Navigating the Evolving Landscape of Human Reliance on Generative AI

5 minutes ago

26 min read

Listen to a Review of this Article:

Abstract: The rapid diffusion of generative artificial intelligence into workplace and personal contexts has created substantial uncertainty regarding how individuals rely on these systems over time. This article examines longitudinal patterns in AI delegation and disclosure behaviors, drawing on evidence from a six-wave study spanning ten months with over 1,000 U.S. participants. Contrary to expectations of increasing familiarity breeding greater reliance, findings reveal declining willingness to delegate tasks and disclose information to AI systems, particularly for personally meaningful activities. Professional writing contexts maintained relatively stable delegation patterns, while personal contexts showed pronounced declines. Trust, anthropomorphic perceptions, and positive attitudes toward AI emerged as consistent predictors of both delegation and disclosure behaviors. These patterns suggest that human-AI interaction reflects a calibration process rather than simple habituation, with important implications for organizational AI adoption strategies, training programs, and technology design. Organizations must recognize that sustainable AI integration depends less on repeated exposure and more on cultivating well-calibrated trust, contextually appropriate applications, and meaningful transparency mechanisms.

The workplace integration of generative AI has accelerated at unprecedented speed since late 2022, fundamentally altering how knowledge workers approach complex cognitive tasks. Large language models now assist with activities ranging from content generation and data analysis to decision support and customer service, positioning AI not merely as a productivity tool but as an interaction partner in knowledge work. Recent survey evidence indicates that approximately 15-20% of U.S. workers now use AI chatbots regularly in their jobs, with information search, content editing, and document drafting among the most common workplace applications.

Yet this rapid diffusion has outpaced our understanding of how individuals actually rely on these systems across time and context. While cross-sectional research has identified numerous factors influencing initial AI adoption—including perceived usefulness, trust, and technical competence—far less is known about the temporal dynamics of AI reliance. Do individuals become more comfortable delegating tasks to AI with repeated exposure? Do disclosure behaviors increase as familiarity grows? Or do users instead calibrate their expectations downward as they encounter system limitations?

These questions carry substantial practical significance. Organizations investing in AI technologies need evidence-based guidance on how user behaviors evolve, which factors sustainably promote appropriate reliance, and how contextual characteristics shape delegation decisions. The present article addresses these gaps by examining longitudinal patterns in two critical dimensions of AI reliance: willingness to delegate writing tasks and willingness to disclose personal information. Drawing on six waves of survey data collected over ten months, we analyze how task characteristics, user attributes, and perceptions of AI systems jointly influence these behaviors—and how these relationships change across repeated interactions.

The stakes extend beyond individual productivity. Organizations must balance efficiency gains against risks including over-reliance on potentially inaccurate outputs, erosion of critical thinking skills, and privacy concerns associated with data disclosure. Understanding temporal patterns in AI reliance is therefore essential for developing training programs, governance frameworks, and technology designs that foster appropriate—rather than excessive or insufficient—dependence on AI systems.

The Generative AI Adoption Landscape

Defining Delegation and Disclosure in AI Contexts

Two behavioral dimensions capture distinct aspects of how individuals rely on AI systems: delegation and disclosure. Delegation refers to the extent to which individuals transfer responsibility for task execution to an AI system. This encompasses decisions about whether to use AI assistance at all, how much control to retain versus relinquish, and which specific activities to assign to the system versus performing personally. Delegation exists along a continuum from no AI involvement through various collaborative configurations to full automation.

Importantly, delegation is not always a fully deliberate process. While users may consciously decide to outsource specific tasks, reliance can also emerge gradually through repeated interactions that foster increasing trust and reduce monitoring of AI outputs. This progressive reduction in oversight can lead to automation bias—the tendency to favor AI recommendations even when they conflict with contradictory information or personal judgment.

Disclosure concerns the information individuals share with AI systems, whether through intentional revelation of personal details, behavioral patterns observable through usage, or data inferrable from linguistic inputs and interaction contexts. As AI systems become more conversational and socially interactive, disclosure takes on relational dimensions similar to human self-disclosure, requiring users to evaluate not only functional benefits but also privacy risks, vulnerability, and trust.

The distinction between delegation and disclosure proves consequential. Delegation primarily involves functional decisions about task performance and outcome quality, making perceived competence and reliability especially salient. Disclosure, by contrast, involves relational considerations including trust, social presence, and emotional safety. These differential emphases suggest that the two behaviors may respond to partially distinct organizational interventions.

Prevalence, Drivers, and Distribution of AI Use

Generative AI adoption has followed a distinctive diffusion pattern characterized by rapid initial uptake concentrated among knowledge workers, followed by gradual expansion into broader occupational categories. Current estimates suggest that 15-20% of U.S. workers use AI chatbots with some regularity in their professional roles, with substantially higher penetration in technology, consulting, marketing, and creative industries.

