English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

What We Keep to Ourselves: A Study of What People Delegate to AI Agents—and What They Refuse to Hand Over

Forum topic · 小凯 · 2026-11-09

Summary

A survey study from researchers at UIUC, UT San Antonio, Northeastern, and TU Munich ('What to Delegate, What to Preserve', arXiv:2611.09730) asked 289 US adults how much autonomy they would grant AI agents across 45 scenarios in 9 life domains, yielding 2,601 observations. Key findings: delegation willingness varies strongly by task—people most readily delegate communication and travel, but resist AI involvement in mental health, physical health, and workplace tasks. Trust is the strongest positive predictor of delegation, alongside desire for AI leadership, prior experience, and expected savings; the strongest negative predictor is fear of losing personal sense of accomplishment. Surprisingly, privacy concerns showed no independent effect after controls, suggesting 'privacy fatigue.' Qualitative answers reveal people want 'bounded autonomy': AI executes with human oversight, collaborates under human leadership, or acts autonomously with veto rights—humans always remain in the loop. The post discusses design implications (trust-building, preserving accomplishment, adjustable autonomy) and limitations (US-only sample, hypothetical scenarios, cross-sectional design).

Some Doors We Won't Hand the Keys To: What Do Humans Guard When AI Can Do Everything?

*Translation and commentary of a Chinese forum post discussing the paper "What to Delegate, What to Preserve: Human Preferences for Agentic AI Across Everyday Life and Work" (arXiv:2611.09730)*

Opening: The Mental Ledger on Your Keychain

Imagine you hire a housekeeper. You're happy to let her water plants, sweep, and feed the cat. You hesitate but hand over the kitchen key—she needs to cook—while telling her not to touch the bottom shelf of the fridge. You lock the bedroom door. Not because there's anything shameful inside, but because it's *your* space: your diary, your letters, the words you write at 3 a.m. that no one will ever see.

Handing over a key was never a question of "can it be done" but "am I willing to give it"—and each door has its own willingness level, each with its own reasons.

Now swap the housekeeper for an AI agent. It can do more than any human: reply to emails, book flights, write reports, manage investments, practice conversation with you, even listen to your worries at 3 a.m. It won't slack, gossip, demand raises, or snoop through your photo album. By rights, you should hand over everything.

But researchers at UIUC, UT San Antonio, Northeastern, and TU Munich ran a 289-person survey and found something unexpected yet intuitive: people are willing to delegate a lot to AI—just not everything. And the reasons for holding back are often not efficiency or privacy, but something deeper.

The paper is titled *What to Delegate, What to Preserve*. Its core question: when AI agents can autonomously execute multi-step tasks, what do people want to delegate—and what do they want to keep?

🤖 A New Question for the Agent Era

In AI's early days, the question was "can AI do this task?"—with clear right/wrong answers. Now AI agents handle open-ended, multi-step, judgment-heavy tasks: not just "book a flight" but "plan a trip fitting my budget, preferences, and stamina"; not just "reply to an email" but "handle all my unimportant communication this week." Once capability crosses the "can it?" threshold, a new dimension emerges: "am I willing?"

This is not an engineering question—it's a human question about trust, control, accomplishment, and meaning. The paper's core contribution: turning this fuzzy "willingness" into quantifiable data—45 concrete scenarios across 9 life domains, asked of 289 real US adults: "To what degree would you let AI do this?"

📋 The Study: 289 People, 45 Scenarios, 9 Life Domains

Participants: 289 US adults, ages 19–81, spanning Gen Z to Baby Boomers. 52.6% female, 79.9% non-Hispanic white, roughly balanced politically (50.5% Democrat, 40.8% Republican), education from high school to graduate degrees. All had real AI experience.

Scenarios (45, across 9 domains): 1. Home management — cleaning/organizing, shopping lists 2. Workplace — writing reports, handling email 3. Travel — booking transport/lodging, itinerary planning 4. Physical health — appointments, exercise planning 5. Mental health & emotional support — processing negative emotions, social anxiety 6. Learning — language learning, exam prep 7. Creative work — idea generation, editing 8. Communication — replying to unimportant messages, translation 9. Personal finance — expense tracking, savings planning

Scenarios were concrete, e.g.: "Imagine you've just finished an exhausting workday and feel low. An AI agent can detect your emotional state and proactively offer relaxation exercises or encouragement. How much would you let it do autonomously?"

