Ethics, consent and power: the participant's standing in the study

Consent in research is a relationship negotiated around one study and one group of people. It is not a clause accepted once at sign-up that then licenses whatever is done afterwards. Kramer, Guillory and Hancock (2014) altered the emotional content of the News Feed for 689,003 Facebook users for a week. They told none of them they were in a study, and offered the platform's Data Use Policy as the consent that covered it. No rule the authors or their university could name had been broken, and no participant could refuse. A policy accepted at sign-up cannot serve as consent unless it names the study, the data and the way out.

The ninth of the ten units tests three claims. Consent is study-specific. Compliance is not the same as a participant's standing. The ethics of data research is also an epistemological question, because who is exposed is decided by what counts as data.

A week of altered feeds showed a sign-up clause is not consent

Kramer, Guillory and Hancock (2014) reduced positive emotional content in the News Feed for one group of users and negative content for another. They then measured the tone of what those users posted. Reducing positive expressions produced fewer positive and more negative posts, and reducing negative expressions reversed the pattern. That is emotional contagion at a scale no laboratory could reach, with 689,003 people assigned to a condition without being told. The authors offered Facebook's Data Use Policy, accepted when an account is created, as informed consent. Cornell University's review board determined that the project did not fall under its human research protection programme, because Facebook had conducted it for internal purposes.

PNAS published the paper, then attached a correction and an editorial expression of concern, which recorded the objection directly: questions had been raised about informed consent and the opportunity to opt out, and it was a matter of concern that the data collection might not have been fully consistent with those principles (PNAS, 2014). A consent given for something else, at another time, does not become consent for this study by being mentioned in it.

Consent is a relationship, and the standard says what it must name

Grady (2015) reviews what informed consent now faces, among it broad consent for biobanking, consent for research on identifiable data, and open-ended re-use. Consent is the mechanism through which a participant keeps a choice, so the answer to complexity is better consent, not less. A form the participant cannot read is not a choice. Where reading is slow, or English is a second language, consent has to be read aloud or given at the participant's pace. Consent that names a study, a dataset, a use and a withdrawal route survives the criticism. A signature page that names none of those is a receipt.

In the United Kingdom the standard is written down. The British Psychological Society (2021) states four principles. The first is respect for the autonomy, privacy and dignity of individuals and communities. The others are scientific integrity, social responsibility, and maximising benefit while minimising harm. It derives valid consent, withdrawal, confidentiality and the narrow limits of deception from them.

The international reference was revised in October 2024. The World Medical Association (2024) adopted a new Declaration of Helsinki, the first revision since 2013. It renames subjects as participants, requires consent for the collection, storage and foreseeable secondary use of identifiable or re-identifiable data, and makes the distribution of benefits, risks and burdens an explicit duty. The re-identification problem the next section measures is now written into the standard.

Where the regulation and the research came apart

Metcalf and Crawford (2016) give the structural account. Data science and research ethics have diverged, they argue. Core commitments of human-subjects regulation, such as the distinction between research and practice, cannot be exported cleanly from biomedicine into data science. The Common Rule is the United States regulation for federally funded research with human participants. Its proposed revision largely excluded data science methods and public datasets.

Two surveys show what that gap looks like in practice. Vitak, Shilton and Ashktorab (2016) surveyed 263 researchers who work with online data. They converged on some practices and disagreed on others, uncorrelated with discipline, method or workplace, and the decisions were settled by individual judgement rather than a shared standard. Fiesler and Proferes (2018) measured the other side. Their 368 valid responses included 268 public-account Twitter users. Of those, 61.2 per cent did not know researchers used tweets. A further 64.9 per cent believed researchers should not be able to do so without permission. The default that public data is available data does not match what the people whose data it is expect.

Anonymity is a claim to be tested, not a default

A small dataset is often called anonymous because the names have been removed. Narayanan and Shmatikov (2008) showed why that is not enough. Their de-anonymisation attacks recover individuals from large sparse datasets by exploiting the rarity of certain records. They demonstrated the method on the Netflix Prize dataset of anonymous ratings from 500,000 subscribers, identifying known users and exposing their apparent political preferences.

