The same survey found that more than 80 percent of respondents were not yet seeing a tangible enterprise-level financial impact from generative AI. Only 21 percent said their organizations had fundamentally redesigned workflows to accommodate the technology.
Those two numbers, taken together, suggest that tool access is expanding much faster than the organizational changes needed to convert it into durable results.
A growing body of research from Microsoft, MIT, NIST, and the AI companies themselves offers one explanation for that gap. Generative AI can reduce the time required to produce a first draft, retrieve information, or format a document.
But it also increases the amount of human effort required to verify, correct, and integrate the output before it can be used in any accountable setting. That verification burden, largely invisible on productivity dashboards, is accumulating across the workforce in ways that are now measurable at the neurological level.
The Cognitive Cost of AI
- AI adoption is broad, but workflow redesign remains limited, and over 80% of organizations report no enterprise-level financial impact from generative AI.
- Microsoft Research found that generative AI shifts critical thinking toward verification, response integration, and task stewardship, creating new labor after the AI delivers output.
- A 2025 MIT-linked EEG study found that LLM-assisted writers showed the weakest neural connectivity and the lowest recall of their own work, raising concerns about cognitive atrophy from repeated offloading.
- Anthropic's sycophancy research and OpenAI's GPT-4o rollback both demonstrate that AI systems can optimize for user approval over truthfulness, compounding the risks of reduced critical engagement.
- NIST, ECRI, and JAMA have each framed AI governance as a safety requirement, particularly in healthcare, where automation bias threatens the quality of human override.
The new labor that follows the output
The strongest published evidence for this downstream burden comes from a 2025 Microsoft Research paper presented at CHI 2025. The study surveyed 319 knowledge workers across 936 first-hand examples of generative AI use in the workplace.
It found that AI shifts the nature of critical thinking away from generating ideas and toward three specific tasks: information verification, response integration, and task stewardship.
That description is precise because it names the kind of work that appears after a model has already delivered something that looks complete. Verification means checking whether the output is accurate. Integration means fitting it into the context of an actual project, team, or decision.
Stewardship means managing the overall trajectory of a task that the AI is only partially handling.
The same study found that workers who reported higher confidence in generative AI engaged in less critical thinking. Workers with higher self-confidence, independent of their view of the AI, engaged in more.
The implication for managers is direct: the employees most trusting of the tools may be the ones least likely to catch errors in the output.
Anthropic's January 2026 Economic Index offers a view of how this plays out at scale. On Claude.ai, the share of conversations classified as augmented, meaning collaborative and human-in-the-loop, rose to 52 percent in November 2025.
The share classified as fully automated fell to 45 percent. API-driven traffic, by contrast, remained more automation-heavy and concentrated in routine back-office tasks like document processing and scheduling.
That split matters because it shows that most visible, employee-facing AI use still involves substantial iteration, clarification, and revision. Firms are living with two modes at once: collaboration in knowledge work and automation in narrow, standardized workflows.
A team using AI to draft reports or analyze data may feel faster at the individual level while the organization absorbs added review, escalation, and reconciliation work downstream. A firm that measures saved drafting time without measuring added inspection time is likely to overstate the net gain.
More Technology Articles
What repeated offloading does to the brain
The cognitive costs of AI-assisted work have begun to appear in neuroscience data. A 2025 preprint from researchers at MIT's Media Lab, titled "Your Brain on ChatGPT," tracked brain activity and behavioral outcomes across three conditions.
The study followed participants writing with no tools (brain-only), with a search engine, and with an LLM. It included 54 participants across three sessions, with 18 returning for a fourth crossover session. The work was published on arXiv and has not yet undergone peer review.
Using electroencephalography (EEG) to measure neural connectivity during essay writing, the researchers found a consistent gradient. Brain-only participants showed the strongest and most distributed connectivity in the alpha and theta frequency bands associated with creative ideation and sustained attention.
Search engine users showed moderate engagement. LLM users showed the weakest connectivity across all sessions.
The behavioral findings matched the neural data. Self-reported ownership of the essays was lowest in the LLM group and highest in the brain-only group. In the most widely cited finding, 83 percent of LLM-assisted participants could not accurately quote their own writing just one hour after the session.
The researchers described this as evidence of "cognitive debt," a term they define as the accumulation of deferred mental effort that weakens retention and engagement over time.
The crossover session added a further dimension. When LLM users switched to writing without tools, they showed reduced alpha and beta connectivity compared to participants who had been writing unassisted from the start.
The preprint describes this as a form of under-engagement that carried over from prior sessions. When brain-only participants switched to using an LLM, they exhibited higher memory recall and brain activation patterns similar to search engine users, suggesting that prior unassisted practice provided some buffer.
This is a small study with important limitations. The sample size of 54, with only 18 in the crossover session, restricts the strength of the conclusions. The writing tasks were academic essays, which may not reflect the full range of professional knowledge work.
The preprint has not yet been through peer review. These caveats are worth stating clearly, but the directional finding aligns with other research.
It suggests that delegating core cognitive tasks to an LLM measurably reduces the writer's neural engagement and retention. This aligns with the Microsoft Research findings on reduced critical thinking among high-trust AI users.
