Chapter 6

Critical thinking

20 min read

Critical thinking can be regarded as the disciplined practice of evaluating information, arguments, and beliefs using logic, evidence, and intellectual humility. Though the term is widely used by educators, professionals, and policymakers, it is frequently misunderstood. It is often used without giving it sufficient thought and acknowledging, for example, disciplinary differences and difference between day-to-day critical thinking and professional or research induced thinking. In view of the paragraphs about the risks of generative AI, the general feeling is that generative AI will come at the cost of critical thinking abilities. Some researchers however see opportunities to involve generative AI to improve critical thinking. We will dive into this in the following paragraphs.

#6.1Knowledge is conditional for critical thinking

Before discussing the role of knowledge in critical thinking, it helps to name a debate that often remains implicit. Is critical thinking a general, transferable skill that students can apply across domains once they have learned it? Or is it tied to expertise within a specific domain? Much of the public and educational debate assumes the former, but research in cognitive science points in the opposite direction: the ability to reason well, solve problems, and make sound judgments depends heavily on prior knowledge within the relevant field (Van Gelder, 2005; Willingham, 2020). Someone who reasons sharply about economics does not automatically reason sharply about literary criticism or molecular biology. This is why a foundation of domain knowledge is not optional but conditional for critical thinking.

Box 6.1

Six lessons for teaching critical thinking

Van Gelder (2005) drew six lessons from cognitive science that are still useful as a practical checklist for course design. Teachers can use them to review whether a course or program genuinely supports the development of critical thinking, or merely assumes it happens automatically.

1. Critical thinking is hard. It is a skill that only becomes possible once more basic skills, such as reading comprehension and foundational domain knowledge, are sufficiently developed.

2. Deliberate practice is essential. Exposure to a subject does not automatically produce critical thinkers. Students need explicit, structured opportunities to practice the skill itself, preferably using evidence-based techniques such as spaced practice and retrieval practice with appropriate feedback to improve performance.

3. Transfer must be practiced separately. The ability to apply critical thinking in a new context is itself a skill that needs active teaching. Students who have learned to argue well in one course often cannot do so in another without explicit bridging instruction.

4. Both practical and theoretical knowledge contribute. Hands-on experience in a domain deepens judgment, but conceptual knowledge of argumentation structures and logical fallacies strengthens that judgment further. The two reinforce each other.

5. Visualizing arguments, through argument mapping, can support critical thinking by making abstract reasoning concrete. The evidence on effectiveness on argument mapping is mixed rather than conclusive, so teachers should treat it as one useful technique among several rather than as a silver bullet (Davies, 2011; Nesbit & Adesope, 2006).

6. Beliefs get in the way. Students find it difficult to reason critically about claims that conflict with their existing convictions. Confirmation bias and belief preservation deserve explicit attention, for example by asking students to construct the strongest possible version of an argument they disagree with.

The main critique on generative AI is that it is a statistical machine and that it consequentially can make outright mistakes. It can provide incorrect facts and present them as correct. Many teachers focus on this characteristic of language models and, rightly so, fear that students will not be aware of these errors or be able to detect them. Teachers focusing on correctness of information (or concepts, procedures, numbers, proofs etc.) sometimes label the inability to detect errors as a ‘lack of critical thinking’ by the students. However, this ignores what education is for. Students enter higher education with only a limited ability to think critically. Teaching them that ability is the task, and educators are responsible for having increased it by the time students leave. A great and worthwhile task.

It is true that to analyze a problem effectively, students must first possess sufficient prior understanding of the relevant domain. But knowledge by itself is not enough for students to get to critical thinking or questions. At the same time students are still learners, it's important to acknowledge that they still lack certain knowledge. As a teacher, you are there to guide them in this process and not judge their lack upfront.

For example, a student cannot critically evaluate a historical argument without knowing the relevant facts, nor assess a scientific claim without grasping the underlying principles. Without this foundation, learners risk engaging with questions they are not yet equipped to address meaningfully (disruptive friction, see paragraph 4.4.4). So it's important to acknowledge and communicate to students the importance of this knowledge as a stepping stone for learning and reaching a level of critical thinking. And not just point out the questions that you may perceive as uninformed.

Oakley et al. (2025) describes this in detail in the article ‘The Memory Paradox: Why Our Brains Need Knowledge in an Age of AI’. For humans to effectively work with AI, they need to maintain strong internal cognitive models (or "schemata" and neural manifolds), that allow them to critically evaluate, refine, and guide AI-generated output.

