Ironies of Generative AI: Understanding and mitigating productivity loss in human-AI interactions

1 INTRODUCTION

• GenAI users report increased cognitive load, frustration, and time spent on tasks the tools are intended to support, while productivity gains may not be guaranteed or evenly distributed.Examples include difficulty prompting and debugging generated code, interruptions, and reviewing or editing suggestions that are ultimately deleted.Source: §1 · p. 2
• The paper synthesizes over 30 years of Human Factors research and recent GenAI studies to show that current usability challenges parallel those seen when automation entered workflows.Its focus is primarily programming, with emerging studies from healthcare, writing, and design also considered.Source: §1 · pp. 2–3
• The paper groups productivity-loss challenges into the production-to-evaluation shift, unhelpful workflow restructuring, task interruptions, and task-complexity polarization.It also identifies situational awareness, cognitive workload, and workflow disruption as areas for further study.Source: §1 · p. 3
• The proposed design directions include continuous feedback, personalization, ecological interface design, task stabilization and timing, and clear task allocation.The paper emphasizes feedback and flexibility as Human Factors principles for GenAI design.Source: §1 · p. 3

2 PRODUCTIVITY CHALLENGES OF GENERATIVE AI AUTOMATION

• The discussion focuses on GenAI systems as integrated models and interfaces, with prompting challenges pertaining to models and interruption challenges pertaining to interface design.Source: §2 · p. 3
• The section identifies four productivity challenges: production-to-evaluation shifts, unhelpful workflow restructuring, task interruptions, and task-complexity polarization.The last category describes automation making easy tasks easier while making hard tasks harder.Source: §2 · p. 3
• The paper connects the four challenges to Human Factors research on automation and examines their emergence in GenAI user studies.Source: §2 · p. 3

2.1 The production-to-evaluation shift

• Copilot users sometimes failed to complete coding tasks more often than autocomplete users and, when completing them, were no faster.Reviewing and validating suggestions could displace writing code or running tests and create an efficiency bottleneck through debugging.Source: §2.1 · pp. 4–5

2.1.1 Reduced situational awareness.

• GenAI output evaluation can reduce situational awareness, making it harder to understand generated code, spot errors, and continue the main task.Users reported refactoring or abandoning code when they could not rely on intuition about bugs; novices may find this especially challenging.Source: §2.1.1 · p. 5

2.1.2 High automation capacity.

• GenAI’s high output capacity makes evaluation challenging, especially when suggestions are separated from the code context needed to compare them.Users may quickly assess many code suggestions through pattern matching, while a separate multi-suggestion pane can increase cognitive load.Source: §2.1.2 · p. 6

2.1.3 Automation complexity and opaqueness.

• GenAI complexity and opacity complicate both prompting and evaluation because users may not know what inputs the system relies on or how they shape its output.The opaque prompt-to-output relationship is termed “fuzzy abstraction matching” [109].Source: §2.1.3 · p. 6

2.1.4 Automation reliability.

• Non-deterministic outputs and subtle errors make GenAI less reliable from users’ perspective and help explain why workers review its outputs.Reported concerns include code errors, violations of brand standards and copyrights, and inaccuracies in legal documents.Source: §2.1.4 · p. 6

2.1.5 Potential complacency and over-reliance.

• Some GenAI users skim rather than rigorously evaluate outputs, and missed errors can require later debugging; over-reliance has also been associated with poorer consultant performance.Skimming and overconfidence are reported across programming, advertising, data science, law, and design.Source: §2.1.5 · p. 7

2.2 Unhelpful workflow restructuring

• Workflow restructuring can add tasks, disrupt familiar sequences, and remove feedback, leading users to change strategies and potentially use freed time and cognitive resources ineffectively.When automation adds work or disrupts steps, users may struggle to adapt, make errors, or avoid using the system during high-workload situations.Source: §2.2 · p. 7

2.2.1 Prompting as a new task.

• Prompting can become a cognitively demanding task that competes with users’ other workflow tasks.Users experiment with wording and scope, build a mental model of what the system can handle, and may rewrite comments or shift attention from their own narrative or argument [8] [29] [109] [129].Source: §2.2.1 · p. 8

