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

This paper reveals a critical expectation gap in the freelance economy where workers often withhold AI use due to a mistaken belief that clients can detect it, while clients prefer proactive disclosure and lack clear policies, ultimately highlighting the urgent need for standardized guidelines to rebuild trust and accountability in AI-mediated work.

The Gist

Imagine the world of freelance work as a massive, bustling open-air marketplace. In this market, independent workers (freelancers) set up stalls to sell their skills—writing, coding, designing—to customers (clients) who walk by looking for help.

Now, imagine a new, magical tool has arrived in this marketplace: AI. It's like a super-fast, invisible assistant that can draft emails, write code, or design logos in seconds.

This paper is a story about the awkward silence and misunderstandings that happen when workers start using this magic tool, but nobody has agreed on the rules of the game yet. The authors call this the "Better Ask for Forgiveness than Permission" dilemma.

Here is the breakdown of what's happening, using simple analogies:

1. The Great "Can You Tell?" Mismatch

Imagine you are a freelancer using AI to write a story. You think, "I bet my client can tell I used a robot because the writing is so perfect." So, you stay quiet, hoping they won't ask.

But the client is thinking, "I have no idea if this was written by a human or a robot. I'm actually quite bad at spotting the difference."

• The Reality: The study found that workers are overconfident (thinking clients can spot AI easily), while clients are underconfident (they can't spot it at all).
• The Result: Workers stay silent, thinking they are being smart. Clients get suspicious, thinking the worker is hiding something.

2. The "Passive Disclosure" Habit

Because of the fear of getting in trouble, most freelancers have adopted a strategy the authors call "Passive Disclosure."

Think of it like this: You are driving a car with a new, experimental autopilot feature. You don't tell the passenger (the client) you turned it on. You only admit it if the passenger asks, "Hey, did you drive this yourself?"

• Why? Workers feel that bringing it up unprompted is annoying, like telling your boss, "I used a calculator to do my math." They assume if the client doesn't ask, it's a "silent yes."
• The Problem: Clients actually want to know! They prefer an honest conversation upfront, but they rarely know how to ask without sounding suspicious.

3. The "Minor vs. Major" Confusion

Clients and workers speak different languages when it comes to what AI is allowed to do.

• The Client's View: They might say, "You can use AI for small things, but not for the big, important stuff."
• The Worker's View: They hear that and think, "Okay, I can use AI to write the email, but not the final report."
• The Clash: The study found that they often disagree on what counts as "small" or "big."
  • Example: A worker thinks "researching facts" is a small task (minor). A client thinks "researching facts" is the core of the job (major).
  • Result: The worker uses AI for research, thinking they are following the rules. The client gets angry, thinking the worker broke the rules.

4. The "Vague Rulebook" Problem

When clients do try to write rules (policies), they are often written in foggy, confusing language.

• The Analogy: Imagine a client hands you a rulebook that says: "Use common sense with AI," or "Don't use AI for sensitive topics."
• The Problem: What is "common sense"? What is "sensitive"? It's like being told, "Don't eat the red berries," without being told which berries are red or if the red ones are poisonous.
• The Outcome: Workers guess. Sometimes they guess right; often, they guess wrong. Because the rules are so vague, workers end up either being too scared to use AI (even when allowed) or using it too freely (when forbidden).

5. The "Forgiveness" Strategy

The title of the paper, "Better Ask for Forgiveness than Permission," perfectly captures the freelancers' mindset.

• The Mindset: Freelancers think, "If I ask for permission to use AI, the client might say no, and I'll lose the job. But if I just use it and do a great job, the client will be happy. If they get mad later, I'll just apologize."
• The Risk: This is a high-stakes gamble. If the client finds out later, trust is broken, and the freelancer might lose their reputation in the marketplace.

The Big Takeaway: What Needs to Change?

The authors argue that we can't just rely on freelancers and clients to figure this out on their own. The current system is broken because:

1. Rules are too vague.
2. Expectations don't match.
3. Trust is fragile.

The Solution?
We need better "traffic signs" for the AI marketplace. Instead of saying "No AI" or "Use AI," we need clear, specific guides:

• "You can use AI to check grammar (like a spellchecker)."
• "You cannot use AI to write the final story."
• "If you use AI for the outline, just put a little note saying 'Outline assisted by AI'."

In short: The paper says that for the freelance world to work smoothly with AI, we need to stop playing "guessing games." We need clear, honest, and specific rules so that workers don't have to choose between being efficient and being honest.

Technical Summary

Here is a detailed technical summary of the paper "Better Ask for Forgiveness than Permission: Practices and Policies of AI Disclosure in Freelance Work."

1. Problem Statement

The rapid adoption of Generative AI in the freelance economy has created a critical expectation gap regarding AI use and disclosure between workers and clients. Unlike traditional corporate environments with established protocols, freelance relationships are often short-term, decentralized, and lack clear institutional guidance.

• The Core Conflict: Freelancers often adopt passive disclosure strategies (revealing AI use only when asked), assuming clients can detect AI or that silence implies consent. Conversely, clients often lack the ability to detect AI and prefer proactive disclosure.
• The Risk: This misalignment leads to "perceptual harms" (negative evaluation of work quality), eroded trust, and potential contract disputes. Furthermore, existing AI policies are often vague, binary (allow/prohibit), or non-existent, leading to systematic misinterpretation by workers.
• The Stakes: In freelance work, reputation is paramount. Workers fear that disclosing AI use will lead to contract termination or negative reviews, leading many to conceal usage ("ask for forgiveness later") despite the potential for long-term trust damage.