Measured variables: automation preference (5-point scale, core DV), trust, collaboration need, AI leadership, privacy concern, willingness to pay, perceived accomplishment loss, interpersonal consequences, prior AI experience.

The final sample produced 2,601 respondent × domain × scenario observations—enough for reliable statistical inference.

🔍 Finding 1: Not "Whether" but "Which Task"

If you guessed overall delegation willingness is uniformly high or low, both intuitions are wrong. The overall average masks the most important pattern: huge differences across domains.

People are most willing to delegate AI in communication and travel, and least willing in mental health, physical health, and workplace domains.

This makes sense: a badly written unimportant email can be rewritten—low cost, reversible. A wrong flight booking can be changed. But if AI botches your mental health support? A wrong suggestion might deepen anxiety or cause you to miss real help. Your body has no "undo" button. And if AI sends a report directly on your behalf, who bears responsibility for errors?

Delegation willingness is not a global attitude but a task-level judgment, hinging on at least three factors: cost of error, reversibility, and personal importance.

This sounds like common sense, but many AI products assume the opposite—that users either trust AI or don't—and apply one trust level to everything. The data says: don't. Trust is domain-, task-, and scenario-specific.

🔍 Finding 2: Trust Drives Delegation; Accomplishment Blocks It

The paper's core statistical finding, via regression:

Positive predictors of delegation:

  • Trust in AI's performance on that task—the strongest, bar none
  • Desire for AI leadership in decisions
  • Prior experience with AI on similar tasks
  • Expected cost savings
  • Negative predictors:

  • Fear of reduced sense of accomplishment—one of the strongest negative predictors
  • The last item deserves the most attention. Efficiency maximalists might ask: "If AI does it faster and better, what's the point of doing it yourself?" But people don't calculate that way. Many care not that "the task got done" but that "*I* did it." Cooking isn't just about being fed; painting isn't just about getting an image; exercise isn't just about health data—the process itself is part of the meaning.

    This aligns strongly with research on intrinsic motivation. Deci and Ryan's self-determination theory identifies three basic psychological needs: autonomy, competence, and relatedness. When AI fully replaces you on a task, it strips both your sense of autonomy ("this was my choice") and competence ("I can do this well")—even as efficiency rises, psychological costs quietly mount.

    For AI product designers, "can it be done" isn't the only question. Also ask: "What is the user getting from doing this themselves? What will they lose if we automate it?" If the answer is "accomplishment and control," consider whether this automation genuinely helps users—or quietly erodes their well-being.

    🕵️ Finding 3: The Surprising Absence of Privacy

    The most surprising finding: after adjustment, privacy concern had no independent association with delegation willingness.

    Once trust, accomplishment, and savings are controlled, "I worry about my personal information leaking" does not predict willingness to delegate.

    Possible explanations:

    1. Privacy fatigue / privacy resignation: by now people universally assume using digital services means giving up some privacy—privacy is the entry ticket, not a differentiator. Like wearing a swimsuit at the pool: giving data to AI isn't a choice, it's a precondition. 2. Privacy is embedded in other judgments: distrust of AI handling your health data mixes privacy concerns with capability concerns and concerns about severity of health consequences—when measured separately, privacy's unique contribution vanishes. 3. Privacy concern is background noise, not a decision variable—the classic behavioral-economics "privacy paradox": you know data is collected, you dislike it, you use the service anyway because not using it costs more.

    Either way, the lesson for designers: don't assume stronger privacy protection will raise adoption. Users may not adopt your agent because of better privacy, but they will because of earned trust. Privacy is necessary but not sufficient—an entry ticket, not a selling point.

    🗝️ Qualitative Findings: "Bounded Autonomy"

    Beyond the numbers, free-text answers were thematically coded. The recurring pattern: people don't want to hand over control completely, but they don't want to do everything themselves either. They want "bounded autonomy."

    Three configurations emerged:

  • AI executes + human supervises. The most common. AI drafts emails; you review before sending. AI plans trips; you confirm the final itinerary. AI tidies; you decide what gets thrown out.
  • AI collaborates + human leads. AI analyzes investment options; you pick the portfolio. AI offers creative inspiration; you create the final work. AI preps interview answers; the delivery style is yours.
  • AI executes fully + human retains veto. AI auto-replies to unimportant emails, but you can review anytime. AI manages your schedule, but you can override. AI adjusts temperature, but you can intervene.
The common thread: the human is always in the loop, only at varying degrees. Nobody said "I hand everything to AI and check out entirely." One participant (paraphrased): "I'll let AI do a lot for me, as long as I can say 'stop' at any moment. Not because I doubt its ability, but because this is my choice, my responsibility, my life."