Research into inquiry-based learning consistently shows that it is less suitable for efficient knowledge building. This is because unguided discovery learning is problematic for most learners due to the too high cognitive demands of the inquiry process itself (Van Joolingen et al., 2007). Therefore, direct instruction and deliberate practice remain essential (Kirschner et al., 2006). These practices are the most suited to effectively acquire foundational knowledge that is functionally sufficiently automated (second nature, or working knowledge), so that cognitive resources are freed to process and acquire new information and insights. Only then can learners productively explore the boundaries of what they know and do not know.

Viewing tip 6.1 See this video of Richard Feynman in which he explains the characteristic of productive questions for trying to understand the working of magnets. — *Viewing tip* 6.1 See this video of Richard Feynman in which he explains the characteristic of productive questions for trying to understand the working of magnets.

Generative AI adds a further layer of potential misinformation: by their nature, generative AI systems produce artifacts that carry some degree of incorrectness while seeming correct at first glance. For any curriculum that pays attention to critical thinking, a focus on logical flaws alone is therefore not enough. It has to go together with sufficient knowledge of the domain, or with the inclination to acquire the facts, principles, methods and theories quickly.

However, this brings us also back to the paragraph 3.1 about the inverted Bloom’s taxonomy. That paragraph makes two arguments. The first is that learning is a messy process, in which learners move back and forth in cycles of understanding and schemata building. The second is that generative AI may let students approach more complex problems than before, because the technology generates and analyzes explanations, arguments and lines of thinking so much faster.

The line between premature questioning and productive inquiry is getting thinner given the presence of generative AI, but one cannot always know in advance on which side a particular question falls. This uncertainty is not a flaw, but exactly touches the reason of education: educators are there to guide students along this path.

#6.2Logic

Besides knowledge, critical thinking requires a good understanding of logic. For example, deduction draws specific conclusions from general principles: if the premises are true, the conclusion must follow (e.g., "All humans are mortal; Socrates is human; therefore, Socrates is mortal"). Induction moves in the opposite direction, from observations to broader generalizations (e.g., observing repeated sunrises and inferring the sun will rise tomorrow). Understanding these and other forms of reasoning and their limitations helps learners to construct sound arguments and identify flawed ones.

A critical thinker identifies claims and the evidence that support the claims, recognizes common fallacies, detects hidden assumptions, and seriously considers alternative explanations. The guiding questions are: What exactly is being claimed? What evidence supports it? What would change my mind?

Generative AI likely calls for a stronger emphasis in education on training students in logic. And logic is not easily learned. Specialized courses, and ongoing attention during all courses and (research) projects are needed.

#6.3Logical fallacies

Critical thinking therefore requires deliberate training, even more so when using generative AI undermines or challenges it. This training can be incorporated into formal education (such as university educational programs and courses) and students can be given sufficient time for training and application in projects.

This is different when looking at daily or even professional life. And we think it is necessary here to emphasize that the political and societal attention for AI Literacy stems most from that background. Citizens, also well educated individuals, seem more and more to be unaware (and sometimes even unwilling) of their lack of knowledge and lack of effort to put in reasoning and are victim of logical fallacies.

In daily or professional lives, there are no program and course guard rails, and individuals must apply critical thinking independently themselves. That is not easy. It demands a conscious effort to pause before drawing immediate conclusions when confronted with a claim or event, whether from an advertisement, a politician, stakeholders with different interests, a company, or, indeed, an artifact produced by generative AI. Without that pause, reasoning quickly becomes unsound and prone to logical flaws. Or even worse people will react based only on their gut feeling. In the age of algorithmic social media bubbles, gut feelings are very easily confirmed and further inflated.

Box 6.2

Two systems of thinking

The distinction between intuitive and deliberate reasoning, articulated by Daniel Kahneman (2011) as System 1 and System 2 thinking, offers a useful framework for understanding why logical flaws are so persistent. System 1 operates automatically and quickly, relying on heuristics and pattern recognition that served humans well in evolutionary terms but can lead to systematic errors in complex reasoning tasks. These are the gut reactions, snap judgments, and the seductive sense that an argument just feels right. It is also the system most prone to cherry-picking and other biases, since it naturally gravitates toward information that confirms existing beliefs while filtering out dissonant data.