2. Methodology

The authors conducted a three-stage mixed-method study involving 386 participants (freelance workers and clients) across multiple freelance platforms.

• Study 1.1: Qualitative Interviews (N=41 Workers)

  • Goal: To profile current AI disclosure practices and concerns.
  • Method: 45–60 minute Zoom interviews analyzed via open coding.
  • Outcome: Identified five distinct disclosure strategies and the rationale behind "passive" approaches.

• Study 1.2: Worker Survey (N=100 Workers)

  • Goal: To validate findings from Study 1.1 on a larger scale.
  • Method: Quantitative survey on AI use frequency, disclosure types (none, passive, active), and experiences with client policies.

• Study 2: Client Survey (N=145 Clients)

  • Goal: To mirror worker perspectives and understand client expectations.
  • Method: Surveyed clients on their AI policies, detection confidence, and preferred disclosure practices.
  • Key Output: Collected actual text of client AI policies for analysis.

• Study 3: Policy Interpretation Experiment (N=100 New Workers)

  • Goal: To test how workers interpret specific client policies.
  • Method: Workers were presented with five policy categories (e.g., "No AI," "AI for minor tasks," "AI for major tasks," "Use AI as much as possible," "Full Automation") and asked to interpret permitted use and required disclosure.
  • Analysis: Compared worker interpretations against the clients' stated intentions to measure the "interpretation gap."

3. Key Contributions & Findings

A. The "Passive Disclosure" Norm

The study identified five types of AI disclosure practices, with Passive Disclosure being the most prevalent (39% of workers).

• Passive Disclosure: Workers reveal AI use only when explicitly asked. They view proactive disclosure as unnecessary, awkward, or a sign of micromanagement.
• Other Strategies: Non-disclosure (7%), Situational/Conditional (20%), Qualified/Framed (22%), and Proactive/Full Transparency (12%).
• Driver: Workers assume clients can detect AI (63.5% believe clients can detect it 75% of the time), whereas clients report very low confidence in detection (Mean = 3.11/5).

B. The Expectation Gap

There is a statistically significant misalignment between worker and client expectations:

• Disclosure: 53.8% of workers believe disclosure is unnecessary, compared to only 40.7% of clients. Conversely, 25.9% of clients expect proactive disclosure, while only 8.6% of workers intend to provide it.
• Perception of AI Capability: Clients often hold "inflated expectations" of AI, assuming workers can deliver faster/better results using AI, while workers fear that using AI will be penalized.

C. Systematic Policy Misinterpretation

When explicit policies exist, workers frequently misinterpret them, particularly regarding the scope of permitted use:

• Restrictive Policies: When clients say "Use AI for minor tasks," workers often overestimate what is allowed, interpreting "minor" more broadly than intended.
• Permissive Policies: When clients say "Use AI for major tasks" or "As much as possible," workers often underestimate the permission, fearing they are still restricted.
• Specific Use Cases: Disagreement is highest on tasks like "Idea Generation" (clients view as major; workers view as minor if used for brainstorming) and "Written Communication" (clients view client-facing emails as major; workers view drafting as minor).

D. Flaws in Client AI Policies

Qualitative analysis of client policies revealed three structural failures:

1. Vague Language: Reliance on terms like "common sense," "human judgment," or "sensitive topics" without definitions.
2. Binary Framing: Policies often use "No AI" or "Full AI" without nuance, failing to account for the spectrum of AI assistance (e.g., proofreading vs. content generation).
3. Process vs. Outcome: Policies often focus on "final decision stages" (e.g., "don't let AI make the final call") but ignore how earlier workflow steps (research, drafting) critically shape the output.

4. Significance and Implications

Theoretical Contribution

The paper challenges the binary view of AI disclosure (used vs. not used). It argues that disclosure is a relational and contextual negotiation, not just a compliance task. It highlights that "perceptual harms" (fear of being judged as less skilled) drive secrecy, even when AI use is permitted.

Practical Design Implications

The authors propose shifting from binary compliance to proportional, stage-aware frameworks:

• Disclosure Ladder: Instead of a yes/no checkbox, policies should require disclosure based on the materiality of AI involvement (e.g., no disclosure for grammar checks; summary required for drafting; pre-approval for full automation).
• Policy Templates: Platforms should provide specific examples of "Permitted," "Prohibited," and "Disclosure Required" scenarios (e.g., "AI allowed for brainstorming, prohibited for final client-facing copy").
• Shared Literacy: Platforms must act as intermediaries to build "AI regulation literacy," offering resources that clarify what "minor" vs. "major" tasks mean in specific domains.

Broader Impact

While focused on freelancers, the findings offer a "micro-laboratory" for understanding AI governance in decentralized work. The misalignments observed here (ambiguity, lack of trust, policy failure) are likely to scale as AI regulation (e.g., EU AI Act) becomes mandatory in corporate settings. The paper suggests that effective AI governance requires usability-focused policy design that bridges the gap between high-level rules and ground-level workflow realities.

Conclusion

The paper concludes that the current state of AI disclosure in freelance work is characterized by a "better ask for forgiveness than permission" mentality, driven by policy ambiguity and fear of reputational damage. To resolve this, stakeholders must move beyond vague prohibitions toward granular, scenario-based guidelines that normalize transparency and align the expectations of workers and clients through shared literacy and platform-supported negotiation.