"Better Ask for Forgiveness than Permission": Practices and Policies of AI Disclosure in Freelance Work

Angel Hsing-Chi Hwang; Senya Wong; Baixiao Chen; Jessica He; Hyo Jin Do

cs.HCMar 8, 2026

"Better Ask for Forgiveness than Permission": Practices and Policies of AI Disclosure in Freelance Work

Authors: Angel Hsing-Chi Hwang, Senya Wong, Baixiao Chen, Jessica He, Hyo Jin Do

Organizations: University of Southern California · Emory University · IBM Research

Abstract

The growing use of AI applications among freelance workers is reshaping trust and relationships with clients. This paper investigates how both workers and clients perceive AI use and disclosure in the freelance economy through a three-stage study: interviews with workers and two survey studies with workers and clients. Findings first reveal a key expectation gap around disclosure: Workers often adopt passive disclosure practices, revealing AI use only when asked, as they assume clients can already detect it. Clients, however, are far less confident in recognizing AI-assisted work and prefer proactive disclosure. A second finding highlights the role of unclear or absent client AI policies, which leave workers consistently misinterpreting clients' expectations for AI use and disclosure. Together, these gaps point to the need for clearer guidelines and practices for AI disclosure. Insights extend beyond freelancing, offering implications for trust, accountability, and policy design in other AI-mediated work domains.

Figures & tables

Figure 1. Workers and clients’ perceptions of common freelance tasks as major or minor tasks for AI use. A grid chart compares how workers (blue circles labeled “W”) and clients (green circles labeled “C”) classify different work tasks as minor or major when considering AI use. The chart has three vertical categories: Definitely Minor Task, Could Be Minor or Major Task, and Definitely Major Task. Five tasks are listed on the y-axis: Communication, Research, Data Analysis, Idea Generation, and Text Editing. For each task, one or two circles appear, showing how each group categorizes it: Communication: Workers classify it as definitely a minor task. Clients classify it as possibly minor or major. Research: Workers place it in the possibly minor or major category. Clients place it in the definitely major category. Data Analysis: Workers classify it as possibly minor or major. Clients also classify it as possibly minor or major. Idea Generation: Workers classify it as possibly minor or major. Clients classify it as definitely major. Text Editing: Workers classify it as definitely minor. Clients classify it as definitely minor. A legend at the bottom identifies blue circles as Workers and green circles as Clients. Overall, the figure highlights areas of divergence—especially in Research and Idea Generation—where clients view tasks as more major than workers do, revealing mismatched mental models of what constitutes a “major” task for AI involvement.

Type of disclosure practice	Description
Non-Disclosure or Avoidance n = 3 (7%)	A minority of workers stated they would not disclose AI use at all. Reasons included believing it was irrelevant (“it’s just a tool”), fear of clients misinterpreting or undervaluing their work, or concern that disclosure might reduce trust or jeopardize future contracts.
Passive Disclosure (Disclose only if asked) n = 16, (39%)	The most common stance is that workers disclose AI use only when clients explicitly ask about it. Many framed this as a passive agreement—AI use does not need to be mentioned unless disclosure is requested. In these cases, workers typically give broad or high-level explanations (e.g., “I used AI to draft ideas” or “I used it as a helper”) without going into technical detail.
Situational or Conditional Disclosure n = 8, (20%)	Workers also described disclosure as context-dependent. For sensitive or competitive projects, they felt disclosure might be more important, while for routine or obvious AI-supported tasks (e.g., transcription, AI-generated voiceovers), disclosure seemed unnecessary. Some reported tailoring the level of detail based on the client’s familiarity with AI or the perceived risks.
Qualified or Framed Disclosure n = 9, (22%)	A sizable group disclosed with disclaimers or framing. They reassured clients that AI was used only as a supportive tool—for brainstorming, proofreading, or saving time—while emphasizing that the final decisions, accuracy checks, and creative control remained with them. This strategy aimed to normalize AI as akin to other everyday tools (e.g., “like a calculator or spell-checker”) while preserving credibility.
Proactive or Full Transparency n = 5, (12%)	Some workers preferred open, upfront disclosure. They described mentioning AI use in proposals, contracts, or notes attached to deliverables. For them, honesty and professionalism required clarifying where AI was involved, especially in tasks like writing, editing, or research. A subset even emphasized “disclosing all parts where AI was used,” sometimes because their employers or clients explicitly required it.