2.2.2 Output adaptation as a new task.

• Adapting generated output adds complex work, and its productivity gains or losses remain uncertain.Programmers may edit, break down, or combine suggestions; debugging errors in adapted code can add workload, and some users abandon suggestions [8].Source: §2.2.2 · pp. 8–9

2.2.3 Loss of task sequence.

• GenAI suggestions and prompting can disrupt task sequence, with novices sometimes cycling through adapting and deleting suggestions or debugging incorrect code.In writing, users produced lower-quality results when GenAI generated most of the text rather than providing feedback.Source: §2.2.3 · p. 9

2.2.4 Loss of feedback.

• GenAI-assisted coding can omit the discussions, explanations, comparisons, and social commentary that programmers find on Q&A sites.Users may see code, comments, and data without the richer feedback available through multiple media sources.Source: §2.2.4 · p. 10

2.3 Task interruptions

• Interruptions disrupt thought processes, require time and cognitive resources for task switching, and can break users’ flow state.Long, complex interruptions can make it particularly difficult to resume the original task.Source: §2.3 · p. 10
• Copilot suggestions interrupt programmers’ thinking, leading some to disable suggestions or the tool, while similar interruptions are reported in writing and data science.Some programmers also report being tempted to follow suggestions instead of thinking through the task themselves.Source: §2.3 · p. 10
• Long, multi-line suggestions can break programmers’ flow during acceleration, shifting attention from problem-solving to reviewing code and debugging.This can be especially disruptive for novices, who may feel compelled to read suggestions despite finding them a nuisance.Source: §2.3 · p. 10
• Complex code suggestions are more distracting during acceleration and more helpful during exploration, making timing a key factor.Interruptions scheduled at task-chunk or subtask boundaries can be easier to respond to.Source: §2.3 · p. 10

2.4 Task-complexity polarization

• Task-complexity polarization describes automation making easy tasks easier while failing to reduce—and sometimes increasing—workload on demanding tasks.Difficult work may remain under manual control alongside monitoring, within a more fragmented workflow.Source: §2.4 · p. 11
• GenAI users report greater success or confidence on simple tasks, while current systems are less effective on harder tasks.Examples include boilerplate code and microtasks, while Copilot limitations on larger coding problems have led to more failures in medium and hard tasks.Source: §2.4 · p. 11
• On demanding tasks, shifted roles, disrupted workflows, and interruptions can deprive users of context or disrupt task sequence; novices may find suggestions confusing.The paper also notes that systems can introduce errors users must handle, including subtle bugs in longer code suggestions that complicate debugging.Source: §2.4 · pp. 11–12
• Multiple AI-generated code translations heightened users’ frustration and mental demand, and Copilot’s multi-suggestion pane could overwhelm users during coding exploration.The paper concludes that GenAI may leave users with the same or increased cognitive workload on hard tasks.Source: §2.4 · p. 12

3 HUMAN FACTORS SOLUTIONS

• The proposed approaches include continuous feedback, personalization, ecological interface design, task stabilization and interruption timing, and clear task allocation.The approaches share Human Factors principles of feedback and system flexibility.Source: §3 · p. 12
• The approaches aim to increase user agency, situational awareness, and flexibility in applying GenAI support to tasks.The paper illustrates these approaches primarily in programming.Source: §3 · p. 12

3.1 Continuous feedback

• GenAI can shift users from producing work to reviewing AI-generated outputs for errors, a cognitively demanding role with reduced contextual support and situational awareness.The paper proposes feedback about system behavior to keep users engaged and in the loop.Source: §3.1 · p. 12
• During monitoring, continuous feedback helps operators know whether requests were received, automation is acting properly, and errors are occurring.For GenAI, feedback can convey how inputs are interpreted and how outputs match them.Source: §3.1 · pp. 12–13
• GenAI feedback can help users understand input-output relationships and system behavior, supporting mental models, better prompting, output adaptation, and workflow structuring.The paper recommends feedback that is informative but non-intrusive and adapted to interaction style and problem type.Source: §3.1 · p. 13
• Feedback can overwhelm when poorly presented or excessive, so its design should support context, abstraction, integration, situational awareness, and error detection.The paper points to ecological interface and notification design as relevant approaches.Source: §3.1 · p. 13