Use cases cluster around several core categories:

Information retrieval and research: Quickly gathering background information, identifying relevant resources, and synthesizing content from multiple sources
Content generation and editing: Drafting initial versions of documents, refining language, and adapting tone for different audiences
Data analysis and summarization: Processing large volumes of text, identifying patterns, and generating executive summaries
Brainstorming and ideation: Exploring alternative approaches, generating creative options, and overcoming initial blocks
Communication support: Drafting emails, preparing meeting agendas, and translating content across languages

Demographic patterns reveal important adoption disparities. Younger workers, those with higher educational attainment, and individuals in professional or technical roles show consistently higher AI usage rates. Gender differences appear relatively modest in overall usage but may be more pronounced in specific applications or levels of reliance. These patterns mirror broader technology adoption trends while also reflecting task-related factors—AI systems currently perform best on the types of knowledge work disproportionately performed by educated professionals.

Attitudinal factors significantly shape adoption patterns beyond demographic characteristics. Individuals with more positive general attitudes toward AI, greater perceived usefulness for specific tasks, and higher trust in system capabilities show substantially higher usage rates. Conversely, privacy concerns, fear of job displacement, and skepticism about output quality inhibit adoption. These attitudinal dimensions often matter more than technical knowledge per se, suggesting that organizational interventions targeting perceptions and concerns may prove more effective than those focused solely on skill development.

The context of AI use also varies substantially across organizations and occupational categories. Some workplaces actively encourage AI adoption through formal policies, training programs, and integration with existing workflows. Others maintain more ambivalent or restrictive postures, particularly in heavily regulated industries or roles involving confidential information. This organizational context shapes not only whether individuals use AI but also how they use it—with more supportive environments associated with more sophisticated, integrated applications rather than occasional, supplementary use.

Organizational and Individual Consequences of AI Delegation and Disclosure

Organizational Performance Impacts

The integration of generative AI into workplace processes produces measurable effects on productivity, quality, and operational efficiency, though magnitudes vary substantially across contexts and implementation approaches. Experimental evidence examining writing tasks found that AI assistance improved productivity by approximately 40% and quality ratings by roughly 18% among professional consultants. These gains concentrated particularly among initially lower-performing workers, suggesting AI may help reduce performance variance within teams.

Similar patterns emerge across diverse task domains. Customer service applications show reduced resolution times by 13-14% on average, with more pronounced benefits for less experienced representatives. Code generation tasks demonstrate 30-50% reductions in completion time, though with notable variance in output quality depending on task complexity and the extent of human oversight. Document summarization, data extraction, and report generation likewise show consistent time savings, typically ranging from 20-45% depending on the specific application.

However, these productivity gains carry important caveats. First, improvements tend to be largest for relatively routine or well-structured tasks where evaluation criteria are clear. More complex, ambiguous, or creative challenges show smaller and more variable effects. Second, performance improvements depend heavily on the quality of AI integration into existing workflows rather than the capabilities of the AI system alone. Organizations that simply provide access to AI tools without process redesign or training often observe minimal productivity gains.

Third, and perhaps most critically, short-term productivity improvements may come at the cost of longer-term skill development. When individuals delegate cognitively demanding tasks to AI, they may lose opportunities to develop expertise, make nuanced judgments, and build tacit knowledge. Educational research shows concerning evidence of reduced critical thinking and deeper processing when students rely heavily on AI for assignments. Similar dynamics may operate in workplace contexts, where delegation could gradually erode capabilities essential for handling novel situations or exercising professional judgment.

Quality considerations present a more mixed picture. AI assistance often improves surface-level quality dimensions such as clarity, organization, and grammatical correctness. Professional documents produced with AI support typically receive higher ratings on these attributes. However, AI may simultaneously reduce originality, contextual appropriateness, and the type of creative problem-solving that distinguishes exceptional from merely competent work. Organizations must therefore consider not just whether AI improves average performance but whether it shifts the distribution of outcomes in desirable directions.

From a financial perspective, AI adoption can reduce operational costs through time savings, enabling organizations to handle higher volumes with existing staff or to redeploy personnel toward higher-value activities. A major consulting firm reported saving approximately $500 million annually through AI-assisted document generation and research. Customer service applications show similar savings through reduced handle times and increased first-contact resolution rates.

Individual Wellbeing and Stakeholder Impacts

Beyond organizational performance, AI delegation and disclosure behaviors affect individual wellbeing, professional identity, and career trajectories. These impacts prove more difficult to quantify but are no less consequential for sustainable AI integration.

Autonomy and Control: Delegation inherently involves ceding control over aspects of work that may be central to professional identity and job satisfaction. Knowledge workers often derive meaning from creative problem-solving, expert judgment, and personal authorship of work products. When AI systems perform substantial portions of these activities, individuals may experience reduced autonomy and diminished sense of ownership over outcomes. This is particularly salient for personal writing tasks, where authenticity and self-expression carry inherent value beyond functional outcomes.

Conversely, delegation can also enhance autonomy by automating routine aspects of work and freeing time for more engaging activities. The key distinction appears to be whether individuals choose which tasks to delegate versus having delegation imposed through organizational mandates. Self-directed AI use that genuinely supports personally meaningful work tends to enhance satisfaction, while enforced delegation that reduces control over core activities often diminishes it.

