Clearing the way for participatory data stewardship in artificial intelligence development: a mixed methods approach

Figures & data

Figure 1. Conceptual research model.

Table 1. Descriptive and reliability statistics of scales.

Scale	n	M (SD)	95% CI	α	Range	Skew
PU	319	4.97 (1.29)	[0.76–2.26]	ρ = 0.77^a***	1–7	−0.70
PEOU-ease	320	4.79 (1.48)	[0.45–1.10]	–	1–7	−0.74
Trust	320	3.52 (1.39)	[0.32–0.77]	ρ = 0.59^a***	1–7	0.01
Subjective norm	320	3.70 (1.28)	[0.47–1.17]	0.80***	1–7	−0.18
Social duty	322	3.33 (1.51)	[0.37–0.80]	0.92***	1–7	0.22
Purpose	322	5.20 (1.40)	[0.40–0.83]	0.92***	1–7	−1.19
Guilt	322	2.70 (1.36)	[0.92–2.07]	0.89***	1–7	0.66

Notes: n: valid sample size; M: mean; SD: standard deviation; CI: confidence interval; α: Cronbach’s alpha. 1 = strongly disagree, 7 = strongly agree.

^a Spearman’s rank order (two-tailed) correlation was undertaken as scale comprised of two items.

*** Correlation is significant at p < .001 level, two-tailed.

Table 2. Logistic regression.

		B	SE	df	Sig.	Exp(B)
Personal characteristics
	Age	0.00	0.01	1	0.757	1.00
	Gender	0.53	0.37	1	0.159	0.59
	Race	−0.20	0.37	1	0.599	1.22
	Sexual orientation	0.44	0.38	1	0.249	0.65
eTAM
	PU	−0.27	0.28	1	0.324	1.31
	PEOU-ease	0.35	0.23	1	0.120	0.70
	Trust	0.70	0.22	1	0.002	0.50
	Subjective norms	0.30	0.23	1	0.198	0.74
Data donation profile
	Social duty	0.61	0.20	1	0.002	0.54
	Understanding	0.54	0.19	1	0.003	0.58
	Guilt	−0.32	0.21	1	0.120	1.38

Notes. N = 277. B: unstandardised coefficients; SE: standard error; df: degrees of freedom; Exp(B): exponential coefficients.

Bolded variables were significant predictors in the model.

Table 3. Primary qualitative themes.

Willingness or objection	Theme	Sub-theme	Examples of quotes	Frequency
Willingness
	1. Benefit society		‘If this data will help in the training of AI to help others, such as in medical contexts or to provide samples of other groups (i.e. genders, sexualities, etc)’ (F, 18).	52
		Medical care^a	‘Yes, AI is integral to many aspects of peoples life and if improvement continues can greatly improve their lives like with healthcare such as prosthesis, recognising risks of permanent eye damage, etc’ (gender non-binary, 19).
	2. Technology development		‘If there is bias in the original data set that the AI is trained with, these biases will flow through to the eventual outcomes of the AI. It is important to give the AI the best possible chance to make decisions that are non-discriminatory and safe for all people’ (M, 42).	45
	3. Knowledge		‘As long as I know how the data is being used and what data is required, I don’t see an issue with giving it to help train AI’ (F, 18).	18
Objection
	1. Privacy concerns		‘I am uncomfortable with sharing personal details unnecessarily’ (F, 29).	23
	2. Lack of trust		‘I don’t trust the big companies that control it’ (M, 59).	23
	3. Understanding the purpose		‘I would need to know more about where it was going, who is using it and what for’ (F, 59)	22

^aMedical care was the only scenario (from the educational, organisational, health care and control conditions) mentioned in the qualitative responses.

Table A1. Extended technology acceptance model items.

Measurement items	Items	Adapted from
PU	I think AI would be very useful	Cheng, Lam, and Yeung (2006) Davis (1989)
	I think AI would improve/enhance my ability to live	Cheng, Lam, and Yeung (2006) Davis (1989)
PEOU	I think AI would make my daily life easier	Cheng, Lam, and Yeung (2006) Davis (1989)
	I think interacting with an AI device would be difficult for me	Cheng, Lam, and Yeung (2006) Davis (1989)
Trust	I trust AI to make predictions, recommendations, or decisions influencing real or virtual environments	Choung, David, and Ross (2022) Dinev and Hart (2006)
	AI is a trustworthy channel for me to share my personal details	Dinev and Hart (2006)
Subjective norms	If people close to me used AI, I would too	Ajzen (1991)
	Most people who influence my behaviour would think that I should use AI for daily life	Ajzen (1991)
	Most people whose opinions I value would approve of me using AI for daily life	Ajzen (1991)

Table B1. Data donation items.