Table 1. Five types of workers’ AI disclosure practice

Figure 2. Clients and workers also differed in their attitudes toward disclosure. Significantly more workers than clients believed that disclosure of AI use is not required, whereas significantly more clients expected workers to actively disclose when they use AI. A stacked bar chart compares clients’ and workers’ AI disclosure practices. Each bar is divided into three categories: Disclose actively (light green, top), Disclose passively (medium gray-green, middle), and No disclosure (dark blue, bottom). For clients, the bar shows roughly one-third “No disclosure,” about one-third “Disclose passively,” and a smaller portion “Disclose actively.” For workers, the bar shows a larger share of “No disclosure,” a substantial share of “Disclose passively,” and a small portion “Disclose actively.” The visualization highlights that workers are less likely than clients to disclose their AI use proactively, with a higher proportion choosing not to disclose at all.

Figure 3. Clients’ and workers’ responses to whether workers are allowed to use AI for their freelance work. The majority of clients and workers both believe AI use is allowed. The portion of clients allowing AI use is slightly larger than workers’ expectations. Alt-text: A stacked bar chart titled “Can workers use AI for work?” shows two vertical bars: one for clients and one for workers. Each bar is divided into “Yes” (green) and “No” (dark blue). For clients, the bar is approximately 78\% Yes and 22\% No. For workers, the bar is about 66\% Yes and 34\% No. The figure highlights that workers underestimate how often clients permit the use of AI in contract work. — Figure 3. Clients’ and workers’ responses to whether workers are allowed to use AI for their freelance work. The majority of clients and workers both believe AI use is allowed. The portion of clients allowing AI use is slightly larger than workers’ expectations. Alt-text: A stacked bar chart titled “Can workers use AI for work?” shows two vertical bars: one for clients and one for workers. Each bar is divided into “Yes” (green) and “No” (dark blue). For clients, the bar is approximately 78% Yes and 22% No. For workers, the bar is about 66% Yes and 34% No. The figure highlights that workers underestimate how often clients permit the use of AI in contract work.

Figure 4. Permitted AI use according to clients’ AI policies vs. workers’ interpretation of permitted AI use based on each policy. A faceted bar chart compares workers’ interpretations of how much AI use is permitted with the actual AI allowances stated in client policies. The horizontal axis in each panel shows five possible interpretations: No AI, Use AI for minor tasks, Use AI for major tasks, Use AI as much as possible, and Use AI to automate work. Bars represent the proportion of workers who interpreted the policy in each way, with percentages labeled. There are five panels, each representing a different actual policy allowance: First, No AI Allowed: Most workers interpret the policy correctly as “No AI,” with small proportions misinterpreting it as allowing minor tasks or automating work. Second, Use AI for Minor Tasks: Workers’ interpretations spread across categories, with the largest share interpreting it as “Use AI for minor tasks,” and substantial percentages believing major tasks or heavy AI use is also allowed. Use AI for Major Tasks: Workers again show distributed interpretations, with notable shares choosing “minor tasks,” “major tasks,” and “use AI as much as possible.” Third, Use AI as Much as Possible: Workers’ interpretations vary widely, but the largest proportion selects “Use AI as much as possible,” with others selecting adjacent categories. Fifth, Use AI to Automate Work: Many workers correctly interpret the policy as allowing automation, but others interpret it as permitting heavy use or major task support. Overall, the figure highlights pervasive misalignment between what policies actually allow and how workers interpret them, with misinterpretations occurring in every policy category.

