work with group activities, the withdrawal of one participant may have consequences for the participation and data collection of several others; this requires special attention.

Please note that not all human beings are capable of consent. Individuals with the capacity or competence to consent: 

• Are 16 years or older (adult); 
• Have the capacity to make choices about a proposed course of action; 
• Know about the risks, benefits, and alternatives; 
• Understand that consent is ‘voluntary and continuing permission’; 
• Understand that consent 'can be withdrawn at any time'.  

Research with less-than-competent participants: If an individual is non-competent or less competent to provide consent, you should ask for informed consent from the individual’s legal guardian or representative(s). For adults who are unable to provide informed consent (incapacitated subjects), their legal guardian or representative must sign consent for participation. If at all possible, you must additionally ask/verify with the participant and break off the participation if during the experiment the participant shows clear signs of objection; in that case describe how you do this. See also https://english.ccmo.nl/human-subjects/informed-consent/incapacitated-adults. 

Research with minors: according to Dutch law, minors younger than 12 years of age are not able to give informed consent, and parent(s) or legal representative(s) must sign consent for the child’s participation. In the case of minors older than 11 and younger than 16 years of age, informed consent is obtained from both the minor and the parent(s) or legal representatives). In both cases, consent from one parent/legal representative is considered sufficient, unless the Ethics Commission decides that a particular research project requires consent from both parents. From 16 years of age, consent is only obtained from the participant. For some types of research, it may nevertheless be good practice to inform the parents or legal representatives.  

The formality and reliability of the (documentation of) obtained consent partially relates to the severity of the ethical issues involved in the research.  

 Some possible forms for obtaining consent are: 

• Signed, written consent form prior to participation 
• Active, non-anonymous online consent before the start of the research (e.g. as part of the opening statement of the online survey including a checkbox for the appropriate forms of consent) 
• Active, anonymous online consent before the start of the research (e.g. as part of the opening statement of the online survey including a checkbox for the appropriate forms of consent) 
• Oral, non-recorded consent prior to the research 
• Oral, recorded consent prior to the research (e.g., interview) 
• Passive/tacit consent (opt-out). This type of consent assumes individuals’ consent if they do not explicitly object to participation after they have been informed about the research. Note: this is a dispreferred option! If you intend to use this form of consent, make a clear case for (a) why this is necessary and (b) why this is acceptable, and (c) why you think sufficient action will be taken to inform participants (and possibly legal representatives) of the research. 
• No consent: only in exceptional cases; requires very good motivation including identification of, and measures to mitigate, potential undesirable consequences. 

Participants may be offered proportional compensation. If you are intending to use incentives/payments, keep in mind that such rewards should be modest, in order to avoid enticing individuals to participate. 

 Some forms of reward that have been used include:

·       travel expenses to a specified maximum

·       voucher of a specified value

·       lottery amongst participants to win a prize of specified nature and value

·       financial reward: amount per activity/time

·       no reward

When you use an existing data set for your research, the data set has originally been collected for a purpose other than your research project. This original purpose may have been research, as well, or something else (e.g., Twitter data, medical dossiers, or websites, which are typically not created for the purpose of scientific research). Repurposing data beyond the original scope for which they were created has to be done with care. 

Different types of data sources are:

• Publicly available sources: Examples of such data: public documents and reports, newspapers, public websites, open web forums or web platforms including Twitter communication, and other public media and/or publicly published data sets. Inclusion of fully publicly available data does generally not lead to additional ethical concerns. 
• Semi-public sources: Examples of such data: data on web forums, discussion forums, online chatrooms, Facebook and other social media that are accessible only to members of an organization or to registered users with a password. These data are not automatically available to all. You may obtain access if you register as a user. You should nevertheless be aware that individuals who use such platforms expect that other users will have purposes similar to their own. They may not wish to be observed or cited by researchers. In such cases, you should start by consulting the terms and conditions of use of the specific platform to determine whether there is a gatekeeper/administrator who you should approach for approval or advice. 
• Private sources: Some examples of such data: personal private data (e.g. medical or police files on individuals); data from Statistics Netherlands (CBS); personal records; data from previous research activities or corporate data that was not publicly released. For private sources, you need to explain for which purpose the data was originally generated, how you will obtain access to the data, what are the conditions to use, and whether individuals that the data pertains to have already consented to this kind of additional later use of the data. 