Skill Development and Career Progression: Over-reliance on AI for complex cognitive tasks creates risks for long-term skill development. When individuals consistently delegate challenging aspects of their work, they may fail to develop expertise necessary for career advancement or for handling situations where AI systems prove inadequate. This concern mirrors historical patterns with calculator use in mathematics or GPS navigation in spatial cognition—tools that provide immediate benefits while potentially atrophying underlying capabilities.

Educational contexts provide cautionary evidence. Students who extensively use AI for writing assignments show reduced development of critical thinking, argument construction, and analytical skills compared to those who use AI more selectively as a supplementary tool. While workplace contexts differ in important respects, similar patterns may emerge if organizations fail to ensure that employees retain opportunities to develop and exercise complex judgment.

Privacy and Vulnerability: Information disclosure to AI systems creates privacy risks extending beyond traditional data security concerns. Language-based AI systems process and potentially retain substantial personal information, including sensitive details users may not realize they are revealing. Research examining conversations with AI chatbots found that users frequently disclosed personal information about relationships, health conditions, and other private matters—information that could be used for purposes beyond the immediate interaction.

Organizations bear responsibility for ensuring that workplace AI use does not pressure employees to disclose sensitive information, either explicitly through data requirements or implicitly through system designs that function poorly without substantial personal information. This becomes particularly important as AI systems grow more sophisticated at eliciting disclosure through conversational engagement and perceived empathy.

Job Security Concerns: Delegation to AI inevitably raises concerns about job displacement, even when automation is partial rather than complete. Workers who delegate substantial portions of their responsibilities may wonder whether their roles remain necessary or whether their contributions are sufficiently distinctive to justify continued employment. These concerns can create psychological stress and may paradoxically discourage adoption of AI tools that could genuinely enhance productivity without threatening job security.

Organizations can partially address these concerns through transparent communication about AI's intended role, explicit commitments regarding employment stability, and active redeployment of time saved through AI use toward activities that build rather than erode job security. However, when automation genuinely does reduce labor requirements, organizations face difficult tradeoffs between efficiency and employment that warrant careful ethical deliberation.

Evidence-Based Organizational Responses

Table 1: Organizational Case Studies and Strategies for Generative AI Integration

Organization	Integration Approach	Specific Use Cases	Key Outcomes and Performance Impacts	Governance and Risk Controls	Employee Training and Capability Building	Evidence-Based Strategy Recommended
Microsoft	Transparent communication and expectation management	AI assistance within Microsoft 365 suite; research and drafting tasks.	Sustainable adoption patterns and reduced instances of over-reliance on AI outputs.	Internal documentation identifying tasks for human oversight vs. AI; feedback mechanisms for reporting errors and successes.	Capability demonstrations featuring both successful outputs and error examples.	Cultivate well-calibrated trust by proactively addressing limitations and maintaining continuous communication as capabilities evolve.
Unilever	Procedural justice and participatory implementation	Marketing content generation.	Adoption rates exceeding 75%; high employee satisfaction; appropriate skepticism regarding AI outputs.	Cross-functional AI councils (marketing, legal, brand, technical) determining use cases and review requirements.	Pilot testing and structured feedback loops that allow frontline workers to shape implementation decisions.	Utilize participatory processes and pre-implementation consultation to incorporate frontline expertise and reduce resistance.
JPMorgan Chase	Top-down governance and risk-based controls	Internal operations and customer-facing services.	Enabled extensive AI adoption while maintaining risk controls; identification and correction of inappropriate AI use through audits.	Three-tier risk framework (automated, assisted, prohibited); AI governance office; quarterly usage audits and registry of approved applications.	Individual coaching based on audit findings; clarification of system-wide guidance for inappropriate tasks.	Implement risk-based categorization of tasks and mandatory human review for high-consequence categories.
IBM	Privacy-protective design and employee-centric transparency	Internal AI assistants for employees.	High levels of employee trust; stable disclosure willingness (avoiding the typical decline seen in other research).	Local device processing to minimize server exposure; 7-day data retention limit; automatic stripping of identifying metadata.	Personal privacy dashboards allowing employees to review and delete records; education on privacy-protective practices.	Support appropriate disclosure through technical safeguards, opt-in defaults, and clear communication about data usage.
Deloitte	Contextualized training and role-specific adoption	Consulting workflows, ranging from research/drafting to complex analysis.	Not in source	Tiered oversight based on consultant seniority; heavy oversight for entry-level research tasks.	Multi-tiered capability program with role-specific modules; judgment development via case studies and quarterly refresher sessions.	Focus training on judgment (when to use AI and how to evaluate it) rather than just technical mechanics or prompt engineering.

Transparent Communication and Expectation Management

Organizations must recognize that sustainable AI adoption depends on well-calibrated expectations rather than uncritical enthusiasm. Transparent communication about both capabilities and limitations proves essential for fostering appropriate reliance.

