The concept of “cognitive offloading” – when humans outsource challenging or effortful mental tasks to external tools like AI chatbots – has increasingly been used to describe how AI might be making us, well, stupid. When people develop such a habit, especially with more complex tasks, it can undermine critical thinking and contextual expertise, compromise decision-making capabilities and lead to more homogenous thinking. 

What does this mean for organisations? As we wrote in a recent article in Harvard Business Review, most of them recognise the value of employees who ask questions and challenge assumptions – behaviours that are crucial for organisational agility and innovation. The irony is that despite growing evidence that using AI can erode this capacity, many companies continue to uncritically infuse AI into how work gets done. 

To help address this problem, we build upon existing research (including some of our own), organisational studies and interviews with executives to offer a new approach to designing, implementing and applying AI in a way that encourages diverse and critical thinking in the long term. 

The risks when AI influences how we think 

Many researchers worry that AI usage will lead to a decontextualisation of expertise. Although AI can readily access facts and data, it's ultimately an algorithm – meaning that it doesn’t experience the data directly and can’t contextually interpret and apply knowledge meaningfully. This is because it overlooks nuances that can only be recognised by people with intuition, emotion and lived experience. 

For instance, the fact that AI can pass a Bar examination doesn’t indicate that it fully comprehends the justice system, its stakeholders and how other forms of expertise relate to the law. Without an understanding of both the specific case and a nuanced judicial system, lawyers relying on AI expertise may risk misinterpreting evidence and making mistakes. 

In fact, a recent study suggests that reliance on AI increases the need for people who, based on domain expertise, can validate AI outputs, adapt to new situations and plan for potential issues. Another concern is that as companies scrap entry-level positions to cut costs, they risk eventually losing the pool of middle managers on whose contextual understanding they now depend on.

AI also encourages homogeneous thinking. Studies have shown that large language models (LLMs) tend to fixate on the first idea generated and exhibit a bias towards averaging. This encourages users to ignore surprising or unexpected information, even though this is exactly what sparks creative insights. Generative AI (GenAI) use thus leads to the development of monocultures that lack the diverse experiences and perspectives essential for innovation. It may also impact companies’ agility, as people are more likely to miss signals in the environment inviting change.

How humans can shape AI 

To control the risks, we need to treat technology not just as a tool, but as part of a system that humans can shape. That means findings ways to design AI-enabled work environments that preserve human agency. Here are three possible ways to achieve this. 

1. Think of new ways to approach AI 

The Greek philosopher Socrates used to ask his followers questions as a way to lead them to independently find answers to problems; or, in some cases, to admit their ignorance. This approach can be used with AI, as shown in the following examples. 

The Prompt With Me Challenge: This was launched by global pharmaceutical company AstraZeneca in response to the issues around using AI in the workplace. The 10-day programme was designed to encourage employees to understand and experiment with their usage of AI. The goal was to develop critical thinking – not just by learning how to prompt better, but by developing ways to use AI to dig deeper and learn more.  

The 1,500 participants were challenged to be self-reflective and introspective on how they used and interacted with AI via their prompts. The initiative grew rapidly within the organisation and became a template for how to approach other workflows, functions and AI tools. Today, it has been adopted by more than 50 teams and functions within AstraZeneca. 

The BUILD2GETHER–AI-SDG Hackathon: ETH Zürich, Nanyang Technological University and Sairam Engineering College organised a 24-hour hackathon that featured nearly 2,000 participants across close to 500 teams. Participants received different types of AI feedback on proposals for initiatives to help the United Nations achieve its Sustainable Development Goals.  

Half the teams got feedback on how to improve their ideas to better reach and fit the predefined goals. The other half gained feedback that identified redundancies between their proposals and projects that had already been implemented, encouraging them to experiment with alternative concepts. When AI challenged participants from the second group, they came up with a wider range of more innovative ideas. 

2. Understand the right place and time to use AI 

In establishing AI governance frameworks, companies often focus on optimising where they can implement AI. Rather, they might want to consider where and when AI should not be used. Here are some suggestions. 

Create AI-free space: Organisations can encourage employees to refrain from using GenAI tools during specific time blocks or provide them with “gated interfaces that unlock assistance only after the user deposits their own context”, as the authors of this article put it. Another group of researchers reported how an Australian telecommunications firm required mid-level managers to engage in “AI-free strategy sessions” before using AI tools, forcing them to first brainstorm using their own judgement and experience. 

Centre human agency: Technology is ubiquitous in many organisations’ development plans and growth strategies – so much so that its use can become unquestioned. Ensuring that humans can shape technology starts with giving employees, especially entry-level ones, a chance to develop proficiency and contextual knowledge. This can be achieved by requiring them to perform tasks without the help of AI. 

In a study on junior investment bankers, researcher Callen Anthony found that those who learnt through practice were better able to do the work and more likely to spot mistakes and suggest alternatives. 

Anthony also found variations among senior bankers: One group saw algorithmic technology as merely a tool for producing outputs, thus weakening their expertise; another made sense of the different technologies and how they shaped their work. The latter was able to deepen their expertise and meaningfully engage with junior analysts, asking them to justify their tool choices and explain the assumptions and calculations beyond the numbers they presented. 

Process in parallel: In certain situations, it may be useful to have AI and a team of humans perform the same task independently in parallel. For example, design consultancy With Company once deployed two teams to work on the same branding challenge: one using a more traditional design process, and another relying heavily on AI-supported workflows.  

The latter team generated close to 1,700 images before arriving at a final proposal, dramatically expanding the exploration space and accelerating divergence. The human-led team produced fewer routes – but with stronger coherence, sharper narratives and greater contextual framing from the start. 

The human-led team’s proposal was ultimately selected, but the more general lesson was that parallel processing revealed differences in the relative strengths of AI and human teams at different stages in the creative process.

3. Experiment with different interfaces 

Many GenAI and LLM models are formatted around “question and response” chatbot interactions. This often devolves into a messy “conversation” where, for instance, a wall of AI-generated text is simplified, restructured and reinterpreted by an office worker who’s overwhelmed by the sheer amount of information being presented. GenAI then reinforces that confusion by responding in a way that structurally mimics the user’s response. 

To get around this, we invite companies to challenge the fact that AI is just a chatbot. Indeed, a recent paper has proposed moving away from the “frictionless user interface” and replacing it with one that surfaces raw data and competing evidence, forcing the user to make sense of and deliberate on the proposed alternatives. 

Such interfaces can work well in contexts like mathematics or coding, where users are already experts. Take Google. After realising that current chat interfaces didn't best assist mathematical research, it created an AI tool with a flexible interface and annotation system that helps mathematicians develop workstreams, verify proofs and check for errors. 

Don’t relinquish control to AI 

As AI continues to be deployed in organisations, leaders must recognise that technology isn’t neutral, and that it can influence how we think and interact in negative ways. Managers must therefore go beyond the hype and adopt a holistic, balanced view of how to use AI. 

Instead of focusing on short-term gains, companies should support employees’ development as active knowledge producers by strategically experimenting with AI to identify how it can bolster work, preserve expertise and nurture diverse thinking. In other words: Shape AI, rather than be shaped by it. 

This is an adaptation of an article published in Harvard Business Review.