AI is scaffolding, not a substitute: protecting human intent, expertise and capability.

Randstad Advisory Signals from the Edge: what is ‘cognitive offloading’ and how can you avoid the challenges it presents?

key findings
When people outsource their critical thinking to artificial intelligence, they risk losing essential skills and producing unoriginal work that doesn’t do what it sets out to do. Research shows that relying largely on AI often reduces problem-solving ability and efficiency on complex projects. To protect professional expertise, teams must use AI as a collaborative tool rather than a replacement for critical thinking. By keeping human intent and judgment at the center of technology use, organizations can preserve authentic communication and drive meaningful results.

In 2030, we’ll look back and wonder how we ever expected to build trust between the creators and consumers of content after outsourcing our intent to machines.

Today, employers and employees who use generative tools to stand in for real expertise, but don’t put human intent behind it, will produce AI-generated content that doesn’t  persuade anyone. 

The cost of this approach will arrive in two forms. The first is visible: Clients and stakeholders see the “AI slop” and lower their opinions of the people who send it to them. 

The second is quieter: The people producing the slop stop building the skills they are paid for and that are required to remain relevant.

Both come from the same habit — using AI to avoid thinking instead of using it to think better, deepen curiosity and create understanding. How we use AI, not just where and when makes a significant difference

a growing intolerance for the signs of the algorithm

People are now able to recognize the "algorithmic texture" of content almost immediately. They notice a predictable rhythm, call out dramatic phrasing and sigh at the long and polite answers that say very little. But quality outputs aren’t just long and polite; they provide real answers, avoid “fluff” and share ideas you haven’t already heard elsewhere.

Analyses of large language models have identified that the use of the word “delve” in scientific abstracts has increased by 1,400% between 2020 – 2024. The use of “intricate” has risen more than 600%. When outputs are filled up with these marker words, readers get the message that the content is AI-generated. 

Increasingly this means that content doesn’t achieve what it sets out to do. Instead of building trust and authority, it’s discarded by the decision-makers it was written for. They doubt its value. 

They ask themselves: If a partner, colleague or leader can’t be bothered to think a problem through, can they be trusted to solve one? Some will reduce or end relationships because of it. In fact, according to Edelman, 70% of C-suite leaders say a piece of thought leadership has led them to question whether to keep working with an existing supplier.

what is cognitive offloading?

“Cognitive offloading” means handing all the thinking required to complete a task to AI. It is important to ensure that people do not skip the effort needed to architect an outcome that would be the result of their intentional thinking. 

Preventing AI outputs that don’t resonate starts with preventing cognitive offloading. It’s critical to equip the people who use the AI with the training, time and resources they need to be truly effective, not to replace their thinking or expertise. This means ensuring measures within organizations are not just focused on executing a task at pace. It requires skilling on how to retain skills.

It also means making sure your people have the expertise to be able to judge whether AI outputs are right. Since the model can supply the structure and baseline knowledge less experienced talent may lack, it can make people appear more capable than they are.   

Stronger performers already have that baseline knowledge and structure though. As AI pulls everyone toward its own standard,  the gap that used to set the best people apart narrows.

43% vs. 17%
According to research published by Harvard Business School, the quality gain of using AI is 43% for below-average performers, compared with 17% for above-average performers.

The problem was bigger than this: The models can be convincingly incorrect, misreading problems, putting weight in the wrong places and drawing the wrong conclusions, all with complete confidence. Even users who know better, can be easily sucked in when they don’t commit effort to a proper review and just skim the output before publishing.

Because the output looks better, it passes internal review, and the lack of substance only shows once it reaches a client or end-reader who actually read it properly. 

The cost to our own capabilities will take longer to show. An MIT study that monitored people’s brain activity found weaker engagement in people who let the machine draft for them.  We must be careful not to produce fragile experts: people who can write a prompt but don't look past the surface of the answer before pressing "publish." When the tool is taken away, could the people still do the work?

