When Clarity Masquerades as Truth

Rethinking How We Use AI

I am currently planning an ATV trip to the mountains. Like many people, I saw it as a perfect opportunity to use AI. I entered some basic details about the kind of trip I was looking for and asked it to suggest a few possibilities.

Within seconds, I had several complete travel plans. Each included a destination, suggested trails, places to stay, and activities. The plans were organized, detailed, and convincing. They did not feel like rough ideas. They felt finished.

But I had never been to any of the places it recommended. I was also planning the trip for more than just myself. A couple of other people are coming with me, and they would be investing their time and money based, at least partly, on my judgment. I felt responsible for making sure the trip was enjoyable, appropriate, and worth the cost.

That sense of responsibility changed how I approached the recommendations. It was no longer enough for the plans to sound polished. I needed to know what they were based on. Was the information current? Were the trails suitable for our group? What had been left out? Were there local conditions that could change the decision?

So I started reconstructing the path behind the answer.

I visited tourism and trail websites, read firsthand accounts, examined maps, and watched videos of people riding several of the suggested routes. I tried to understand not only whether the trails looked appealing, but whether they would actually work for our group.

When I checked current wildfire maps and advisories, I discovered active wildfires in and around several of the recommended areas. The suggested rides might ordinarily have been excellent. The trails had not suddenly become less interesting, and the scenery had not become less impressive. But under the current conditions, they were not appropriate recommendations for our trip. What struck me was how complete the answer had felt before I began checking it. 

The itinerary was detailed enough to create the impression that the planning was finished, yet some of the most consequential information was missing. The answer did not need to be entirely wrong to lead me toward a poor decision. Much of it may have been accurate, but it was not sufficiently current or complete for the decision I needed to make.

Twenty years ago, I studied how people decided whether to trust health information on the internet. At the time, the challenge was evaluating a website. Who wrote it? Who sponsored it? Was the information current? What evidence supported its claims? Were other perspectives available?

The internet gave people access to more information than ever before, but it also placed greater responsibility on them to decide what deserved to be trusted.

Today, we might not even see the websites.

We ask AI a question and receive an answer. Depending on the tool, that response may be generated from training data, retrieved from current sources, or produced through a combination of the two. Either way, the searching, selecting, comparing, and interpreting often remain largely invisible. We may never see the competing possibilities, disagreements in the evidence, assumptions that shaped the response, or information that was excluded.

This is often described as an accuracy problem. AI can be wrong. It can rely on outdated information, generate nonexistent or misattributed sources, overlook important context, or present uncertain claims with confidence. But I think the deeper problem is about critical thinking.

Critical thinking does not begin and end with determining whether an individual fact is correct. It also requires us to examine how a question has been framed, what evidence has been used, what assumptions connect the evidence to the conclusion, what alternatives have been overlooked, and how much confidence the information justifies.

In my case, I investigated because the decision mattered. Other people were relying on me, and I understood that I would remain responsible for what happened if we acted on the recommendation. AI could help me plan the trip, but it could not assume responsibility for the consequences of the plan.

But what happens when we do not immediately recognize the consequences of accepting an answer: when convenience matters more than verification, when we know little about the subject, or when the response is so clear and complete that further investigation feels unnecessary?

Twenty years ago, my research examined why people critically evaluated some of the information they encountered online while accepting other information with little scrutiny. Today, the challenge is more difficult. We must first recognize that an apparently complete answer may have concealed much of the process we need to evaluate.

When the source disappears from view, critical thinking requires us to look for it. When the reasoning is hidden, we have to reconstruct it. And when an answer arrives before we have encountered the evidence or considered the alternatives, we have to resist the feeling that the thinking has already been done.

Critical Thinking Has Never Been Automatic

When I conducted my earlier research, the internet had already changed how people accessed information. Knowledge that had once been difficult to locate was suddenly available to almost anyone with an internet connection.

This was especially significant in health. People could search for symptoms, investigate diagnoses, compare treatments, and access some of the same research used by health professionals. 

