The more you use AI, the more user-friendly it feels, but people's critical thinking ability is degrading.
The more you use AI, the weaker your ability to think independently may become. Research shows that over-reliance on AI can disconnect professional knowledge from real-world contexts and make people's ideas increasingly homogeneous. This article puts forward three solutions: let AI challenge your assumptions in reverse, deliberately create time and space without AI, and try interaction methods beyond the chat box. Remember: AI should be a tool to help you think, not a master that thinks for you.
Most companies are well aware that employees cannot only passively execute instructions and mechanically apply rules. People with critical thinking are good at raising questions and daring to challenge inherent assumptions — such behaviors are crucial for organizations to maintain agility and stimulate innovation. However, growing evidence shows that the application of AI may weaken this human ability.
To help companies address this issue, we integrate existing research in management (including some of our own results), cognitive science, human-computer interaction, and literature related to work and technology in organizations. At the same time, we interviewed a group of corporate executives who are experimenting with various AI implementation methods, and proposed a brand-new set of ideas to guide companies in designing, deploying, and using AI, with a view to promoting the collision of diverse perspectives and cultivating critical thinking in the long run.
Companies should focus on guarding against two major organizational risks caused by the weakening of critical thinking: the loss of professional competence and the loss of cognitive diversity. Both of these risks can undermine the foundation of an enterprise's agility and innovation. Managers must recognize clearly: Technology is not value-neutral, and it can reshape people's thinking patterns and collaboration methods in negative ways. Therefore, leaders should not only focus on short-term benefits (such as expected efficiency improvements and cost reductions) but also guide users not to take technical outputs for granted, take the initiative to explore and practice — to actively shape artificial intelligence rather than be shaped by it.
How AI Reshapes Human Thinking
Long before the generative AI boom, multiple studies had revealed the phenomenon of the "black box effect": practitioners blindly trust analytical tools without understanding them, which over time leads to the loss of professional judgment. Recent research further shows that using AI reduces people's willingness to actively verify information, raising academic concerns about cognitive outsourcing and even cognitive compliance. A 2025 study on the cognitive cost of large language models by the MIT Media Lab and a 2026 study by the Wharton School both confirmed that people who use AI tend to no longer question or verify AI outputs. Although the long-term impact of AI on human cognition remains to be continuously demonstrated, existing research has sent a clear signal: the use of AI is effectively changing organizational capabilities.
First, many researchers worry that the popularization of AI will cause the professional knowledge within enterprises to be decontextualized. The core logic is: AI can easily retrieve facts and data, but it cannot make meaningful interpretations and applications of knowledge in combination with specific contexts. Data often contains a large amount of subtle information that can only be captured by people with intuition, empathy, and cross-domain practical experience. The reason why humans can understand data is that they have personal experiences and an understanding of the real scenarios where the data is located. AI is just a set of algorithms and cannot directly experience reality, so it is easy to ignore these subtle and critical pieces of information.
For example: Even if AI can pass the bar exam, it does not mean that it understands all the nuances in the judicial system, the demands of various stakeholders, the differentiated causes and logic of different cases, and how other professional fields are connected to the legal system. Once legal professional knowledge is separated from specific contexts, risks will follow: if lawyers simply rely on AI, are not familiar with the details of the case, and do not understand the complex judicial environment, they are likely to misinterpret evidence and make judgment errors.
In fact, a recent study shows that precisely because people are increasingly relying on AI, enterprises actually need more talents with mastery of professional domain knowledge to be responsible for verifying AI outputs, flexibly responding to new situations, and predicting potential risks. Another hidden danger is that enterprises cut entry-level positions to reduce costs, which seems to save expenses in the short term, but in the future, there may be a talent gap among middle managers — and enterprises are currently highly dependent on the contextual understanding capabilities of this group of managers.
In addition, AI tends to foster homogeneous thinking. Multiple studies have shown that large language models often compress the diversity of perspectives: on the one hand, models tend to stick to the first idea that comes up, and on the other hand, they have an "averaging" tendency. This will induce users to ignore anomalous and unexpected information, yet innovative insights often emerge precisely from these unexpected pieces of information. It can be seen that the proliferation of generative AI easily forms a single thinking circle, depriving organizations of the diverse experiences and perspectives that are indispensable for innovation. At the same time, employees are more likely to ignore external change signals, weakening the company's ability to adapt to the environment.