System 2, by contrast, is slower, more effortful, and capable of logical analysis, but it requires deliberate activation and is easily bypassed when cognitive resources are limited, emotions are high, or when the conclusion aligns with what we already believe. Effective critical thinking training, in this light, is largely about teaching learners when and how to engage System 2: to recognize the situations in which intuitive reasoning is likely to mislead, and to develop the mental habits that trigger more careful analysis before reaching a conclusion.

So, it is not surprising that individuals, and thus also generative AI, frequently exhibit a lack of critical thinking and produce logical flaws in their reasoning, sometimes knowingly but more often without awareness.

According to the A Broader Mind concept of VU Amsterdam, it is important that we pay attention to this form of critical thinking capacity and attitude in day-to-day normal life and profession. Being mindful of the following examples is important, but even more so in guiding students when using generative AI – which sometimes reinforces these biases.

Confirmation bias is a flaw where one selectively attends to evidence that supports pre-existing beliefs while discounting contradictory information.
Its close relative is Cherry picking, which involves deliberately selecting only those data points or examples that support one's argument while ignoring the broader body of evidence that might complicate or contradict it.
And it is close to Causal oversimplification or Single cause fallacy, which occurs when a multi-causal problem is attributed to one visible factor while larger but less visible contributors are left out.
Other frequent biases include:
- ad hominem reasoning, which dismisses arguments based on the character of the person making them object of the argumentation rather than their content;
- false dichotomies or black and white-thinking, which reduces complex issues to two opposing options when more nuanced positions exist;
- hasty generalization, drawing broad conclusions from limited examples;
- falsely appealing to authority: learners also commonly accept claims simply because an authoritative figure endorses them without examining the underlying evidence;
- post hoc reasoning: if because one event followed another, the first must have caused the second. Or assuming causation when there is only correlation. These flaws are pervasive precisely because they often feel intuitively correct and can be socially reinforced within groups that share similar assumptions.

Box 6.3

Common reasoning errors in debates about generative AI in education

Public and institutional debates about generative AI in higher education contain many positions that can be based on logical fallacies rather than careful reasoning. Both enthusiasts and skeptics commit them. Recognizing the pattern in a familiar context is good practice for spotting the same patterns elsewhere. Below are some examples to be mindful of:

False equivalence treats AI use in student work as morally identical to having a friend help with homework, ignoring substantive differences in transparency, attribution, and the relationship between the help and what the student is supposed to learn.
Equivocation shifts between "AI as a tool that supports thinking" and "AI as a substitute for thinking", using the same word for fundamentally different uses.
Straw man arguments characterize skeptics as wanting to ban all educational technology, rather than addressing their actual concerns about whether specific learning outcomes can still be reliably assessed.
Status quo bias assumes that pre-AI examination practices were already optimal, without examining what those practices already failed to assess.
Tu quoque, or whataboutism, dismisses AI concerns by pointing out that students have always cheated, without engaging with whether the present case is qualitatively different in scale, detectability, or consequence.
Slippery slope claims that any AI use in coursework will inevitably lead to graduates who cannot think for themselves, without evidence that this follows from carefully designed AI integration.

Again, training learners to identify and avoid logical fallacies and become critical thinkers requires deliberate and structured interventions. It can be specific courses in logic, critical thinking or debating. Many institutions use the Critical Thinking Mini Guide from Paul & Elder (2019) for that purpose.

Box 6.4

Critical Thinking mini guide

The Miniature Guide to Critical Thinking Concepts and Tools (Paul & Elder, 2019) is a gentle and brief introduction to core critical thinking concepts and principles for every profession, and domain of study. The book dives into how to deconstruct thinking through the elements of reasoning and how to assess the quality of our thinking. It distinguishes Elements of Thought and Intellectual Standards. The guide details these foundations, and how they can be applied in instruction to improve teaching and learning. The application differs substantially across disciplines, but the underlying structure remains consistent.

Figure 6.1 Image of a YouTube of University of Amsterdam explaining the main concepts needed for critical thinking.

When applied to daily decision-making, the framework encourages individuals to explicitly identify what they are trying to accomplish (purpose), what question they are addressing, what information they are using, and what assumptions underlie their thinking.

In professional contexts, the framework addresses a common challenge. Many professionals make decisions based on incomplete analysis or unexamined assumptions. By systematically working through the elements, practitioners can identify gaps in their reasoning before committing to action. For instance, a manager evaluating a policy change would benefit from asking whether the question has been clearly defined, whether relevant information has been gathered, and what implications the decision carries for different stakeholders.