Evidence Base: Longitudinal survey evidence tracking 1,008 U.S. adults across six measurement waves found that willingness to delegate writing tasks to AI declined slightly over time rather than increasing as hypothesized. This pattern suggests that users calibrate expectations downward as they encounter system limitations through repeated interactions. Organizations that fail to proactively address limitations may therefore experience declining adoption and engagement as initial enthusiasm gives way to disappointment.

Research on trust calibration in human-automation interaction similarly demonstrates that over-trust—initially common with novel AI systems—typically decreases after users experience errors or limitations. This recalibration can be beneficial when it corrects excessive reliance, but may overcorrect into under-trust that prevents appropriate use of genuinely helpful capabilities.

Effective Approaches:

Capability demonstrations with error examples: Rather than showcasing only successful AI outputs, organizations should include examples of common errors, limitations, and inappropriate applications. One technology company implements "failure showcase" sessions where employees share instances where AI produced incorrect or unhelpful outputs alongside successful applications.
Task-appropriate usage guidelines: Clear frameworks help employees determine which tasks are well-suited for delegation versus those requiring human judgment. A financial services firm developed a three-category system: "AI-suitable" (routine data extraction and formatting), "AI-assisted" (draft generation with mandatory human review), and "Human-led" (complex analysis requiring professional judgment).
Ongoing communication as capabilities evolve: AI system capabilities change rapidly through updates and new model releases. Organizations should establish regular communication channels—monthly briefings, internal newsletters, or dedicated collaboration platforms—that keep employees informed about capability changes rather than treating initial training as sufficient.
Emphasis on complementarity over replacement: Messaging should consistently position AI as augmenting rather than replacing human capabilities. A consulting firm reframed AI adoption from "automation" to "amplification," emphasizing how AI extends what consultants can accomplish rather than making them redundant.

Microsoft: Following the integration of AI assistance into its Microsoft 365 suite, the company implemented a comprehensive communication strategy addressing both capabilities and limitations. Internal documentation explicitly identifies task categories where AI performs well versus those where human oversight remains essential. The company established feedback mechanisms allowing employees to report both successful applications and limitations, creating a continuous learning loop that informs both technology development and usage guidance. This transparent approach has been associated with more sustainable adoption patterns and reduced instances of over-reliance on AI outputs.

Procedural Justice and Participatory Implementation

How organizations implement AI systems matters as much as what they implement. Research consistently demonstrates that participatory processes that give affected employees voice in adoption decisions result in better outcomes across multiple dimensions.

Evidence Base: Organizational change research shows that participation in implementation decisions enhances acceptance, reduces resistance, and improves the quality of implementation by incorporating frontline expertise. This general principle applies particularly strongly to AI adoption, where workers often possess crucial tacit knowledge about task requirements, contextual factors, and practical constraints that technology designers may overlook.

Studies examining AI implementation in healthcare, customer service, and manufacturing contexts consistently find that participatory approaches—where frontline workers help determine which tasks to automate, what information AI systems need, and how to integrate AI into existing workflows—produce better outcomes than top-down mandates. These benefits include higher adoption rates, more appropriate usage patterns, and better alignment between AI capabilities and actual work requirements.

Effective Approaches:

Pre-implementation consultation and piloting: Organizations should engage employees in the assessment phase before committing to specific AI tools. This involves discussing pain points in current workflows, evaluating potential AI applications collaboratively, and conducting small-scale pilots with representative users before organization-wide rollout.
Frontline worker involvement in configuration: Rather than imposing standardized AI implementations, organizations should involve workers in customizing systems for their specific contexts. This might include choosing prompt templates, determining what information gets shared with AI systems, or establishing review protocols appropriate for different task types.
Establishment of ongoing governance mechanisms: Effective participation extends beyond initial implementation. Organizations should create standing committees or working groups—including frontline employees alongside management and technical staff—that continuously assess AI performance, address emerging concerns, and recommend adjustments.
Transparent decision-making about what not to automate: Organizations should explicitly communicate decisions about tasks that will remain human-led despite technical feasibility, explaining the reasoning based on factors such as stakeholder preferences, ethical considerations, or strategic importance of human judgment.

Unilever: The consumer goods company established cross-functional AI councils when implementing AI assistance for marketing content generation. These councils—comprising marketing professionals, legal staff, brand managers, and technical experts—collectively determined appropriate use cases, established review requirements, and developed guidelines for maintaining brand voice and values in AI-assisted content. Employees participated in pilot testing and provided structured feedback that shaped implementation decisions. This participatory approach has been credited with achieving adoption rates exceeding 75% while maintaining high employee satisfaction and appropriate skepticism about AI outputs.

Capability Building Through Contextualized Training

Traditional technology training often emphasizes technical mechanics—how to use specific features or execute particular commands. Effective AI training instead focuses on judgment: determining when to use AI, how to evaluate outputs, and how to maintain critical engagement with AI-generated content.

Evidence Base: Survey research demonstrates that objective knowledge about AI—measured through assessments of technical understanding—shows inconsistent relationships with delegation and disclosure behaviors. In contrast, experiential familiarity with AI systems and positive attitudes toward AI use consistently predict greater willingness to delegate tasks and disclose information. This pattern suggests that effective capability building emphasizes hands-on experience and critical evaluation over abstract technical knowledge.

