Cognitive load in AI-assisted work

Cognitive load in AI-assisted work tracks the mental burden created by the task itself, the AI interface, and the coordination work required to keep human and model aligned. The useful distinction is between intrinsic load, which comes from genuine task difficulty, and extraneous load, which comes from unnecessary interaction friction: task switching, information sprawl, incoherent model initiative, missing context, or outputs that force the user to repair the conversation instead of progressing the work.

2026-lepine-precision-proactivity gives this topic its first direct work-domain evidence. In a study of 34 financial professionals completing a complex valuation task with GPT-4o, AI-generated content use was positively associated with output quality, but extraneous load had the largest negative association with quality, roughly three times the magnitude of intrinsic load. The important design claim is precise: AI can help with content while still making the work cognitively harder through badly timed or badly structured proactivity.

The paper also makes cognitive load observable in human-AI traces. Lepine et al. estimate load from transcript features grounded in task decomposition, semantic coherence, dependency structure, and knowledge-graph representations. That gives the wiki a measurement bridge between broad claims about AI-mediated effort and concrete interaction diagnostics: when the model initiates task switches, spreads information across too many directions, or truncates needed structure, load becomes a design signal rather than a vague complaint.

2026-zhang-fatigue-aware-learning-defer adds a sequential allocation mechanism. Its fatigue-aware learning-to-defer framework treats cumulative human workload as part of the system state, so the value of human involvement changes over time as prior deferrals accumulate. That makes cognitive load more than a subjective experience variable: in some human-AI systems, load becomes an input to task routing and cooperation-budget decisions.

2023-arcot-cognitive-load-optimization-contact-center-agents reports a field-oriented version of that allocation logic: wearable physiological signals and voice features were used to infer contact-center agents’ stress and cognitive-load states, then route calls and balance workloads dynamically. Its reported 500-agent randomized trial makes the mechanism more concrete than a simulated cooperation budget, but it also makes the measurement and governance problem unavoidable. Inference from biometric and voice data is not merely a better workload signal; it creates a worker-monitoring system whose consent, validity, access, purpose limitation, and consequences for control must be designed before routing can be treated as supportive rather than extractive.

2026-shen-ai-impacts-skill-formation adds a learning-effort distinction. Some AI interaction patterns looked cognitively light because users delegated code generation or debugging, but those patterns produced poor learning outcomes; other patterns required explanation-seeking, conceptual questioning, or comprehension checks and preserved more skill formation. For this page, the important distinction is productive cognitive effort versus extraneous burden: reducing all effort may improve immediate ease while removing the friction that builds competence.

2026-mansuroglu-technostress-employee-well-being gives cognitive load a broader occupational-health frame. In its systematic review, techno-overload appeared in 122 studies, techno-complexity in 105, and emerging constructs such as digital fatigue, information anxiety, communication platform strain, and ICT hassles were associated with sleep problems, concentration loss, impaired cognitive skills, and reduced well-being. That expands the page beyond local model-interaction load: AI systems can create extraneous burden through the whole digital work environment, including platform proliferation, always-on communication, unreliable tools, and boundary invasion.

2026-ruttenberg-neurodivergent-expertise-ai-work adds an individual-differences boundary condition. For autistic and neurodivergent knowledge workers, the same AI notification stream, context-switching demand, or passive offloading pattern may consume capacity sooner because sensory filtering, social inferencing, and executive-function demands are already part of the work ecology. The paper’s reflective-scaffolding proposal is a useful design distinction for this page: reduce extraneous load through pacing control and sensory-aware interfaces, but do not remove the productive effort needed for comprehension, synthesis, and transfer.

2026-ruttenberg-cognitive-debt-ai-research sharpens the productive-effort boundary. In its model, cognitive debt grows when AI use reduces active engagement through attentional erosion, effort displacement, and affective depletion; reflective scaffolding is protective only if it lowers unnecessary burden while keeping the user in contact with the hard parts of the work. That makes load reduction a construct-validity problem: a low-effort interface may either remove extraneous friction or remove the cognitive struggle that produces transfer and judgment.

Connections

• Relates to human-ai-collaboration because collaboration quality depends on whether model initiative reduces burden or creates coordination work the human must absorb.
• Relates to automation-complacency because low effort and high load are different failure modes: some AI systems suppress scrutiny, while others overwhelm the user with extraneous coordination demands.
• Relates to critical-thinking because verification and stewardship require enough cognitive capacity to evaluate AI output rather than merely cope with conversational disorder.
• Relates to ai-mediated-work-experience because cognitive load is one channel through which AI changes workload, competence, control, and felt job quality.
• Relates to work-redesign because task sequencing, escalation rules, and interface defaults can either suppress extraneous load or push it onto workers.
• Relates to human-ai-task-allocation because workload and fatigue can change which actor should handle the next task instance.
• Relates to digital-work-strain because cumulative digital demands can become cognitive fatigue, overload, recovery failure, and occupational-health risk.
• Relates to cognitive-debt because load design determines whether AI assistance preserves productive effort or silently displaces the cognitive work needed for later transfer and judgment.
• Relates to inclusive-hr-systems because cognitive-load assumptions built into AI tools can encode neurotypical work rhythms and make some workers’ sustainability costs invisible.

Contradictions & open questions

• 2026-lepine-precision-proactivity is observational and domain-specific. Its framework is promising because it uses real transcripts from professionals, but its load measures are computational proxies that still need validation against established cognitive-load assessments.
• AI-generated content can improve quality while extraneous load harms it. That means “AI makes the task easier” is too coarse: assistance can reduce one burden while creating another.
• The open design question is when proactivity should be suppressed, delayed, or transformed into repair. More initiative is not automatically better; proactivity needs to be state-dependent and load-sensitive.
• The framework depends on domain-specific task decompositions and knowledge graphs, so portability to software engineering, healthcare, HR, or strategy work remains an empirical problem rather than an assumption.
• 2026-zhang-fatigue-aware-learning-defer uses simulated fatigue curves and image-classification benchmarks, while 2023-arcot-cognitive-load-optimization-contact-center-agents reports live multimodal monitoring but does not provide independently verified validation of its stress and load measures. The open question is whether workplace fatigue or cognitive-load signals can be measured validly enough to drive allocation without creating surveillance, fairness, or agency problems.
• 2026-mansuroglu-technostress-employee-well-being synthesizes recurring overload and fatigue patterns but does not validate AI-interaction-level cognitive-load measures; the bridge from technostress surveys to trace-based AI workflow diagnostics remains open.
• 2026-ruttenberg-cognitive-debt-ai-research and 2026-ruttenberg-neurodivergent-expertise-ai-work propose cognitive-debt and neurotype-specific thresholds but do not yet validate them; the open measurement problem is how to detect harmful threshold crossing without turning accommodation-sensitive load signals into surveillance or deficit labeling.