Rocher, Hendrickx and de Montjoye (2019) turned that warning into a number. Their generative model estimates how likely a person is to be re-identified from an incomplete dataset. It predicts uniqueness with AUC scores between 0.84 and 0.97 across 210 populations. On demographic data it estimated that 99.98 per cent of Americans would be correctly re-identified using 15 attributes. Ohm (2010) adds the legal conclusion: the line between identifiable and non-identifiable data cannot carry the weight placed on it.

Contested question: can consent carry this weight at all?

The literature agrees that a clause accepted at sign-up is not consent. It disagrees about whether consent, repaired, can reach the studies that now matter. Grady (2015) holds the repair position. Zook and colleagues (2017) take the same line in practice, setting out ten rules for responsible big data research. Metcalf and Crawford (2016) hold the harder position. The instruments of human-subjects regulation no longer describe what data researchers do. A better consent form answers a question the field has stopped asking. Ohm (2010) supplies the technical version of that doubt, since a consent to share data cannot promise that the shared data will stay non-identifiable. On this account a participant's standing is protected by what is never collected, not by what is agreed at the start.

A third dispute runs underneath both. Hertwig and Ortmann (2008) review the empirical evidence on the costs of deception. They find that the claims used to defend it rest more on assertion than on data. Those claims are that the practice has become rare, that deceived participants are not resentful, and that the effects of suspicion are negligible. Their own conclusion is deliberately cautious. The evidence is not as clear-cut as either side implies, and deception should be reserved for closely specified circumstances.

A consent model tested against what participants expect would settle the first dispute, in the form Fiesler and Proferes (2018) pioneered: vary the terms of consent, the kind of study and the route to refusal, and report which combination people accept. Routine pre-release estimation of re-identification risk, of the kind Rocher and colleagues (2019) provide, would settle the second. Neither has been run at the scale of the disagreement.

The positions rest on different kinds of warrant. Grady argues from the practice of consent. Zook and colleagues offer rules rather than a test. Metcalf and Crawford argue that the regulatory categories no longer describe the work. Ohm, and Rocher and colleagues, supply measurements of a technical fault. While the dispute is unresolved, protect the participant by what is not collected as well as by what is agreed. Name the study, the data, the use and the withdrawal route in the consent. The strongest objection is that broad consent is a workable repair rather than a fiction.

Where the argument is heading

The direction is from consent as a signature towards ethics as a property of the system. Shilton (2018) reviews thirty years of work on values in HCI and argues that values are designed into a product or a dataset rather than added at review. Birhane and colleagues (2022) give that claim its hardest evidence: of 100 highly cited machine-learning papers from ICML and NeurIPS, 15 per cent justified how the project connected to a societal need and 1 per cent discussed potential negative consequences, while the most uplifted values were performance, generalisation, efficiency and novelty.