Educational research examining AI use in academic contexts similarly finds that training focused on evaluating AI outputs, identifying limitations, and making appropriate delegation decisions proves more effective than instruction in prompt engineering or technical capabilities alone.

Effective Approaches:

Task-based learning with realistic scenarios: Training should center on actual work tasks employees perform, demonstrating AI application in authentic contexts rather than decontextualized examples. Employees should practice with representative scenarios, generate outputs, evaluate quality, and revise as necessary.
Explicit instruction in output evaluation: Organizations should teach specific evaluation criteria relevant to different task types. For writing tasks, this includes assessing factual accuracy, logical coherence, appropriate tone, and alignment with organizational standards. For data analysis, evaluation focuses on methodological soundness, consideration of alternative interpretations, and appropriate caveats.
Comparative analysis exercises: Training should include exercises where employees compare human-generated, AI-generated, and collaboratively-produced outputs, discussing strengths and weaknesses of each approach. This builds calibrated understanding of comparative advantages rather than uncritical adoption or blanket rejection.
Longitudinal skill development rather than one-time training: Organizations should establish ongoing learning opportunities as employees gain experience and as AI capabilities evolve. This might include monthly workshops, peer learning sessions where employees share effective practices, or online communities of practice.

Deloitte: The professional services firm implemented a multi-tiered AI capability program when integrating generative AI into consulting workflows. Rather than standardized training, the program offers role-specific modules addressing different levels of AI sophistication. Entry-level consultants focus on using AI for research and initial drafts with heavy oversight, while senior consultants learn to guide AI for complex analysis while maintaining critical evaluation. The program emphasizes judgment development through case studies where participants evaluate actual AI outputs, identify errors or limitations, and discuss appropriate use. Quarterly refresher sessions address emerging capabilities and share lessons learned across the organization.

Governance Frameworks and Appropriate Use Controls

While capability building addresses individual judgment, organizational governance provides structural safeguards ensuring AI use remains aligned with strategic objectives, ethical standards, and quality requirements.

Evidence Base: Research examining delegation patterns across professional and personal writing tasks found consistently higher willingness to delegate professional tasks, but with notable variance depending on task characteristics. Professional tasks perceived as routine or standardized showed highest delegation willingness, while those involving complex judgment, sensitive information, or significant consequences showed greater preference for human control. This suggests that appropriate governance should differentiate among tasks rather than applying uniform policies.

Studies of automation bias and over-reliance demonstrate that even well-intentioned users often fail to detect errors in AI outputs, particularly after developing trust through repeated interactions. Structural controls therefore complement individual judgment rather than serving as redundant safeguards.

Effective Approaches:

Risk-based categorization of tasks and applications: Organizations should develop frameworks that classify tasks based on potential consequences of errors, sensitivity of information involved, and extent of external visibility. High-risk categories require more stringent review, while lower-risk applications allow greater automation.
Mandatory human review for specified categories: Policies should clearly identify tasks where AI may assist but not independently complete work. This includes client-facing communications, financial analysis informing major decisions, legal advice, and personnel matters. Review requirements should specify not just that review occurs but what reviewers should assess.
Output attribution and audit trails: Systems should maintain records of AI involvement in work products, including what prompts were used, what outputs were generated, and what modifications humans made. This serves both accountability purposes and organizational learning about effective practices.
Periodic audits of actual usage patterns: Organizations should regularly assess whether AI use in practice aligns with policies and guidelines. This includes identifying frequent misuse patterns that may indicate either unclear guidance or impractical policies requiring revision.
Escalation pathways for novel situations: Governance frameworks should include clear mechanisms for employees to raise questions or concerns about appropriate AI use in ambiguous situations, with designated decision-makers who provide timely guidance.

JPMorgan Chase: The financial institution implemented a comprehensive governance framework when deploying AI assistance for internal operations and customer-facing services. The framework categorizes applications across three risk tiers with corresponding review requirements: automated (routine data processing with monthly spot checks), assisted (human-in-the-loop with mandatory review of all outputs), and prohibited (decisions requiring fiduciary judgment). The bank established an AI governance office that approves new use cases, conducts quarterly usage audits, and maintains a registry of approved applications. Audit findings revealed several instances where well-intentioned employees used AI for inappropriate tasks, leading to both individual coaching and clarification of system-wide guidance. This governance structure has enabled extensive AI adoption while maintaining strong risk controls appropriate for a regulated financial institution.

Supporting Appropriate Disclosure Through Privacy-Protective Design

Willingness to disclose information to AI systems depends heavily on trust, perceived privacy protection, and clarity about how information will be used. Organizations must implement both technical safeguards and clear communication about data practices.

Evidence Base: Longitudinal research found that willingness to disclose information to AI declined over time similarly to delegation willingness, though disclosure showed stronger relationships with perceived anthropomorphism and social actor perceptions. This suggests that disclosure responds particularly strongly to relational aspects of AI interaction rather than purely functional considerations.

