Critical Thinking When the Machine Answers for Us

A scene that has become familiar: someone comments confidently on an article he has not read, because he fed its link to an AI assistant and took the "summary". Laziness is not the only problem; there is this question: if we delegate understanding and inference to the machine, who remains to judge what has been written?

Does the machine weaken our criticism?

We cannot be certain, but the evidence deserves attention. A study published on the Microsoft Research website and presented at CHI 2025 surveyed 319 knowledge workers and gathered from them 936 examples of using generative AI at work. It found that greater confidence in the tool is associated with less critical thinking while using it, and that greater confidence in one's own ability to do the task is associated with more. In scientific fairness: this is a survey resting on what participants say about themselves, and a correlation in it does not prove that the tool is the cause.

Human-factors scientists studied the same phenomenon years before language models appeared, under the name "automation bias". In 2010 Raja Parasuraman and Dietrich Manzey reviewed the available studies and concluded that people make errors of omission and of commission when automated aid is imperfect, and that this persists despite training and instructions.

Why does it err with confidence?

The authors of "Why Language Models Hallucinate" (2025) say that hallucination, a plausible but false statement, arises partly because training and evaluation reward confident guessing over expressing doubt, so models are optimised to be "good test-takers". This means fluency does not indicate knowledge, and a model that answers without hesitation is not necessarily more truthful than one that hesitates. We explained what these models do in the article "How Do Large Language Models (LLMs) Like ChatGPT Work?".

Why does demand for judgement grow?

When everyone is equal in the speed of producing text, discernment becomes the scarce effort. In the World Economic Forum's The Future of Jobs Report 2025, based on a survey of more than a thousand employers representing over 14 million workers in 55 economies, analytical thinking came first among core skills, and seven in ten employers considered it essential. This is the result of a survey of employers' opinions, not a final verdict, but it shows the direction of demand.

An experiment to do today

Ask an AI assistant to summarise a text you know well, then compare the summary with the original line by line. You will notice where it abbreviated well, where it added what is not in the text, and where it dropped an important qualification. This small experiment reveals the nature of the tool more honestly than any description, gives you a personal gauge of how far to trust it for this kind of task, and reminds you that a "summary" is not the "text".

Five small habits

1. Conditional trust: treat every answer as a first draft, ask it for its source, and open the source yourself.
2. Check numbers and names before passing them on; these details are easy to get wrong and hard to catch.
3. Ask better questions: framing a question is itself a critical exercise.
4. Keep the skill alive: do not delegate to the machine what you ought to master; a skill not practised weakens, and the survey result above perhaps hints at this caution.
5. Compare and disagree: set the answer against a second source and grant yourself the right to differ from the machine; the final decision must be borne by a responsible human, accountable before God and then before people.

In this way the machine does not take our minds; it is we who choose to hand them over or keep them. The most useful habit we can bequeath our children is not memorising answers but asking "who said it? what is the evidence?" before believing.