3.2 System personalization

• Constrained personalization increases operators’ cognitive demands and associated productivity loss, while flexible controls can let users adapt systems to their tasks and ways of working.Users could choose when suggestions appear, indicate their work state, or adjust the context provided to the system.Source: §3.2 · pp. 13–14
• In creative writing, choosing when to receive GenAI feedback rather than receiving generated text preserved creativity and alleviated anchoring effects and over-reliance.Systems could tailor suggestions to a user’s stated goals for complexity, variety, length, and frequency.Source: §3.2 · p. 13
• Users may benefit from different kinds and amounts of information: novices may want help only when stuck, while experts may prefer line completion.Personalization matters because task and domain expertise affect user preferences and needs.Source: §3.2 · p. 14

3.3 Ecological interface design

• Workflow disruption can increase cognitive load, reduce productivity, and prevent users from exercising expertise or benefiting from AI support.Ecological interface design aligns interfaces with domain-specific perceptual constraints through real-time signals, consistent mappings, and visible system relationships.Source: §3.3 · p. 14
• Systems should fit users’ mental models and work-domain expectations, while clearly specifying which external sources inform their inputs and when.For Copilot, this could include code beyond the current file, with sources made clear to support real-time control.Source: §3.3 · p. 14
• GenAI should respect natural task sequences, avoid suggestions that interfere with planning, and show how code changes affect the wider codebase.For physicians’ administrative tasks, outputs should include patients’ medical histories and physicians’ clinical reasoning.Source: §3.3 · p. 14
• Hierarchical displays can present multiple information levels, letting users shift attention according to expertise and task demands.For programming, suggestions could range from high-level pseudocode to low-level implementations, organized visually.Source: §3.3 · p. 14
• Domain-specific explanations can help users form accurate mental models; explanations lacking relevant context confused expert dermatologists.The passage also reports similar results in radiology.Source: §3.3 · p. 15

3.4 Main task stabilization and interruption timing

• Systems should preserve users’ flow states through task stabilization or by timing interruptions around those states, rather than forcing users to disable distracting tools.Some programmers and writers avoid suggestions or tools because of interruptions.Source: §3.4 · p. 15

3.4.1 Task stabilization via attention guidance.

• Attention guidance can make interruptions easier to manage by directing users to relevant task elements and helping them parse long suggestions into logical units.Cues such as highlighting keywords may support pattern matching; without appropriate cues, programmers may enter debugging rabbitholes.Source: §3.4.1 · p. 15

3.4.2 Task stabilization via pre-interruption alerts.

• Pre-interruption alerts can help users set place-keepers and resume their main task faster; GenAI could pair alerts with context-relevant information.The cited study found faster resumption after alerts, and the passage suggests alerts before long code suggestions.Source: §3.4.2 · p. 15

3.4.3 Timing of interruptions.

• When users are focused, systems should limit interruptions to contextual alerts and align suggestions with users’ flow; outside flow, they can offer prompts, feedback, or guidance.Recognizing users’ state could support shorter suggestions during flow and more extensive support when users are not in flow.Source: §3.4.3 · p. 16

3.5 Clear task allocation

• Task-complexity polarization describes GenAI systems making easy tasks easier and hard tasks harder, which task allocation is proposed to address.Clear allocation can distribute workload according to human and system strengths and reduce the effort of discerning responsibilities moment by moment.Source: §3.5 · p. 16
• Users could proactively assign repetitive or familiar tasks to GenAI while retaining high-level planning or exploratory work for more complex tasks.Examples include allocating boilerplate code to the system, or asking for open-ended ideas rather than explicit suggestions in creative work.Source: §3.5 · pp. 16–17
• Effective task allocation depends on the system understanding work-domain context, which ecological interface design can support.The paper presents these Human Factors approaches as working in synergy for human–GenAI interaction and productivity.Source: §3.5 · p. 17

• Shifts from active production to passive evaluation, unhelpful workflow restructuring, interruptions, and task-complexity polarization can stifle human performance and effective GenAI implementation.Source: §4 · p. 17
• Design approaches include ecological interface design, continuous feedback, flexible task allocation, user-guided personalization, and attention guidance during interruptions.Source: §4 · p. 17
• The paper is an initial bridge; more nuanced Human Factors research may help address productivity challenges, while future HCI research may open new areas for Human Factors.Source: §4 · p. 17