Privacy research demonstrates that disclosure behaviors often reflect incomplete understanding of actual risks. Users frequently underestimate how much information they reveal through conversational AI, what can be inferred from interaction patterns, and how disclosed information might be used beyond the immediate context. Organizations therefore cannot rely solely on user judgment but must implement structural protections.

Effective Approaches:

Clear data governance policies: Organizations should establish and communicate explicit policies about what data is collected from AI interactions, how long it is retained, who can access it, and for what purposes it may be used. Policies should specifically address whether interaction data is used for training AI models, performance evaluation, or other secondary purposes.
Technical privacy protections: Where possible, organizations should implement technical measures that limit data collection and retention. This includes local processing that avoids sending sensitive information to external servers, automatic deletion of interaction logs after specified periods, and anonymization of data used for system improvement.
User controls over information sharing: Employees should have meaningful choices about what information they share with AI systems. This might include opt-in rather than opt-out defaults for collecting detailed usage data, options to use AI capabilities without creating permanent records, or tiered access where more capable AI features require accepting greater data sharing.
Transparency about inference and profiling: Organizations should communicate not just what explicit information is collected but what implicit attributes or patterns might be inferred from AI interactions. This includes behavioral patterns, performance indicators, or personal characteristics that could be derived from how individuals use AI systems.
Education about privacy-protective practices: Training should include guidance on protecting sensitive information when using AI, such as removing identifying details before submitting confidential information, using alternative phrasing to avoid revealing personal information, and being aware of what context AI systems might retain from earlier in conversations.

IBM: When deploying internal AI assistants for employees, the company implemented a privacy-protective design approach that minimizes data collection while maintaining functionality. The system processes most queries locally on employee devices rather than sending all interactions to central servers, reducing exposure of potentially sensitive information. For interactions that do require server processing, the system automatically strips identifying metadata and applies a seven-day retention limit unless users explicitly opt in to longer retention for their own records. IBM provides employees with a personal privacy dashboard showing what data has been collected from their AI use, with options to review and delete records. Internal surveys show high levels of employee trust in the system, with disclosure willingness remaining stable rather than declining as observed in external research.

Building Long-Term Organizational AI Capability

Cultivating Calibrated Trust Through Transparency

The longitudinal pattern of declining willingness to delegate and disclose suggests that initial enthusiasm often gives way to disappointment as users encounter limitations. Organizations should therefore focus not on maximizing initial adoption but on fostering calibrated trust that accurately reflects both capabilities and constraints.

Evidence Base: Research consistently identifies trust as a central predictor of both delegation and disclosure behaviors. However, the nature of trust matters substantially. Well-calibrated trust—where confidence in AI corresponds appropriately to actual reliability and capability—enables appropriate reliance, whereas both over-trust and under-trust produce suboptimal outcomes.

Studies examining trust calibration across multiple interaction episodes demonstrate that transparency about system limitations helps users develop realistic expectations that remain stable rather than deteriorating. When users understand what AI can and cannot do reliably, they maintain appropriate usage patterns across repeated interactions, whereas those with initially inflated expectations show steeper declines in willingness to rely on AI systems.

Organizational Practices:

Capability transparency: Organizations should communicate clearly and specifically about what AI systems can reliably accomplish, what tasks involve higher error rates, and what completely exceeds current capabilities. This includes acknowledging uncertainty—being explicit about categories of tasks where AI performance is inconsistent or difficult to predict.
Limitation disclosure: Training and documentation should prominently feature common failure modes and error types rather than focusing exclusively on successful applications. Users who understand typical limitations develop better strategies for verification and appropriate skepticism.
Accuracy metrics for specific applications: Where feasible, organizations should provide concrete data about AI accuracy for different task types. For example, rather than claiming an AI system "helps with writing," organizations might specify that it reduces grammar errors by 87%, improves clarity ratings by 12%, but requires human review for factual accuracy and appropriate tone.
Explanation of confidence indicators: When AI systems provide confidence estimates or uncertainty indicators, organizations should explain what these actually mean and how to interpret them. Research shows that users often misinterpret confidence scores, treating them as guarantees rather than probabilistic estimates.

Developing Human-AI Collaboration Competencies

Long-term organizational capability depends less on accessing powerful AI systems and more on developing workforce competencies in effective human-AI collaboration. This represents a new category of professional skill that organizations must deliberately cultivate.

Evidence Base: Research examining writing processes with AI assistance identified distinct patterns of engagement ranging from passive reproduction of AI-generated text to active integration involving substantial human modification and critical evaluation. Participants showing more active, integrative patterns produced higher-quality outputs and reported greater learning from the process compared to those who primarily reproduced AI content with minimal modification.

These findings suggest that organizations should focus not just on AI adoption rates but on the quality of human-AI collaboration, emphasizing patterns that leverage AI capabilities while maintaining human judgment, creativity, and accountability.