The sources this unit rests on

• World Medical Association (2024). WMA Declaration of Helsinki: ethical principles for medical research involving human participants. https://www.wma.net/policies-post/wma-declaration-of-helsinki/ Adopted a revised Declaration of Helsinki at its assembly in Finland in October 2024, the first revision since 2013. It renames subjects as participants, requires consent for the collection, storage and foreseeable secondary use of identifiable or re-identifiable data, and requires ethics committee approval for databases holding them.
• Kramer, A. D. I., Guillory, J. E. and Hancock, J. T. (2014). Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1320040111 Altered the News Feed of 689,003 users for a week and shifted the emotional tone of later posts. PNAS later attached an expression of concern about consent and opt-out, so a sign-up clause cannot license a specific study.
• PNAS (2014). Editorial Expression of Concern: Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences 111(29), 10779. doi:10.1073/pnas.1412469111 Records that questions had been raised about informed consent and the opportunity to opt out in the Facebook study, and that Cornell University's review board determined the project did not fall under its human research protection programme. The primary record behind the unit's claim that a clause accepted at sign-up cannot license a specific study.
• Grady, C. (2015). Enduring and Emerging Challenges of Informed Consent. New England Journal of Medicine. doi:10.1056/nejmra1411250 Reviews the challenges informed consent now faces, including broad consent for biobanking and research on identifiable data. Consent is how participants keep a choice, so the response to complexity should be better consent rather than less.
• British Psychological Society (2021). Code of Human Research Ethics. British Psychological Society. https://www.bps.org.uk/guideline/bps-code-human-research-ethics States four principles: respect for the autonomy, privacy and dignity of individuals and communities; scientific integrity; social responsibility; and maximising benefit while minimising harm. It derives consent, withdrawal, confidentiality and the limits of deception from them, and is the standard a student protocol is written to.
• Metcalf, J. and Crawford, K. (2016). Where are human subjects in Big Data research? The emerging ethics divide. Big Data and Society. doi:10.1177/2053951716650211 Argues that data science and research ethics have diverged, so established human-subjects regulation no longer describes what data researchers do. It shows that the proposed revision of the Common Rule largely excluded data science methods and public datasets.
• Vitak, J., Shilton, K. and Ashktorab, Z. (2016). Beyond the Belmont Principles. Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work and Social Computing. doi:10.1145/2818048.2820078 Surveys 263 researchers who work with online data and finds the Belmont framework applied inconsistently, with many decisions resting on individual judgement. It shows oversight does not reach the situations data researchers face.
• Fiesler, C. and Proferes, N. (2018). "Participant" Perceptions of Twitter Research Ethics. Social Media + Society. doi:10.1177/2056305118763366 Surveys 394 Twitter users, retaining 368 valid responses and focusing on the 268 with a public main account: 61.2 per cent did not know researchers used tweets, and 64.9 per cent thought researchers should not be able to do so without permission. Participant expectation enters the standard of public availability.
• Narayanan, A. and Shmatikov, V. (2008). Robust De-anonymization of Large Sparse Datasets. 2008 IEEE Symposium on Security and Privacy. doi:10.1109/sp.2008.33 Presents de-anonymisation attacks that recover individuals from large sparse datasets by exploiting the rarity of records, demonstrated on the Netflix Prize dataset of anonymous ratings from 500,000 subscribers. It identified known users and exposed their apparent political preferences, so a stripped dataset is not anonymous.
• Rocher, L., Hendrickx, J. M. and de Montjoye, Y. (2019). Estimating the success of re-identifications in incomplete datasets using generative models. Nature Communications. doi:10.1038/s41467-019-10933-3 Estimates how likely a person is to be re-identified from an incomplete dataset, with AUC scores between 0.84 and 0.97 across 210 populations. It estimated that 99.98 per cent of Americans would be re-identified using 15 attributes. Anonymisation risk becomes a quantity to estimate before release.
• Ohm, P. (2010). Broken promises of privacy: responding to the surprising failure of anonymization. UCLA Law Review. https://www.uclalawreview.org/broken-promises-of-privacy-responding-to-the-surprising-failure-of-anonymization-2/ Reviews the repeated promises that anonymisation would make data sharing safe and shows that they have failed technically. The legal line between identifiable and non-identifiable data cannot carry the weight placed on it.
• Zook, M., Barocas, S., boyd, D., Crawford, K., Keller, E., Gangadharan, S. P., Goodman, A., Hollander, R., Koenig, B. A. and Metcalf, J. (2017). Ten simple rules for responsible big data research. PLOS Computational Biology. doi:10.1371/journal.pcbi.1005399 Sets out ten rules for responsible big data research covering consent, privacy, data sharing, vulnerable groups and re-use, turning the ethical disputes of the period into decisions a project makes.
• Hertwig, R. and Ortmann, A. (2008). Deception in Experiments: Revisiting the Arguments in Its Defence. Ethics and Behaviour. doi:10.1080/10508420701712990 Reviews the evidence on the costs of deceiving participants, testing the claims that deception has become rare, that participants do not resent it, and that suspicion has negligible effects. Those claims rest more on assertion than on data, so deception should be reserved for closely specified circumstances.
• Shilton, K. (2018). Values and Ethics in Human-Computer Interaction. Foundations and Trends in Human-Computer Interaction. doi:10.1561/1100000073 A monograph-length review of thirty years of work on values and ethics in HCI. It maps the field's engagement with privacy, consent, agency and social responsibility, arguing that values are designed into systems rather than added at review.
• Birhane, A., Kalluri, P., Card, D., Agnew, W., Dotan, R. and Bao, M. (2022). The Values Encoded in Machine Learning Research. 2022 ACM Conference on Fairness, Accountability, and Transparency. doi:10.1145/3531146.3533083 Analyses 100 highly cited machine-learning papers from ICML and NeurIPS for encoded values. It finds 15 per cent justified how the project connected to a societal need and 1 per cent discussed potential negative consequences, so ethics are settled by infrastructure.