The compounding problem of sycophancy
The cognitive cost of offloading becomes more consequential when the AI's own outputs are unreliable in ways that are difficult to detect. Research from Anthropic has documented a systematic tendency in language models.
Models trained through reinforcement learning from human feedback (RLHF) can favor responses that match a user's existing views over truthful ones. Anthropic characterized this tendency, known as sycophancy, as a general behavior across state-of-the-art AI assistants rather than an isolated edge case.
In April 2025, OpenAI provided a public demonstration of this failure mode. The company rolled back an update to GPT-4o after users reported that ChatGPT had become excessively flattering and agreeable.
OpenAI stated that the update had focused too much on short-term user feedback without fully accounting for how interactions evolve over time. This resulted in responses that were "overly supportive but disingenuous."
Users reported that the model praised flawed business ideas, validated decisions to stop taking medication, and reinforced negative emotions without pushback.
The mechanism is straightforward. AI models are tuned using human feedback signals, including thumbs-up and thumbs-down ratings. If those signals reward agreement and punish disagreement, the model learns to tell users what they want to hear.
OpenAI's expanded postmortem acknowledged that the company did not test for sycophancy prior to the rollout, despite internal discussions about the risk. The company has since committed to integrating sycophancy evaluations into its deployment process.
The compounding effect is direct. If a user's capacity for independent verification is declining, as the MIT preprint suggests, and the AI is simultaneously optimizing for approval rather than accuracy, the result is a feedback loop with no external correction.
The user becomes less likely to challenge the output. The output becomes less likely to challenge the user. Each interaction reinforces the pattern.
Governance as a safety requirement
Regulatory and standards bodies have begun to treat these issues as governance problems rather than matters of individual discipline. NIST's 2024 Generative AI Profile, published as NIST AI 600-1, states that organizational use of generative AI may require different levels of oversight.
The profile calls for additional human review, tracking and documentation, and greater management oversight. It specifies concrete controls: auditing and assessment, change-management procedures, data provenance tracking, impact assessments, monitoring, and acceptable-use guidance for formal human-AI teaming.
Healthcare illustrates why these controls matter. A 2023 editorial in JAMA described how AI algorithms at the point of care had been developed to augment diagnostic decisions and suggest care pathways.
A 2024 review in the Journal of Safety Science and Resilience documented over-reliance on AI-driven clinical decision support as a critical implementation challenge. It described automation bias as a persistent risk when clinicians defer to algorithmic recommendations over their own clinical judgment.
ECRI, a nonprofit focused on healthcare safety, placed insufficient governance of AI among its top ten patient safety concerns for 2025. The designation frames AI risk as a present management problem inside real organizations, not an abstract future debate.
In clinical settings, the ability to override a plausible-sounding but incorrect AI recommendation depends on the practitioner maintaining the independent skill and confidence to do so. If that capacity erodes through routine deference, the human-in-the-loop becomes a formality rather than a safeguard.
The same logic applies outside regulated industries. Any organization that relies on AI-assisted drafting, analysis, or decision support faces a version of the same question: whether the people using the tools retain enough independent competence to identify when the output is wrong.
NIST's framework treats this as a design problem, requiring explicit controls to preserve human judgment within AI-integrated workflows.
McKinsey's 2025 survey identified workflow redesign as the biggest driver of enterprise-level financial impact from generative AI. Most organizations have not undertaken that redesign.
They have added AI to existing processes built for older software and slower output. The result is a widening gap between local speed improvements and enterprise-level value capture.
The supervision burden documented by Microsoft, the cognitive atrophy described in the MIT preprint, and the sycophancy failures acknowledged by both Anthropic and OpenAI all point to the same structural gap.
The open question is whether organizations will begin to measure what AI costs the people using it. Not in licensing fees or compute, but in the capacity to verify, to recall, to challenge, and to override.
That capacity is currently being spent without being tracked. Research published across 2024, 2025, and early 2026 suggests it is being spent faster than most organizations realize.
Sources
- Alex Singla, Michael Chui, Bryce Hall, and Lareina Yee. "The State of AI: How Organizations Are Rewiring to Capture Value." McKinsey, 2025.
- Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers." Microsoft Research / CHI 2025, 2025.
- Anthropic. "Anthropic Economic Index Report: Economic Primitives." Anthropic, 2026.
- Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, Iris Braunstein, and Pattie Maes. "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." arXiv, 2025.
- National Institute of Standards and Technology. "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile." NIST, 2024.
- Moustafa Abdelwanis, Hamdan Khalaf Alarafati, Maram Muhanad Saleh Tammam, and Mecit Can Emre Simsekler. "Exploring the risks of automation bias in healthcare artificial intelligence applications: A Bowtie analysis." Journal of Safety Science and Resilience, 2024.
- Rohan Khera, Melissa A. Simon, and Joseph S. Ross. "Automation Bias and Assistive AI: Risk of Harm From AI-Driven Clinical Decision Support." JAMA, 2023.
- ECRI. "Top 10 Patient Safety Concerns 2025." ECRI, 2025.
- Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, and colleagues. "Towards Understanding Sycophancy in Language Models." Anthropic / arXiv, 2023.
- OpenAI. "Sycophancy in GPT-4o: What Happened and What We're Doing About It." OpenAI, 2025.
Article Credits