Organizational Practices:

Collaborative workflow design: Rather than positioning AI as independently completing tasks or humans as passively accepting AI outputs, organizations should design workflows that specify distinct contributions from human and AI participants. For example, AI might generate multiple alternative approaches, with humans selecting among them based on contextual factors AI cannot access; or humans might create initial frameworks that AI develops into detailed drafts for subsequent human refinement.
Prompt literacy development: Effective AI use increasingly depends on prompt engineering—crafting inputs that elicit desired outputs. Organizations should help employees develop systematic approaches to prompting, including iterative refinement, providing relevant context, and specifying desired output characteristics.
Critical evaluation skills: Organizations should cultivate specific competencies in evaluating AI outputs, including detecting plausible-sounding but inaccurate information (hallucinations), identifying inappropriate tone or framing, and recognizing outputs that technically satisfy prompts while missing broader objectives.
Meta-cognitive awareness: Employees should develop reflexive awareness of their own AI usage patterns, including recognizing when they are over-relying on AI, when their critical engagement is waning, and when they might benefit from reduced AI involvement to maintain skill development.

Establishing Continuous Learning Systems

AI capabilities evolve rapidly, rendering one-time training insufficient. Organizations must establish continuous learning systems that help employees keep pace with capability changes while sharing effective practices across the organization.

Evidence Base: The present longitudinal research found that the effects of user-related characteristics (experience, use frequency, attitudes) and AI-related evaluations (trust, anthropomorphism) remained relatively stable across six measurement waves spanning ten months. This stability suggests that initial training creates lasting impact but also implies that updating skills and perceptions requires ongoing intervention rather than single educational sessions.

Research on technology adoption consistently demonstrates that peer learning and knowledge sharing prove more effective than formal training alone in helping workers develop sophisticated, contextualized competencies. This appears particularly important for AI systems where appropriate use depends heavily on tacit knowledge and situated judgment that resist codification in formal documentation.

Organizational Practices:

Communities of practice: Organizations should establish forums—whether physical spaces or digital platforms—where employees share AI usage experiences, discuss challenges, and collaboratively develop solutions. These communities function most effectively when they include diverse participants across hierarchical levels and functional areas.
Regular capability updates: As AI systems are updated with new features or improved performance, organizations should communicate changes through multiple channels and provide opportunities for hands-on experimentation. This might include brief demonstration sessions, updated documentation highlighting key changes, or sandbox environments where employees can explore new capabilities without production consequences.
Case repositories and pattern libraries: Organizations should systematically capture and share effective practices, developing libraries of successful prompts, use cases, and integration approaches that employees can adapt to their own contexts. This organizational memory accelerates learning and reduces duplicated effort across teams.
Bidirectional feedback mechanisms: Organizations should establish clear pathways for employees to report both successful applications and limitations or failures, with this information informing both technology selection and usage guidance. When employees see their feedback leading to tangible changes, they become more engaged in the continuous improvement process.

Balancing Efficiency and Capability Development

Perhaps the most challenging long-term issue concerns balancing immediate productivity gains from AI delegation against potential erosion of human capabilities essential for handling novel situations, exercising professional judgment, and maintaining organizational resilience.

Evidence Base: Educational research provides concerning evidence that extensive AI use for academic assignments is associated with reduced development of critical thinking, argument construction, and analytical skills. While workplace contexts differ in important respects, similar dynamics may operate when employees consistently delegate complex cognitive tasks that would otherwise build expertise.

This concern extends beyond individual skill development to organizational capability. When AI systems handle increasing portions of knowledge work, organizations may lose tacit knowledge, erosion of judgment capabilities, and reduced capacity to perform work without AI assistance. These risks become particularly salient during system failures, when legacy AI systems become outdated, or when novel situations exceed AI capabilities.

Organizational Practices:

Intentional task allocation: Organizations should explicitly consider developmental impact when deciding what to delegate to AI versus retaining as human-led. This includes identifying "core capabilities" essential for organizational function and ensuring that employees maintain and develop proficiency in these areas even when AI could handle portions of the work.
Rotational exposure to human-led work: Organizations might implement rotations where employees periodically perform work without AI assistance, maintaining skills that could atrophy through consistent delegation. This approach, analogous to pilots maintaining manual flying skills despite sophisticated autopilot systems, preserves capabilities needed for non-routine situations.
Differentiated delegation by experience level: Organizations might appropriately restrict AI delegation for less experienced employees still developing foundational skills, while allowing more extensive delegation by senior professionals with established expertise. This acknowledges that delegation impacts differ depending on whether individuals have already developed capabilities versus are still acquiring them.
Monitoring for capability erosion: Organizations should establish metrics and assessment approaches that detect potential skill degradation associated with AI delegation. This might include periodic competency assessments, analysis of performance trends during system outages, or structured evaluation of work produced without AI assistance.

Conclusion

The integration of generative AI into organizational workflows represents a profound transformation in knowledge work, with implications extending far beyond immediate productivity gains. The evidence examined in this article reveals a more complex picture than simple narratives of increasing adoption would suggest. Rather than growing steadily more comfortable with AI delegation and disclosure across repeated interactions, individuals appear to engage in a calibration process—adjusting their reliance to better correspond with actual system capabilities and limitations.