Access was empowering, but it also transferred more responsibility to the person searching. Finding information was only the first step. People still had to decide who created it, why it had been published, whether it was current, and what evidence supported it.

At least some of those clues were usually visible. A website might identify an author, organization, sponsor, or publication date. A reader could visit the About page, examine references, or compare its claims with another website. The clues were not always easy to find or complete, but the source itself could be examined.

The problem was that people did not necessarily examine it. Research at the time suggested that credibility judgments were often influenced by visual design, professional appearance, ease of navigation, or confident presentation. These features were weak substitutes for examining authorship, evidence, purpose, accuracy, and currency.

That led me to a question that still feels important today: Why do people critically evaluate some information but accept other information with very little scrutiny?

It is tempting to treat critical thinking as a stable personal quality. We describe someone as a critical thinker, as though they either possess the ability or they do not. But my research suggested that critical appraisal is more conditional than that. Whether we examine information carefully depends on several factors operating together.

First, we need to know how to evaluate it.

Telling someone to “think critically” is not especially helpful if they do not know what to look for. A person may read carefully and still never ask who produced the information, what interests may have shaped it, what evidence supports it, or what important details have been left out.

Knowing which questions to ask directs our attention. If we know that currency matters, we are more likely to look for a publication date. If we understand sponsorship, we are more likely to investigate who funded the information or benefits from it. If we know that conclusions should be proportional to the evidence, we are more likely to notice overstated certainty.

Second, we need enough knowledge of the subject to recognize when something may be wrong.

This is one of the persistent difficulties of evaluating information. We often search because we do not already know the answer. Yet the less we know about a subject, the harder it can be to identify what is implausible, incomplete, outdated, or missing.

Someone with experience riding mountain trails, for example, may immediately recognize that a route is unusually difficult, that an estimated travel time is unrealistic, or that a particular type of terrain requires equipment that has not been mentioned. Someone without that experience may see only a detailed and confidently written itinerary.

We do not need to become experts before seeking information. We do need to recognize the limits of our knowledge and know when a decision requires expertise, independent evidence, or closer investigation.

Third, we need the practical ability to investigate.

When websites were the main point of access, this included knowing how to search effectively, navigate a page, trace an author, interpret a web address, locate a publication date, and compare multiple sources.

Interestingly, my research found that knowledge and skill with the Web did not directly predict critical appraisal. Being comfortable with the technology was not the same as being able or willing to judge the information it provided.

That distinction is even more important with AI. A person can become skilled at writing prompts, refining responses, and producing polished material without becoming skilled at evaluating the claims, assumptions, and evidence within it. Technical fluency can make us more efficient users of a tool. It does not necessarily make us more critical users.

Finally, we need a reason to invest the effort.

Critical thinking takes time. It can complicate an answer we would prefer to keep simple. It may require us to open additional sources, compare conflicting perspectives, tolerate uncertainty, or reconsider a conclusion we already like.

In my earlier work, I understood motivation partly in terms of personal relevance and personal responsibility. We are more likely to examine information carefully when the outcome matters to us and when we believe we will be responsible for the consequences of the decision.

That was precisely what happened as I planned the ATV trip. Had the trip been hypothetical, I might have read the itinerary, thought it sounded impressive, and moved on. The quality of the answer would have been the same. What changed was my willingness to question it.

We make countless low-effort judgments every day and cannot investigate everything with the same intensity. We rely on shortcuts, especially when a decision appears routine, the consequences seem distant, or the answer aligns with what we expected.

What AI has changed is the environment in which those judgments are made. A collection of separate websites, authors, publication dates, and competing claims can be compressed into a single response. The answer arrives coherent, organized, and apparently complete.

Critical thinking was never automatic when the source was visible. Now the source may be hidden, and the answer may be more persuasive because of its fluency. The very features that make AI useful can also make further investigation feel unnecessary.

The Fluency Problem

One reason AI-generated answers can be so persuasive is that they are easy to process. They are clear, logically structured, and responsive to the question we asked. Complicated information is organized into headings, summaries, recommendations, and next steps.