Some research methods can lead to accidental discoveries that may be of vital importance to the data subject, such as an irregular heartbeat on an ECG that was included in the data set that you use. If so, outline an exact procedure to be followed in such a case. For instance, whether the subject is to be informed about such a result. 

Accidental findings in existing data sets: Generally not desirable to inform the subject because they never agreed to diagnostic use of the data and may not even be aware that their data is part of your research. 

Bias is a prejudiced and unfair belief about a person, group of people or animal species. It normally relies on an unfavorable preconceived evaluation of a person based on personal characteristics or perceived group membership. Biased decisions result in unfair treatment of those affected by it or otherwise disadvantages them. 

Possible sources of bias are:

• Biased data sampling/content (eg only high educated college students);
• methodological bias (eg not properly and balanced subdividing the data into test and train blocks);
• algorithmic bias (e.g. if a certain subpopulation is easier to model… verified by doing post-performance analysis across different populations)

Often, the performance of an AI is tested on all the test population (for example, data from all patients), and a single performance score is reported. But it can happen that the test population is composed of demographic subcategories (for example, pediatric patients vs adult patients, or women vs men) on which the performance score of that AI is significantly different (“demographic bias”). This difference only becomes clear when the demographic subcategories are tested separately (“stratified testing”).

When people are asked to label extreme or problematic content that might effect their mental health, e.g. Facebook asking people to label content so they can train their models to detect or avoid harmful content, or asking people to label a Telegram database of criminal communications.

.

By misleading people, e.g., crowdsourcing without telling them why. This is unethical, but may also methodologically lead to skewed or invalid label sets because you are not in control of who makes the labels on what ground.

Pruning, as well as quantisation, are techniques by which a dense or large AI model (a neural network) is optimised, such that the model can be executed on machines with much fewer resources, or can consume much less energy when used. A technical summary with examples of pruning and quantisation techniques: https://www.tensorflow.org/model_optimization/guide .

Large AI models with many parameters may retain individual training data points and their labels. This is called memorisation, a form of overfitting which comes with ethical risks, because it can disclose exact training data, which may be personal information, security information, or copyrighted data.

https://www.licenses.ai/ai-licenses

A brief description: https://oecd.ai/en/catalogue/tools/model-cards.

This is how the HuggingFace model repository implements them: https://huggingface.co/docs/hub/model-cards#model-cards.

When it concerns a UT system, please contact responsible-disclosure@utwente.nl (see also: https://www.utwente.nl/en/cyber-safety/responsible/)

When it concerns another system, please follow the procedure for Coordinated Vulnerability Disclosure described here: https://www.utwente.nl/en/eemcs/research/ethics/coordinated-vulnerability-disclosure/

Examples of misuse

• Use of pattern recognition in surveillance
• Generation of fake input
• Misinformation propagation

Harm from AI that did not make an error and was applied correctly: AI can often lead to real decisions that have a real, potentially harmful impact on people. Some examples: people not getting invited for a job interview because the automatic application-analysing AI discarded their letter; algorithms in crime & law keeping informing police or judicial decisions; algorithm-supported decisions in insurance or bank loans; health care decision support systems; with or without bias, in such applications the AI can lead to decisions that can have a real harmful impact on people.

Some example risks are (see also the FAQ):

• collecting data in potentially dangerous environments or through dangerous activities
• dealing with sensitive, distressing, or controversial topics
• being at risk of retaliation for your research
• becoming exposed to media pressure or public controversy through your research
• when working in a setting that may pose other ‘lone worker’ risks