Policy Type	$\beta$	$S.E.$	$t$	$p$
(Intercept)	$0.31$	$0.11$	$2.83$	$0.005^{**}$
Use AI for minor tasks	$0.46$	$0.12$	$3.69$	$<0.001^{***}$
Use AI for major tasks	$-0.45$	$0.14$	$-3.15$	$0.002^{**}$
Use AI as much as possible	$-0.84$	$0.12$	$-7.20$	$<0.001^{***}$
Use AI to automate work	$-1.71$	$0.14$	$-11.94$	$<0.001^{***}$

Table 2. Workers’ interpretation of permitted AI use by policy

Figure 5. Difference between permitted AI use and workers’ interpretations of AI policies. Positive values indicate that workers believed they were allowed to use AI to a greater extent than the policies actually permitted, while negative values indicate that workers interpreted the policies more restrictively than intended. A horizontal bar chart shows the difference between workers’ interpretations of permitted AI use and the actual policy allowances across five policy categories. The x-axis represents the interpretation–policy difference, where positive values indicate workers interpreted the policy as allowing more AI use than intended, and negative values indicate workers interpreted it as allowing less. The y-axis lists the actual policy categories from top to bottom: No AI, Use AI for minor tasks, Use AI for major tasks, Use AI as much as possible, and Use AI to automate work. Each bar shows the mean difference for that category, with a small horizontal black error bar indicating the confidence interval. The pattern indicates: No AI and Use AI for minor tasks both have positive differences, meaning workers tend to overestimate allowed AI use. Use AI for major tasks and Use AI as much as possible are near zero or slightly negative, indicating mixed or slightly conservative interpretations. Use AI to automate work has a large negative difference, meaning workers significantly underestimate how much AI use is allowed. Overall, the figure illustrates systematically different directions of misinterpretation depending on policy strictness.

Policy Type	$t$	$p$	Cohen’s $d$
No AI use	$3.66$	$<0.001^{***}$	0.38
Use AI for minor tasks	$10.30$	$<0.001^{***}$	0.74
Use AI for major tasks	$-1.25$	$0.214$	$-0.13$
Use AI as much as possible	$8.14$	$<0.001^{***}$	$-0.48$
Use AI to fully automate tasks	$13.35$	$<0.001^{***}$	$-1.36$

Table 3. Interpretation gap by AI policy type

Policy Type	$\beta$	$S.E.$	$t$	$p$
Baseline: Use AI for minor tasks
(Intercept)	$0.69$	$0.06$	$12.03$	$<0.001^{***}$
Use AI for major tasks	$0.12$	$0.09$	$1.31$	$0.189$
Use AI as much as possible	$-0.21$	$0.07$	$-3.08$	$0.002^{**}$
Use AI to automate work	$-0.40$	$0.09$	$-4.32$	$<0.001^{***}$

Table 4. Workers’ interpretation of specific AI use case by policy

Figure 6. Workers’ interpretations of clients’ disclosure requirements under different AI policies. Workers were more likely to report proactive disclosure of AI use when policies encouraged AI use. A stacked bar chart shows how workers’ approaches to AI disclosure vary across five types of client AI-use policies. The x-axis lists the policy categories: No AI, Use AI for minor tasks, Use AI for major tasks, Use AI as much as possible, and Use AI to automate work. The y-axis represents the proportion of workers (in percent) within each policy category. Each bar is divided into four colors: Yellow = Disclose actively. Green = Disclose passively. Teal/blue = No need to disclose. Black = No AI use. Patterns across the bars show that: Under No AI policies, a large proportion of workers report not using AI (black) or needing no disclosure (teal), while smaller shares report passive or active disclosure. As policies become more permissive (minor to major tasks to heavy AI use), passive and active disclosure both increase, and the share reporting no AI use decreases to very small levels. Under the most permissive policies (Use AI to automate work), the majority of workers disclose—either actively or passively—with very little reporting “no AI use.” Overall, the figure highlights that more permissive client policies correlate with greater worker openness to AI disclosure, especially active disclosure.