We need AI as scaffolding, not as a substitute. Used without guardrails, it gives us the feeling of practice and productivity while damaging both our learning and the quality of what we produce.

loss of the authentic voice

When machines take over drafting, they can pull everything toward the average. When we hand rough drafts over for polish, we run the risk of smoothing away what makes them ours: the signals of our unique identities that let a reader hear a real person and their authentic voice. 

Writers are noticing. Their sense of ownership fell from 6.74 to 4.29 on a 7-point scale when AI drafted for them. The structure of their writing improved while their critical analysis weakened. Writing with an opinionated model also also pull people's arguments toward the model's position. This is an important realization; a slow process that often doesn’t get noticed until all the pieces come together, highlighting the fracture in narrative structure.

The same flattening can have a disproportionate effect on different people. For example, detection tools flag essays by non-native English writers as machine-generated 61% of the time vs. just 5% for essays by native-speaking U.S. school pupils.

Meanwhile, the market is beginning to pay more for the things that are being removed. Once readers know or suspect a machine has created the content, they discount it: Across 13 experiments with more than 5,000 people, trust fell consistently once someone disclosed they used AI, and trust fell furthest when a detection tool exposed the AI use.

What they valued was human agency: A person had thought it through and would stand behind it.

Fearing that judgement, people are hiding their use of AI. And leaders don’t always see where AI is shaping the work, so their organizations can’t learn from it or manage the risk.

what is the most interesting benefit AI offers?

Productivity is the least interesting benefit that AI can offer us. Used with intent, AI can improve our work and stretch our thinking beyond what we could manage alone. Many of us aren’t seeing that, because we’re using it to skip the work.

AI multiplies capability, so the quality of what comes out depends on what goes in: the expertise, the intent and the effort applied at the start. It was never a shortcut.

On complex work, people expect AI to make them faster. Against the clock it often makes them slower, yet they feel faster because the work feels easier. Many people don't trust what the machine produces though, and many don’t verify it either, so the rework builds up.

Then there is sycophancy. Models are trained to please, and they drop logical rigour to produce a flawed answer that matches a user's wrong assumption. A polite model is a risky one.

The people who get the most from AI treat it as a sparring partner. They give it structure and domain rules, ask it to reason step by step and tell it to challenge their assumptions. They add friction on purpose, addressing the sycophancy, and the quality improves with it.

Technology can scale our output, but only our intent, effort and judgement can scale our capability. How is your organization keeping human intent and judgement at the centre of its AI work today?

signals from the field: the cost of quality

• • -24% vs. +19%: Developers forecasted AI would cut their time on complex tasks by 24%. It made them 19% slower, yet they believed it made them 20% faster.
  • 29%: A leading model produced a "proof" of a false maths statement 29% of the time to please its user instead of pointing out the error.
  • -17% vs +127%: Unrestricted GPT-4 cut student exam scores by 17%. A version with guardrails raised their practice scores by 127% without the exam penalty.

how can we avoid cognitive offloading?

Here are three scenarios to help you test whether or not you’re relying on cognitive offloading:

1. 1. Look at the last complex problem you handed over to an AI.
      If the platform went down today and you had to explain the mechanics of that output from scratch to a stakeholder, could you do it, or did you outsource the thinking along with the typing?
      
   2. For your next project, have a point of view first and apply curiosity and critical thinking to the subject. 
      Change your standard prompt from "write a draft for X" to "act as an expert reviewer and ask me three probing questions to help me write a better draft for X." How does adding this deliberate friction change your mastery of the final product?
      
   3. Audit how your team currently defines AI success.
      Are you rewarding pure speed and output volume (which risks the 17% capability penalty), or are you tracking whether your team's independent problem-solving skills are actively growing (the 127% opportunity)? Understand how your team currently feels about work they perceive to be created by, or with, AI. What does this mean for what you are gaining and what you are losing?