This fluency is one of AI’s greatest strengths. It can make unfamiliar subjects more accessible, translate technical language into plain language, and help us see connections that might otherwise take considerable time to identify.

But fluency can also create an illusion. Research on processing fluency suggests that information that is easier to process may be more likely to be judged as true. When an answer is coherent, we may assume that the reasoning behind it is equally coherent. When it is detailed, we may experience it as comprehensive. When it is delivered with confidence, we may not notice how much uncertainty has been removed from view.

The answer feels complete because it reads as complete.

Yet an answer can be clearly written without being well supported. It can be detailed without including the details most relevant to the decision. It can be factually accurate in several respects while still being misleading as a whole.

In my itinerary, the missing information changed the judgment that should be made from the information that was included. This is why evaluating AI output cannot be reduced to checking isolated facts. We also have to ask whether the answer is sufficiently current, complete, and appropriate for the decision in front of us.

A polished response can make those questions feel less urgent.

If I had planned the trip using a conventional search engine, I would have encountered separate sources: tourism pages, trail maps, rider videos, news reports, and government advisories. The process would have been slower and less orderly, but I would have seen that each source served a different purpose. None would necessarily offer the whole answer. I would have had to assemble it myself.

AI changes that experience. When connected to retrieval or search tools, it can draw on information from multiple places and present it through a single voice. Even when it is not actively searching, it can create the same impression of synthesis. The rough edges disappear. Differences in emphasis become less obvious. AI may blend conflicting accounts into a general conclusion. Gaps in the available information may disappear entirely. .

We receive the result without necessarily seeing the disagreement, uncertainty, or selection decisions behind it. When we search through sources ourselves, we decide which links to open, which claims to compare, and how much confidence the evidence deserves. With AI, many of those decisions may already have been made for us or may simply be invisible.

We regularly rely on summaries because no one has the time or expertise to investigate every question from the beginning. Good summaries can reduce unnecessary effort and direct our attention to what matters.

The concern is that AI does not always make the decisions behind its response visible. We may not know why information was included or excluded, how recent it was, or whether competing interpretations were considered. Unless we deliberately ask, we may not recognize that there were decisions to examine.

The technology therefore does more than provide information. It shapes the conditions under which we encounter and judge that information. This recalls Marshall McLuhan’s observation in Understanding Media that “the medium is the message”: a medium influences us not only through what it communicates, but through how it reorganizes our relationship with information.

A polished answer does not prevent critical thinking. It makes critical thinking feel less necessary.

This matters especially when we know little about the subject. If an answer contradicts something we already understand, we are likely to pause. But when we lack the knowledge needed to recognize a problem, clarity and confidence may become substitutes for evidence. We may judge the quality of the answer by how well it is expressed because we have few other ways to judge it.

There is an uncomfortable paradox here. We often turn to AI because we do not know enough about a subject to answer the question ourselves. Yet that same lack of knowledge can make us less capable of recognizing when the response is incomplete, misleading, or wrong.

The people most in need of assistance may also be the least equipped to evaluate the assistance they receive.

A 2025 survey of 319 knowledge workers found that confidence in generative AI was associated with less reported critical-thinking effort in AI-assisted tasks. The study does not establish that AI caused a decline in critical thinking, but it does reinforce the importance of how much confidence we place in the tool. Read the CHI study.

This does not mean we should stop using AI for unfamiliar subjects. It means our confidence in an answer should not exceed our ability to evaluate it. The less we know, the more important it may be to ask for sources, compare the response with independent information, seek relevant expertise, or treat the answer as a starting point rather than a conclusion.

AI can make information easier to understand, but it cannot guarantee that what we understand is sufficient for the judgment we need to make.

Critical Thinking Is More Than Fact-Checking

When people are encouraged to verify information generated by AI, the advice often comes down to two suggestions: check the facts and ask for sources.

Both are useful, but neither is sufficient.

An answer can contain accurate facts and still lead to a poor conclusion. Even if the facts in my itinerary were accurate, they were incomplete, and the missing information changed what those facts meant.