Policy Type	$\beta$	$S.E.$	$t$	$p$
Baseline: No AI use
(Intercept)	$0.64$	$0.09$	$6.93$	$<0.001^{***}$
Use AI for minor tasks	$1.37$	$0.11$	$12.59$	$<0.001^{***}$
Use AI for major tasks	$1.44$	$0.13$	$11.49$	$<0.001^{***}$
Use AI as much as possible	$1.65$	$0.10$	$16.05$	$<0.001^{***}$
Use AI to automate work	$1.44$	$0.13$	$11.49$	$<0.001^{***}$

Table 5. Workers’ interpretation of required AI disclosure by policy type

Figure 7. Comparison of clients’ expectations and workers’ interpretations of AI disclosure under different policies. Workers frequently misjudged situations where clients deemed disclosure unnecessary, and consistently overestimated the extent to which clients expected proactive disclosure of AI use. A multi-panel bar chart compares clients’ expectations for AI disclosure with workers’ interpretations of those expectations across five types of AI-use policies. Each panel represents one policy category along the y-axis: No AI, Use AI for minor tasks, Use AI for major tasks, Use AI as much as possible, and Use AI to automate work. Within each panel, pairs of bars appear for four disclosure approaches shown on the x-axis: No AI use, No need to disclose, Disclose passively, and Disclose actively. Light gray bars represent client expectations. Dark green bars represent worker interpretations. Percentages are labeled above each bar. Key visual patterns: Under No AI policies, clients expect very little AI use or disclosure, yet workers frequently interpret the policy as requiring passive or even active disclosure. For minor task and major task policies, clients expect mostly passive disclosure, but workers are more split across passive and active disclosure. Under more permissive policies (Use AI as much as possible and automate work), client expectations shift strongly toward active disclosure, and workers broadly mirror this—though workers still overestimate the need to disclose compared to clients in several categories. Overall, the figure highlights systematic mismatches between what clients expect and what workers think clients expect, especially around when disclosure is required and how proactively it should be done.

Clients’ expectation for AI disclosure	$t$	$p$	Cohen’s $d$
No AI use	$2.53$	$0.022^{*}$	0.63
No need to disclose	$15.01$	$<0.001^{***}$	0.96
Disclose passively	$1.48$	$0.140$	$0.08$
Disclose actively	$-12.21$	$<0.001^{***}$	$-0.87$

Table 6. Difference between clients’ expectations vs. workers’ interpretation of AI disclosure.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Demographics	Study 1.1	Study 1.2	Study 2	Study 3
Gender
Male	46.3%	53%	51.38%	50.00%
Female	46.3%	39%	33.33%	48.96%
Non-binary/third gender	4.9%	1%	1.39%	1.04%
Prefer not to disclose	2.4%	7%	2.78%	0%
Age	40.35 $\pm$ 11.29	36.4 $\pm$ 9.8	42.31 $\pm$ 11.20	37.29 $\pm$ 9.88

Appendix

Table 14

Freelance Service Type	Study 1.1	Study 1.2	Study 2	Study 3
Healthcare and Medicine	12%	7%	16%	14%
Information Technology (IT) and Engineering	23%	27%	23%	22%
Education and Training	8%	5%	6%	8%
Sales and Marketing	15%	12%	13%	14%
Finance and Accounting	5%	18%	13%	10%
Human Resources (HR) and Recruitment	15%	2%	16%	1%

Appendix

Table 15

AI Use Disclosure

See all 10 papers

Aug 24, 2026

Expectations and Practices around AI Disclosure in CS Research

As generative AI tools find increasing use in research workflows, ongoing debates on their impact, appropriateness and responsible use have led policymakers to enact policies to disclose AI use at multiple publishing venues. However, are current AI disclosure policies and practices reflective of their purpose? In this work, we first investigate disclosure policies of top computer science venues and find that despite their prevalence, they remain highly under-specified. Secondly, through a survey of computer science researchers (N=

109

), we characterize the necessity of disclosures across different research tasks and levels of human involvement. We learn that researchers find disclosures most necessary for tasks involving research design, and for tasks when the human involvement is low. We also compile expectations that researchers have about the information to be conveyed in AI disclosure statements. Lastly, through an analysis of

13867

disclosure statements from EMNLP

2025

and ICLR

2026

, we reveal a large disconnect between these expectations and AI disclosures in practice---a prime example being writing assistance which is deemed less necessary but is frequently disclosed. We conclude with recommendations to align AI disclosure policies and practices with expectations, suggesting a categorization of research tasks by perceived necessity and a boilerplate template capturing expected details.