In scientific fields, the elements map onto research methodology. Scientists formulate research questions (question at issue), collect data (information), develop theoretical frameworks (concepts), and consider alternative explanations (point of view).

In humanities disciplines, the same elements serve different functions. Literary scholars might use the framework to analyze texts by identifying an author's purpose, examining the concepts employed, uncovering underlying assumptions about human nature or society, and considering how different theoretical perspectives yield different interpretations.

Critical thinking, then, is not about being right. It is a method, a habit of mind, that prioritizes truth over comfort and careful reasoning over reflexive conviction.

Figure 6.1 An AI generated image of A Blueprint for Critical Thinking. If you want to practice with this kind of framework for Critical Thinking, go to this Chatbot on VU Amsterdam’s Cogniti to converse about a topic that interests you or let the bot generate a topic for you.

One effective approach is to have students analyze real-world arguments: political speeches, societal dilemmas, opinion pieces, social media debates, and business, marketing or research plans. They label the reasoning errors they encounter explicitly, including the cases where authors have cherry-picked evidence to build a misleadingly one-sided case.

You can supplement this with exercises where students construct the strongest possible version of an argument they disagree with, thereby forcing them to separate logical validity from personal conviction and to actively seek evidence they might otherwise ignore. Peer review exercises, where students critique each other's reasoning in a structured format, help develop the habit of scrutinizing arguments regardless of their source. An additional system VU Amsterdam has licensed is the system Civinc and (in the process of licensing) the critical thinking and argument mapping systems Kialo-edu. Additionally, teaching learners to slow down and articulate their reasoning steps in writing can expose hidden assumptions, inferential leaps, and selective use of evidence that would otherwise go unnoticed. Over time, these practices can shift flawed reasoning from an invisible default to a recognizable pattern that learners actively resist.

#6.4Moral convictions in relation to generative AI

Even as we encourage students to question assumptions and examine evidence, an unspoken boundary remains: certain positions function as moral convictions. These are the foundational assumptions that underlie ethical reasoning. Just as mathematics starts from basic assumptions that are accepted without proof, moral philosophy starts from claims taken as self-evident, such as 'causing unnecessary suffering is wrong' or 'people deserve equal respect'. They are not derived from other arguments; they are the premises from which moral arguments are built.

Moral convictions matter because they shape which conclusions a line of reasoning can reach. Two people who share the same logic, but start from different convictions will arrive at different conclusions, which helps explain why many ethical disagreements are so persistent: the debate is often not about logic, but about which foundational commitment underlies it. Such convictions can also carry internal inconsistencies, including ones widely held in academia, and recognizing these in our own thinking is part of teaching critical thinking by example.

How do moral convictions play a role in discussing the qualities and impact of generative AI? Mostly it comes down to discussions about bias in language models or the (lack of) applied technical guard rails.

#6.4.1Bias

First, the makers of the LLMs unconsciously or intentionally could reinforce several biases that could be considered a case of moral suspiciousness. This is the position presented by Bender et al. (2021). They reason that generative AI models are trained on large datasets of internet information and images, which likely over-represent stereotypical or the culturally most abundant, but not necessarily correct, information on the internet. And this is indeed the case, and hence Bender et al. (2021) talk about LLMs as stochastic parrots that amplify societal biases present in training data.

Box 6.5

An exercise in bias and moral axioms

Suppose for a workshop you first ask participants what would happen when they would ask a multimodal LLM to generate a picture of a physician before actually generating it?

When trying this out in workshops of the VU Education Lab, many participants think about this and respond that it would likely generate a picture of a white middle-aged male in a white coat, maybe with glasses. They reason that this is because the model was trained on large datasets of internet images, which likely over-represent this stereotype of what a physician looks like.

But, when the participants then generate the image, multiple things happen. Some pictures will show this stereotype. But others could show a female or maybe someone of color. And if students would redo the generation, other images might appear again. This is because this behavior of LLMs is known as such (responses vary because of the statistical nature of LLMs) and developers of newer models have attempted to correct for this through post-training interventions.

Additionally, we discuss what would have happened if we had asked the participants upfront before this exercise to draw a physician by hand?