Critical thinking therefore requires more than determining whether individual statements are correct. It requires us to examine the path from the question to the conclusion.

First, we need to examine how the question was framed.

When we ask AI for the “best” destination, course of action, explanation, or solution, the word “best” can conceal a surprising number of assumptions.Best for whom? Based on what priorities? Under what conditions? What trade-offs are acceptable? What constraints need to be considered?

The best ATV trip for an experienced rider may not be the best trip for a mixed group. The most scenic route may not be the most accessible. The least expensive destination may require more travel time. A trail that is ideal in June may be inaccessible, unsafe, or unpleasant in August.

The quality of an answer depends partly on whether the right question was asked. But we do not always know enough about a subject to ask the right question at the beginning.

AI will often compensate for missing information by making assumptions. This can be helpful. A system that stopped to request clarification about every missing detail would quickly become frustrating to use. The problem is that we may not know which assumptions it has made.

The response can appear to answer our question when it has answered a slightly different one. I thought I was asking, “Where should our group go for an ATV trip?” The AI may have effectively answered, “Which mountain destinations are generally known for good ATV riding?” My actual decision also involved current conditions, travel costs, group needs, and whether the experience would be worth the investment.

The recommendation addressed the general topic without fully addressing the decision. Before deciding whether an answer is correct, we need to ask whether it is answering the question we actually need answered.

Second, we need to examine the evidence behind it.

Asking AI for sources can make some of the hidden path visible, but a list of links is not verification. A source may be real without supporting the claim, credible but outdated, or accurate about typical conditions while saying nothing about present ones. Several sources may also appear independent while repeating information from the same place.

Sources also serve different purposes. A tourism website may accurately describe why a destination is worth visiting, but it is not necessarily the best place to learn about wildfire conditions or road closures. A personal testimonial may reveal what an experience felt like, but it cannot establish that everyone will have the same experience. A government advisory may provide authoritative safety information while saying little about whether the trip will be enjoyable.

Critical thinking involves understanding not only whether a source is credible, but also what kind of claim that source is equipped to support.

Third, we need to examine how the evidence is being used.

AI responses often move smoothly from information to interpretation and from interpretation to recommendation. Because the transitions are well written, we may not notice when the nature of the claim has changed.

Consider the difference between these three statements:

•      A region has an extensive network of ATV trails.

•      The region offers varied riding opportunities for different experience levels.

•  The region would be an excellent destination for your group.

The first statement may be a straightforward fact. The second involves interpretation. The third is a judgment that depends on the group, current conditions, costs, timing, and alternatives. Each requires different support, yet an AI response may present all three with the same confidence. This blurs the distinction between what is known, inferred, and recommended.

A critical reader needs to separate those layers. What information is directly supported? What conclusions have been drawn from that information? What assumptions connect the two? Does the available evidence justify the strength of the recommendation?

Many questions we ask AI are questions of judgment: what we should do, which option is better, or what a set of findings means. These questions involve values, priorities, context, and trade-offs. AI can help organize those considerations, but it cannot decide which should matter most without making assumptions about what we value.

Finally, critical thinking also requires us to pay attention to what is missing.

This is challenging because missing information does not announce itself. We can question a claim that appears in front of us, but it is much harder to question a perspective, condition, or alternative that was never included — and the more coherent the answer, the easier it is to assume nothing important is missing. This is where critical thinking requires deliberate reconstruction. We have to ask questions that make the missing parts of the process more visible:

•  What assumptions is this answer making?

•  What information would I need before acting on it?

•  What credible alternatives should be considered?

•  What evidence might lead to a different conclusion?

•  Whose perspective is represented, and whose might be absent?

•      Which parts of the answer are sensitive to changing conditions?

•  What remains uncertain?

Critical thinking cannot guarantee perfect decisions. These questions can, however, reduce the likelihood that we accept an answer simply because it was presented clearly and confidently. They also help us calibrate how much confidence it deserves.

Not every question requires an exhaustive investigation. If I ask AI to suggest a recipe using ingredients in my kitchen, a poor recommendation may produce a disappointing meal, but the decision is easy to reverse.