Then, how should we manage these risks? The mature literature on work and technology in the organizational field has already provided feasible directions. The core viewpoint is: Technology always requires human interpretation, it is rooted in specific contexts, and users have the initiative to transform and adjust technology.
How Humans Can Actively Shape AI
The current mainstream narrative around generative AI mostly presents it as a single tool with a chat interface, where people ask it questions and seek answers. The mainstream view is that to get better results, you need to learn to optimize prompts and use AI more frequently. There is nothing wrong with learning to ask better questions in itself, but this cannot fundamentally resolve the risks. To truly meet the challenges, we must not only treat technology as a tool, but regard it as a system that humans can shape. This means we need to design an AI-empowered work environment that safeguards human autonomous judgment and dominant position.
Combining the practical experience of various enterprises we have investigated and cooperated with in AI and new technologies, we have summarized three feasible paths.
Use AI in Reverse
The ancient Greek philosopher Socrates was accustomed to asking questions to his followers, guiding them to independently explore the answers to philosophical questions, and sometimes prompting people to realize their own cognitive limitations. This line of thinking can also be applied to AI practice. The following two cases can illustrate this:
"Co-create prompts with me" Challenge. Launched by the large pharmaceutical company AstraZeneca in August 2025, this 10-day project consists of a series of short exercises aimed at encouraging employees to experiment with AI and exercise their critical thinking skills. More than 1,500 employees across the company participated. What participants learned is not only writing prompts, but more importantly, mastering a way of thinking: using prompts to force themselves to conduct in-depth learning and dig deeper.
Feedback from participants shows that this activity has reshaped everyone's way of getting along with AI. One participant pointed out: "This method opens up a whole new perspective of thinking, challenges my inherent cognition, and also makes me learn to reflect on myself before starting any conversation."
Dr. Bonnie Check, AstraZeneca's Global Head of Capability Development, and Dr. Maciej Szymaszek, Head of Enterprise AI Strategy and Innovation, told us in an interview: "This challenge spread rapidly across the entire organization, with various teams and departments participating one after another, making localized adjustments in combination with their own workflows, functions, and AI tools." The first pilot in 2025 won the Brandon Hall Silver Award. A year after the launch of the activity, more than 50 teams and business departments have implemented the project, with thousands of employees participating in the challenge.
Hackathon. A few months ago, research teams from ETH Zurich and Nanyang Technological University of Singapore conducted AI experiments with nearly 2,000 participants (forming nearly 500 teams) in a 24-hour hackathon hosted by the Seram Institution. Each team submitted project proposals that could help the United Nations achieve the Sustainable Development Goals, and AI provided different types of feedback to the teams. Half of the teams received feedback focusing on how to optimize their proposals to better meet the intended goals; the other half received feedback emphasizing the overlaps between their proposals and existing projects, encouraging them to try completely new ideas. The research team found that when AI continuously questioned the second group of participants, helping them distinguish which parts of their proposals were novel and which already existed, these teams produced more diverse and more innovative ideas.
Identify the Right Timing and Scenarios for Using AI
When building an AI governance framework, companies usually focus on where to implement AI. But at the same time, enterprises should think more about: Which scenarios and time periods should not use AI. We put forward the following suggestions for managers:
Create a free-thinking space without AI. This strategy draws on past corporate practices to deal with email information overload. For example, in 2007, Intel launched a pilot program for a group of engineers to alleviate the information overload problem faced by employees: setting aside one day a week to encourage engineers not to use email. (Previously, companies such as Deloitte and T-Mobile USA had also tried to implement "No Email Day/Week".)
Similarly, enterprises and managers can designate specific time periods to require employees not to use generative AI tools, or provide them with "access-based interactive interfaces" — the system will only open the AI-assisted function after users first input their own context-integrated thinking. Another group of researchers reported that an Australian telecom operator they studied required middle managers to hold "AI-free strategic meetings" before using AI tools, forcing everyone to first rely on their own judgment and experience to brainstorm strategic plans.