Arati Mohapatra, Danish Pruthi

Indian Institute of Science Bengaluru, KA, India

AI Use Disclosure AI Governance Responsible AI Scientific Communication AI-Assisted Scientific Research

Ask PDF

Apr 29, 2026

Upskilling with Generative AI: Practices and Challenges for Freelance Knowledge Workers

Freelance workers must continually acquire new skills to remain competitive in online labor markets, yet they lack the organizational training, mentorship, and infrastructure available to traditional employees. Generative AI-powered tools like ChatGPT are reshaping market skill demands while also offering new forms of on-demand learning support to meet those demands. Despite growing interest in AI-powered learning tools, little is known about how freelancers actually use these tools to learn, the challenges they encounter, and how generative AI for learning interacts with precarity and competition in platform-based work. We present a mixed-methods study combining a survey and semi-structured interviews with freelance knowledge workers. Grounded in self-directed learning theory, we examine how freelancers integrate generative AI tools into their learning practices. Our findings show that freelancers increasingly rely on generative AI to structure learning and support exploratory skill acquisition, but do not treat it as their primary learning resource due to inconsistency, lack of contextual relevance, and verification overhead. We identify a shift from learning as growth to learning as survival, where upskilling is oriented toward immediate market viability rather than long-term development. We also surface a structural challenge we term invisible competencies, in which workers acquire skills through generative AI tools but lack credible ways to signal or validate these skills in competitive freelance markets. Based on these insights, we offer design recommendations for generative AI-powered learning tools for freelancers.

Kashif Imteyaz, Isabel Lopez, Nakul Rajpal +2

Civic AI Lab, Northeastern University Boston, Massachusetts, USA · CICESE Ensenada, Baja California, Mexico · Universidad Nacional Autonoma de Mexico (UNAM) Mexico City, Mexico

Generative AI in Education

Ask PDF

Jun 9, 2026

Designed by Journalists, but Is It for Readers? Rethinking AI Disclosures and Transparency in News

As newsrooms integrate generative AI, journalists face a disclosure challenge: how to communicate AI involvement in ways that maintain reader trust. Current practice offers two approaches: brief one-line labels or detailed disclosures specifying human oversight, editorial accountability, and error reporting mechanisms. Neither achieves journalists' goal of building trust through transparency. An existing controlled experiment with 34 news readers show that detailed disclosures trigger a transparency dilemma, reducing trust rather than increasing it, and risk introducing dark patterns that readers scroll past with the illusion of transparency. One-line disclosures avoid this effect but can create an information gap, prompting readers to expend cognitive effort searching for signs of AI involvement that the disclosure indicates but does not explain. Yet readers are not rejecting transparency, they proposed disclosure designs centered on user agency: detail-on-demand interactions, proportional AI-ratio visualizations, outlet-level signals, and explicit "no AI" labels. I argue that this disconnect between what practitioners believe is responsible disclosure and what users actually need is a design problem for the HCI community.

Pooja Prajod

Centrum Wiskunde & Informatica Amsterdam, the Netherlands

Human-AI Interaction Trust in AI AI Use Disclosure

Ask PDF

Abstract

AI Use Disclosure

Explore similar work

Expectations and Practices around AI Disclosure in CS Research

Upskilling with Generative AI: Practices and Challenges for Freelance Knowledge Workers

Designed by Journalists, but Is It for Readers? Rethinking AI Disclosures and Transparency in News