Measurement items	Items
Social duty to help	1. I would donate my data to AI research because I believe that I have a responsibility to help others. 2. If I receive a request to donate my data to AI research, I would consider it a social responsibility to do so. 3. When I receive a request to donate my data to AI research, I would automatically offer my data. 4. I would donate my data to AI research, even if no gratitude was shown in return. 5. I would donate my data to AI research because I feel that I have a duty to give back to the community.
Understanding the purpose	6. I would make the decision to donate my data to AI research depending on the purpose of research. 7. I would make the decision to donate my data to AI research based on how they would deal with my personal data. 8. I would make the decision to donate my data to AI research based on how the data will be used. 9. Before donating my data to AI research, I would seek to understand the purpose of giving data for research. 10. Before donating my data to AI research, I would seek to understand how my data could help others. 11. I would make a decision to donate my data to AI research depending on what they would do with my data. 12. I would make a decision to donate my data to AI research based on who the data would be shared with.
Guilt, reputation, self-image	13. I would donate my data to AI research as this could relieve some guilt felt for being more fortunate than others. 14. Through donating my data to AI research, I would think of those who are unlucky/ill-fated and this would help me to forget how bad I have been feeling myself. 15. I would donate my data to AI research as I wish to be praised and have good reputation. 16. By donating my data to AI research, I would be able to show people that I am a good and kind person. 17. If I did not donate my data to AI research, I would feel less guilty if others did the same. 18. By taking interest in societal issues through donation of my data to AI research, I would feel less stressed about my own problems.
Adapted from Skatova and Goulding (2019)

Table C1. Scenarios.

Condition	Excerpt
Health care	‘This has grave implications when considering AI that is used for health care. It was recently publicised that an AI mole scanner was trained on predominantly Caucasian skin tones, resulting in the machine not detecting cancerous moles on other skin types (e.g. Black, Asian, Indigenous and Hispanic skin tones). This is due to the overrepresentation of cis-gendered, Caucasian men in existing datasets and underrepresentation of other ethnicities, genders and races (to list a few)’.
Organisational	‘This has grave implications when considering AI that is used for organisational use. Notably, companies have developed AI-based recruitment systems that filters through resumes, conducts interviews and recommends candidates. For instance, Amazon, used ten-years of employees’ data to build a model based on terms frequently found in their resumes Proxies for male candidates based on masculine language began to occur due to the overrepresentation of male resumes in the data. As such, female resumes were downgraded and the system only recommended men’.
Education	‘This has grave implications when considering AI that is used for educational use. For instance, Intel has developed AI-software that is used in tandem with Zoom to recognise facial expressions to monitor how students’ are interacting with the content. Negative repercussions of this technology arise when considering neurodivergent students, who may avert eye gaze or exhibit other expressions that could be misinterpreted. As such, accurate representation of all students is required to drive successful use of AI in education’.
Control	‘To counteract the prevalence of large, foundational models, scholars are calling for the rise of participatory data stewardship (PDS) in AI. PDS encourages individuals, including those who are historically disenfranchised, to regain control and rebalance asymmetries of traditional data collection used to train AI. This occurs via the submission of personal data by an individual for the purpose of training AI. Further, by allowing the individual to control the ownership of their data via private use, pooling their data with likeminded people or as a common good’.

Table D1. Pearson correlation coefficients.

	1	2	3	4	5	6	7	8	9	10	11	12
1. Willingness
2. Age	−0.053
3. Gender	0.154***	−0.201***
4. Race	0.018	0.203***	0.108
5. Sexual Orientation	0.061	0.105	−0.062	−0.011
6. Perceived usefulness	0.463***	0.077	−0.163***	0.019	−0.047
7. Perceived ease of use	0.460***	0.108	−0.168***	0.041	−0.042	−0.720***
8. Trust	0.552***	0.100	−0.147***	0.027	0.011	−0.591***	−0.527***
9. Subjective norms	0.475***	0.252***	−0.090	−0.071	0.013	−0.600***	−0.509***	−0.738***
10. Social duty	0.538***	0.101	−0.071	−0.002	−0.007	−0.564***	−0.492***	−0.664***	−0.567***
11. Understanding	0.416***	0.086	−0.024	0.010	−0.020	−0.520***	−0.402***	−0.337***	−0.347***	−0.437***
12. Guilt	0.239***	0.334***	0.120*	−0.039	0.075	−0.286***	−0.279***	−0.498***	−0.515***	−0.581***	−0.113*

Note. n = 278. Gender, race and sexual orientation are measured as binary items.

*** Correlation significant at the 0.001 level (two-tailed). *Correlation significant at the 0.05 level (two-tailed).

Cheng, T.E., D.Y. Lam, and A.C. Yeung. 2006. “Adoption of Internet Banking: An Empirical Study in Hong Kong.” Decision Support Systems 42 (3): 1558–1572. doi:10.1016/j.dss.2006.01.002.

 Web of Science ®Google Scholar

Davis, F.D. 1989. “Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology.” MIS Quarterly 13 (3): 319–340. doi:10.2307/249008.

 Web of Science ®Google Scholar

Choung, H., P. David, and A. Ross. 2022. “Trust in ai and Its Role in the Acceptance of ai Technologies.” International Journal of Human–Computer Interaction 39 (9): 1727–1739. doi:10.1080/10447318.2022.2050543.

 Web of Science ®Google Scholar

Dinev, T., and P. Hart. 2006. “An Extended Privacy Calculus Model for e-Commerce Transactions.” Information Systems Research 17 (1): 61–80. doi:10.1287/isre.1060.0080.

 Web of Science ®Google Scholar

Ajzen, I. 1991. “The Theory of Planned Behavior.” Organizational Behavior and Human Decision Processes 50 (2): 179–211. doi:10.1016/0749-5978(91)90020-T.

 Web of Science ®Google Scholar

Skatova, A., and J. Goulding. 2019. “Psychology of Personal Data Donation.” PLoS One 14 (11): e0224240. doi:10.1371/journal.pone.0224240.

 PubMed Web of Science ®Google Scholar