Other decisions deserve more scrutiny. The stakes are different when an answer concerns health, finances, education, professional responsibilities, or someone’s well-being. They also rise when information is time-sensitive, we know little about the subject, or a mistake would be difficult to undo.

We do not need to be suspicious of every AI-generated sentence. But we should match the depth of our scrutiny to the consequences of being wrong. The greater the stakes, the less appropriate it is to allow fluency to stand in for evidence.

There is one more question that may be particularly important:

Could I explain and defend this conclusion without relying on the AI’s wording?

It is possible to read an answer, understand each sentence, and still not understand the reasoning well enough to make it our own. We may be able to repeat the conclusion without being able to explain why it follows, what evidence supports it, what its limitations are, or what would cause us to reconsider it.

We possess an answer, but we have not yet formed a judgment.

This matters in education, where students may submit ideas they cannot independently explain; in the workplace, where someone may present a recommendation without understanding its assumptions; and in leadership, where decisions must be justified to the people affected by them. It matters whenever we act on advice whose consequences the technology itself will never have to bear.

AI can assist with critical thinking. We can ask it to identify assumptions, distinguish evidence from interpretation, present competing explanations, or describe what would weaken its conclusion. These uses make the tool part of the investigation rather than merely a source of answers.

But asking AI to critique its own response is not the same as independently verifying it. The system that created the original blind spot may reproduce that blind spot when asked to review its work. At some point, important claims still need to be examined against evidence outside the response.

Critical thinking in the age of AI therefore involves a shift in posture. We cannot remain passive recipients of polished answers. We have to become active judges of how those answers were constructed, what they include, what they leave out, and how much confidence they deserve.

Fact-checking is only one part of the work.

The larger task is to recover the question, evidence, assumptions, alternatives, uncertainties, and judgments that were compressed into the answer. Only then can we decide whether the conclusion is one we are prepared to accept as our own.

The Thinking Is Still Ours

AI has made it possible to move from a question to an answer with remarkable speed. It can organize complicated material, identify possibilities, and produce recommendations that might otherwise take hours.

The answer I received for my ATV trip gave me destinations to investigate, trails I had not previously considered, and a starting point for planning. The problem would have been treating that starting point as a finished decision.

Once I looked beyond the itinerary, I was no longer asking whether the recommendation sounded appealing. I was asking whether the evidence justified acting on it. That required current information, independent sources, knowledge of our group, and judgment.

AI can participate in that process, but it cannot assume responsibility for it. It will not spend the money, travel the distance, ride the trails, or face the consequences if important information has been overlooked. It can generate a recommendation, but it cannot care whether the trip succeeds. The responsibility remains with me.

The same is true when we use AI in other parts of our lives. 

A student remains responsible for understanding the ideas submitted under their name. 

A professional remains responsible for a recommendation presented to colleagues or clients. 

A leader remains responsible for decisions that affect other people.

A parent remains responsible for advice acted upon on behalf of a child.  

AI may contribute information, language, analysis, or possibilities, but it does not inherit responsibility for what happens next.

This is one of the most important reasons to preserve critical thinking. It is not simply an academic skill or a method for identifying false information. It is part of how we assume responsibility for our judgments. It is how we move from an answer that was given to us toward a conclusion we understand, can defend, and are prepared to act upon.

Twenty years ago, I was interested in whether people would critically evaluate the websites they encountered. Today, I am concerned that they may never encounter those websites at all. The author, evidence, purpose, context, and competing perspectives may be compressed into a response that appears complete before we have had an opportunity to examine any of them.

AI can help us search. It can help us compare. It can even help us question its own conclusions. But it cannot decide how much confidence those conclusions deserve. It cannot determine which values should guide us. And it cannot accept responsibility for the choices we make.

The source may have disappeared from view, but our responsibility to look for it has not. The reasoning may be hidden, but our responsibility to examine it remains. And although AI can give us an answer, the thinking that turns that answer into a judgment must still be ours.