This calibration manifests most clearly in the declining willingness to delegate personal writing tasks, where authenticity and self-expression carry inherent value beyond functional outcomes. Professional writing contexts show more stable delegation patterns, likely reflecting cost-benefit calculations that favor efficiency despite AI limitations. The distinction between these contexts underscores a fundamental principle: appropriate AI reliance depends on task characteristics, not just system capabilities or user comfort.

For organizational leaders navigating AI adoption, several imperatives emerge from the evidence. First, sustainable implementation depends less on maximizing initial enthusiasm and more on cultivating well-calibrated trust through transparency about both capabilities and limitations. Organizations that proactively address system constraints—rather than discovering them through disappointing user experiences—foster more stable, appropriate long-term adoption patterns.

Second, effective AI integration requires substantial investment in capability development beyond basic technical training. Employees need support developing judgment about when to delegate, skills in evaluating AI outputs critically, and meta-cognitive awareness of their own usage patterns. These competencies prove more consequential than prompt engineering skills or technical understanding of how AI systems function.

Third, organizational governance must balance enabling productivity gains against protecting against over-reliance, privacy risks, and capability erosion. This requires differentiated policies that reflect task characteristics rather than one-size-fits-all approaches, combined with ongoing monitoring of actual usage patterns and their downstream consequences.

The broader trajectory of human-AI collaboration remains uncertain. The present evidence suggests we are in early stages of a longer adjustment process as individuals and organizations learn where AI can genuinely add value versus where it proves disappointing or counterproductive. Organizations that treat AI adoption as an ongoing organizational learning challenge—rather than a one-time technology deployment—position themselves to navigate this uncertainty most effectively.

Ultimately, the goal is not maximum AI adoption but appropriate AI integration that genuinely enhances human capability while preserving essential judgment, creativity, and accountability. Achieving this balance requires moving beyond simplistic metrics of usage rates toward more nuanced assessment of collaboration quality, task appropriateness, and long-term capability development. Organizations that embrace this complexity, investing in the cultural, procedural, and educational infrastructure that effective human-AI collaboration requires, will be best positioned to realize AI's potential while avoiding its pitfalls.

Research Infographic

Mastering AI Calibration Slide Deck

References

Anderl, C., Klein, S. H., Sarigül, B., Schneider, F. M., Han, J., Fiedler, P. L., & Utz, S. (2024). Conversational presentation mode increases credibility judgements during information search with ChatGPT. Scientific Reports, 14, 17127.
Baird, A., & Maruping, L. M. (2021). The next generation of research on IS use: A theoretical framework of delegation to and from agentic IS artifacts. MIS Quarterly, 45(1), 315-341.
Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion. Journal of Marketing Research, 56(6), 809-825.
Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). How people use ChatGPT (Working Paper 34255). National Bureau of Economic Research.
Cress, U., & Kimmerle, J. (2023). Co-constructing knowledge with generative AI tools: Reflections from a CSCL perspective. International Journal of Computer-Supported Collaborative Learning, 18, 607-614.
Essien, A., Bukoye, O. T., O'Dea, X., & Kremantzis, M. (2024). The influence of AI text generators on critical thinking skills in UK business schools. Studies in Higher Education, 49(5), 865-882.
Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627-660.
Kaplan, A. D., Kessler, T. T., Brill, J. C., & Hancock, P. A. (2023). Trust in artificial intelligence: Meta-analytic findings. Human Factors, 65(2), 337-359.
Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50-80.
Lubars, B., & Tan, C. (2019). Ask not what AI can do, but what AI should do: Towards a framework of task delegability. Advances in Neural Information Processing Systems, 32.
Luther, T., Kimmerle, J., & Cress, U. (2024). Teaming up with an AI: Exploring human-AI collaboration in a writing scenario with ChatGPT. AI, 5(3), 1357-1376.
Luther, T., Belcher, K., Kimmerle, J., Meurers, D., & Cress, U. (2026). Collaborative knowledge construction with generative AI: Exploring argumentative co-writing processes through n-gram and cluster analysis. International Journal of Computer-Supported Collaborative Learning.
Maslej, N., Fattorini, L., Perrault, R., Gil, Y., Parli, V., Kariuki, N., Capstick, E., Reuel, A., Brynjolfsson, E., Etchemendy, J., et al. (2025). Artificial Intelligence Index Report 2025. Stanford University Human-Centered AI Institute.
Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81-103.
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192.
Pew Research Center. (2025). Workers' experiences with AI chatbots in their jobs.
Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293-2303.
Venkatesh, V., Morris, M. G., Davis, G. B., & Davis, F. D. (2003). User acceptance of information technology: Toward a unified view. MIS Quarterly, 27(3), 425-478.

Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Co-Founder & Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Professor of Organizational Leadership & Change (UVU). Read Jonathan Westover's executive profile here.

Suggested Citation: Westover, J. H. (2026). Navigating the Evolving Landscape of Human Reliance on Generative AI. Human Capital Leadership Review, 39(2). doi.org/10.70175/hclreview.2020.39.2.2