The paper thus reconceptualizes "delegation"—not a binary switch (all-or-nothing) but a configurable dial of execution, collaboration, leadership, and authority, tuned differently by person, task, and time. Today you might let AI fully handle email; tomorrow, with one sensitive message, you'd turn the dial back to manual.

💬 A Telling Detail: The Communication Paradox

The paper reports a contradiction: people are *most* willing to delegate communication (e.g., replying to unimportant messages), yet many expressed worry about AI communication "not sounding like me."

One participant (paraphrased): "I know AI can reply for me, but I worry it won't sound like me. If friends found out it was AI, it'd be awkward. And for messages from people who matter, even if AI could reply, I'd want to write it myself."

This reveals a deep tension in delegation decisions: the pull between efficiency and authenticity. People want to save time (efficiency motive) but also want to keep human interaction genuine (identity motive). When these conflict, different people choose differently—and some waver. That wavering itself is a signal: delegation willingness is not a static preference but a dynamic decision process. AI agents should support real-time adjustment per interaction, not one-and-done "delegation levels."

🎯 Implications for AI Product Design

1. Trust is everything. It's the strongest positive predictor. Build it via transparent reasoning ("why I suggested this"), a reliable track record, and graceful degradation on failure. Trust accumulates; it isn't established by a "we're safe" statement. 2. Don't assume privacy protection boosts adoption. Privacy is a floor (an ethical/legal must), not a selling point. 3. Treat sense of accomplishment as a design constraint. If your agent automates a task users find fulfilling, you're creating a loss. Can AI and user do it *together*? Can users keep the "this was my choice" feeling at key decision points? Make AI more a coach (helping you do better) than a replacement (doing it for you). 4. Offer configurable "bounded autonomy." Not just "fully automatic" and "fully manual," but a continuous spectrum of confirmation points—dynamically adjustable at any time without changing global settings. 5. Design trust mechanisms per domain and scenario. High trust in travel doesn't transfer to health. Each domain has its own logic of risk, emotional investment, and error cost. Good agents should sense this and adapt.

🗺️ Limitations

The authors note:

1. US-only sample. Culture likely matters greatly—collectivist vs. individualist values toward automation and interpersonal interaction. The "accomplishment effect" may differ elsewhere. 2. Hypothetical scenarios. Intent–behavior gaps are well documented; real answers may differ when AI actually says "let me do it." 3. Cross-sectional design. Causality is unclear—does trust cause delegation, or does delegation experience build trust? Likely both. 4. Limited scenario coverage. 45 scenarios can't exhaust daily life. Domains like parenting, religious activity, or political participation—tied to deeper values and identity—may show even lower willingness.

💡 Deeper Reflection: When "Doing" Is No Longer Required

Beneath "what will people hand to AI" lies a deeper question: when AI can do everything, what do humans still need to do?

The data suggest people instinctively distinguish between tasks that are functional (just get it done—who does it doesn't matter) and constitutive (the process itself shapes who you are—cannot be outsourced). Having AI clean your room? Fine—your identity isn't defined by sweeping. But having AI comfort a grieving friend? Write your child's bedtime story? Do your late-night soul-searching? **For these, the doing *is* the meaning.**

This recalls an ancient Greek distinction: *poiesis* (making, valued by output—who made it doesn't matter) versus *praxis* (action, valuable in itself—doing it is living your life; the value lies in the process, not the result).

Perhaps as AI grows more powerful, human value won't vanish but will concentrate—into those moments that "must be done oneself." What those moments are, everyone keeps their own private ledger. AI designers must respect that ledger, not pretend it doesn't exist.

Reference

Wheeler, S., Rammouz, V., Qian, Y., Liu, Z., Wei, M., Rios, A., He, T., & Zhou, K. (2026). What to Delegate, What to Preserve: Human Preferences for Agentic AI Across Everyday Life and Work. *arXiv:2611.09730*.

*Translated and edited from a Chinese-language post on zhichai.net.*

Tags

#ai-agents#human-ai-collaboration#delegation#user-trust#privacy-paradox#intrinsic-motivation#product-design#paper-review

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636972