Would they have thought about their own internal bias? And if they did so, how would they decide what kind of person they would draw? This thinking process draws them to the need to specify in the prompt what kind of person they would like to have generated in the image (a woman? A person of color? A young person with a green coat?). And this will depend most likely on the context and goal that the picture would need to serve.

So, it is not only within the statistical preference of an LLM what representation will be generated, but it is even more important that the user considers their personal awareness of stereotyping and accordingly decide on specifying the required output.

But it is not always so clear-cut what these kinds of cultural biases are. Think for example about questions related to cultural habits of individuals or groups. It makes a big difference if one were to ask a generative AI system to explain which hand gesture would mean ‘to calm down’ if you did not specify for which culture you are asking this. The same might be true, for example, about how to raise children or how to teach pupils. But also, simply when asked to generate a picture of a person. This person will likely be relatively young as compared to the average age in a society because the internet holds many more pictures of young people than old people.

A somewhat more covert bias will be for a multimodal model to generate a picture that shows ‘Humanitarian aid in Africa’. This particular example below sparked debate whether generative AI functions as a reinforcer of ‘poverty pornography’ (Fuchs, 2026). How does this relate to moral convictions?

Figure 6.2 Image of an AI generated picture of ‘humanitarian aid in Afrika’

#6.4.2Guardrails

With generative AI, much is said about the so-called guardrails. Though not often discussed, it is in the hands of the developers of generative AI models how they methodologically and technically ensure that a user will not get responses for requests that might generate taboo output. Think of outputs on how to build bombs, biological warfare arms, or how to commit suicide. Or in China: what happened during the Tiananmen revolution. The underlying moral axiom is that it is wrong – and even legally prohibited – to provide individuals with such information, as it could harm individuals or societies.

Things do change over time. Technical and ethical hackers have repeatedly demonstrated that guardrails can be circumvented. Or worse: the operators of some commercial models have loosened or removed guardrails themselves. Elon Musk, for example, did this for his Grok models in 2025, allowing the models to produce content that earlier generations of similar models declined to generate. Think of profanity, jokes about taboo subjects, and various image-alteration requests. Decisions like this reflect the moral axioms held by the people who build and steward AI systems, not only the technical capabilities of the models themselves. For students and educators, this raises a broader concern. The answers that a generative AI system provides are partly an artifact of choices made by its developers. Choices that can shift over time and that may differ between platforms.

Figure 6.3 Images of an AI generated picture by Elon Musk showing himself in bikini’s.

#6.5What this could look like in your course

Translating the framework of this chapter into a specific course requires you to make a few design choices, and then pick concrete activities that fit your discipline and the placement of your course in the curriculum. On the new VU EduBase website we assembled a collection of exercises you could use for this purpose. For most teachers, the practical entry point is to combine a few choices about how the course frames critical thinking, with a selection of two or three activities from the companion that students can repeat across the semester.

A first design choice is whether your course will treat critical thinking as a separate subject, with one or two dedicated sessions early in the semester on logic, fallacies, and the role of moral axioms. Or will critical thinking be an integrated running thread with short critical-thinking moments embedded in every session and applied to the content on the agenda. Neither is universally better. Disciplines where reasoning patterns are central to the subject, such as philosophy, law, or economics, could benefit from the integrated version. Disciplines where critical thinking is more related to knowledge and precise procedural methods, such as the natural sciences, could benefit from a dedicated unit followed by targeted reminders. The choice should match the cognitive demands students already face in your content.

A second choice concerns the role of generative AI in the activities themselves. In some activities, students use AI as the object of critique, evaluating AI-generated arguments, summaries, or images for reasoning errors and biases. In other activities, students use AI as a partner to support their reasoning, prompting it to challenge their assumptions, generate counter-arguments, or surface alternative positions, and then critically evaluating what it produces. Both approaches develop different aspects of critical thinking. We assembled a collection of exercises that includes exercises of both kinds, including bias-recognition exercises, AI source-attribution debates, AI-generated argument analyses, and discipline-specific critical evaluation tasks for fields such as biology, engineering, and the social sciences.

A third choice is how critical thinking will be assessed. If you only assess the final product of student reasoning, you measure the outcome but not the process by which the student got there. For a process-aware assessment, we assembled a collection of exercises covering options that include critical reflection logs, annotated AI-output critiques, methodological reviews of AI-generated research designs, and personal AI use policies that students draft and defend. These options allow you to assess the reasoning quality directly, rather than infer it from the final text.