Uphold Human Dominance. Technology is ubiquitous in current development and growth strategies, and sometimes it can so greatly influence corporate decisions and structures that its use becomes unquestionable. To ensure that humans can shape technology, the first step is to provide employees — especially new entrants to the industry — with opportunities to handle specific tasks without AI tools, so as to accumulate professional competence and contextual knowledge.
Understanding processes is not only a required course for front-line employees, but also equally important for middle and senior managers. Anthony also observed a clear divide when studying senior banking practitioners: some executives only regard algorithmic tools as carriers for outputting results, and their own professional competence continues to weaken; other senior practitioners actively understand the principles of various technologies and clarify how technology affects their work. The latter not only continuously improves their professional competence, but also can effectively guide junior analysts — requiring them to explain the basis for tool selection, and clarify the hypothetical logic and calculation process behind the data.
From a more macro organizational perspective, research on technology and work has shown that the effective implementation of new technologies cannot be separated from redesigning workflows on the basis of employees' existing knowledge. In fact, an expert report funded by the Spanish Ministry of Labor embodies this idea, emphasizing the importance of involving front-line employees in discussions when promoting AI implementation.
Parallel Work Mode. In some cases, it may be beneficial to have AI and a human team independently perform the same task in parallel. After that, the team and AI can integrate their outputs to form a hybrid solution. With Company, a transformation design consulting firm headquartered in Lisbon, is continuously experimenting with this parallel model in strategic and creative projects.
In an early experiment with a startup in 2024, two teams completed the same brand planning task simultaneously. One team adopted a more traditional design process, while the other relied heavily on AI-supported workflows. The AI-supported team generated nearly 1,700 images before arriving at the final solution, greatly expanding the exploration space and accelerating the divergent process. The human-led team produced fewer solutions, but they had greater coherence, clearer storytelling, and a more complete contextual framework from the very beginning.
In the end, the project adopted the solution of the human team. But With Company summed up general experience from it: the parallel model clearly demonstrates the relative strengths of AI and human teams in different stages of the creative process. AI excels at information transcription, synthesis, rapid exploration, and expanding possibilities, while framework construction, priority judgment, and value trade-offs still highly rely on human interpretation and contextual understanding. Since then, With Company has continued to apply the parallel work mode in multiple projects.
Try Different Interactive Interfaces
The vast majority of generative AI and large language models adopt a chatbot architecture, providing a "question-and-answer" style interaction in most scenarios. AI's response is usually a large piece of text, which requires users to sort out the key points one by one and filter out the information they need. However, this interface form is not suitable for all scenarios. What often happens in reality is that the chat interface degenerates into a chaotic "conversation": office workers are overwhelmed by the massive summary text output by AI, so they simplify, restructure, and reinterpret the content; AI then continues to answer by imitating the adjusted expression of humans, further exacerbating the chaos.
Human-computer interaction research has long broken the inherent cognition that "human-computer interaction can only rely on a keyboard and mouse". Similarly, we should also break away from the mindset that "AI can only be a chat window". A recent academic paper proposes that we should abandon the "frictionless user interface" and instead create a new type of interface: directly displaying raw data while presenting multiple opposing arguments, forcing users to actively sort out information, compare different alternative solutions, and carefully weigh them.
Empirical research shows that such complex interactive interfaces perform outstandingly in fields where users are already experts, such as mathematics and programming. For example, Google has pointed out that the existing chat interface cannot support the complex, multi-dimensional, and iterative work modes required for mathematical research. To this end, Google developed the "AI Mathematician's Copilot" tool, equipped with a flexible interactive interface, which can not only assist mathematicians in advancing their research work, but also help verify proof processes and identify vulnerabilities. This tool visualizes the complete reasoning steps through an annotation system, making the derivation process clearly visible, thus better supporting iterative research in mathematics. Such tools show that AI does not have to be limited to the general chatbot form that we are so familiar with today.
As AI continues to penetrate into various organizations, managers should set aside industry hype and plan the enterprise's AI applications with a comprehensive and balanced perspective. To help employees become active knowledge creators, enterprises must strategically carry out AI pilot explorations, clarify how to use technology to empower work, uphold the foundation of professional competence, and cultivate diverse thinking. AI should empower people, not replace them.
This article is from the WeChat public account "Harvard Business Review" (ID: hbrchinese), author: HBR-China, published with authorization from 36Kr.