Humanity has never had this much access to information. Almost any question can be answered within seconds. Search engines can find millions of sources. Wikipedia can give us an overview of almost anything. YouTube can teach us skills that once required formal classes. And now AI can read, summarize, compare, translate, and explain information for us.
You would think all of this would make us better readers. Instead, something rather strange is happening: reading performance is getting worse.
The OECD’s PISA 2025 assessment found that reading scores across OECD countries have continued to decline. Compared with 2015, average reading performance fell by 28 points, a larger decline than either science or mathematics over the same period.
Perhaps more interestingly, the OECD highlighted declines in skills such as evaluating information, connecting information across different sources, and thinking critically about what we read. Those happen to be exactly the skills we may need most in the age of AI.
Reading Is No Longer Just About Finding Information
For most of human history, information was scarce. Books were expensive. Libraries were limited by geography. Academic knowledge could be difficult to access outside universities. Even in the early internet era, finding the right information often required knowing where to look.
Being knowledgeable therefore depended partly on how much information you could acquire. That problem has largely flipped.
Today, the average person has access to far more information than they could possibly consume in a lifetime. The challenge is no longer finding information. It is deciding what deserves our attention.
Search for almost any health question, political issue, historical event, or technology topic and you can find thousands of competing explanations. Some are excellent. Some are misleading. Some are technically accurate but missing important context. Others are complete nonsense presented with impressive confidence.
AI adds another layer to this. Instead of searching through ten websites ourselves, we can simply ask a chatbot to digest them and give us an answer. That is incredibly useful. But it changes what reading is for. When information was scarce, reading helped us acquire information. When information is abundant, reading increasingly helps us filter it.
Definitely related: Reading and Listening: Are We Losing These Skills?
The Internet Changed How We Read
The internet didn’t eliminate reading. If anything, many of us probably read more words every day than previous generations did. Messages, emails, social media posts, articles, comments, subtitles, search results, notifications and endless group chats all require reading.
But reading a lot of words isn’t necessarily the same thing as reading deeply. Digital environments reward speed. We scan headlines. We skim paragraphs. We jump between tabs. We search for keywords inside long documents. We scroll until something catches our attention.
None of these behaviours are inherently bad. Skimming is useful when you are trying to find something quickly. The problem appears when skimming becomes our default way of processing everything.
PISA 2025 offers an interesting clue here. The OECD reported that the proportion of what it calls “hasty readers” — students who rush through texts and provide quick but incorrect answers — almost doubled between 2018 and 2025.
At the same time, the report notes that increasing digitalisation, longer screen time and declining rates of reading for enjoyment have coincided with weaker literacy outcomes. That does not prove that smartphones or social media caused reading scores to fall. Education outcomes are influenced by countless factors, and the decline in some areas predates the COVID-19 pandemic.
But the broader pattern is difficult to ignore. Our information environment increasingly rewards consuming more things rather than understanding fewer things deeply.
AI Makes Reading Easier — and Potentially More Important
Generative AI pushes this trend much further. Suppose you encounter a 40-page research paper. Previously, you might read the abstract, introduction and conclusion before digging into the sections that matter to you. Now you can upload the entire document to an AI system and ask: “Summarize this for me.”
Seconds later, you have the major arguments. That is genuinely useful. There is little reason to romanticize spending an hour extracting information that software can accurately summarize in thirty seconds. But there is a catch. How do you know whether the summary is good?
Maybe the AI misunderstood an important qualification. Maybe it flattened a complicated argument into something much stronger than the researchers actually claimed. Maybe the original study found correlation while the summary quietly made it sound causal. Or perhaps the AI is completely correct.
The problem is that judging the answer still requires knowledge. The easier it becomes to generate an explanation, the more valuable the ability to evaluate that explanation becomes. This creates a strange paradox. AI reduces the amount of reading required to obtain information while potentially increasing the importance of reading comprehension itself.
You may read less. But what you do read has to be understood well enough to tell whether the machine is misleading you.
Could be related: Students Who Use AI Every Day May Be Falling Behind
PISA Shows That Technology Isn’t Simply Good or Bad
This is also why arguments about whether technology is “good” or “bad” for education tend to miss the interesting part. PISA 2025 found a much more complicated relationship.
Limited or moderate use of digital devices for learning at school is often associated with better outcomes. But when digital devices are heavily used for leisure at school, performance tends to be lower.
Digital distraction is also widespread. Across OECD countries, 28% of students reported being distracted by other students using digital devices during most or every science lesson. AI follows a similarly messy pattern.
Students who use AI for specific schoolwork tasks do not automatically perform better than those who don’t. Meanwhile, students who frequently use AI alongside opportunities to develop AI literacy tend to perform somewhat better.
In other words, access to the technology itself does not appear to be the magic ingredient. Knowing how to use it matters. This distinction will probably become even more important as AI gets better.
AI Can Give You an Answer Without Giving You Understanding
There is another difference between AI and older information technologies.
Google usually made you choose. Search for a question and you receive a list of websites. You have to decide which result looks credible, open it, read it and perhaps compare it with another source.
AI can remove much of that friction. Ask a question and you receive a clean, coherent answer immediately. That convenience is one of its greatest strengths. It is also why AI can create an illusion of understanding surprisingly easily.
An explanation can feel obvious while you are reading it. Every sentence makes sense. The examples seem clear. You reach the end thinking you understand the subject. Then someone asks you to explain it without looking at the answer. Suddenly, things become considerably less obvious.
Reading comprehension isn’t simply the ability to recognize sentences that make sense. It involves building a mental model of what those sentences mean and how the ideas connect. AI can construct the explanation for us. It cannot guarantee that the model has been constructed inside our heads.
Critical Reading Becomes a Form of Quality Control
This matters beyond education. Consider how much information modern adults encounter about investing, medicine, politics, science, technology and current events.
AI will increasingly sit between us and much of that information. It will summarize reports. Explain legal documents. Compare products. Interpret statistics. Condense meetings. Translate research. Perhaps eventually it will become the primary interface through which many people interact with the internet.
That makes critical reading less about remembering information and more about quality control.
- Does this conclusion actually follow from the evidence?
- Is an important caveat missing?
- Is the source credible?
- Is the language more certain than the underlying data?
- Are two things merely correlated, or does one actually cause the other?
Those questions require something AI cannot conveniently outsource for us, because outsourcing the evaluation creates another question: who evaluates the evaluator? At some point, a human still needs enough understanding to decide whether an answer deserves to be trusted.
The Most Valuable Reader May Not Be the Fastest One
There is an understandable temptation to respond to all of this by rejecting new technology. Maybe students should stop using AI. Maybe everyone should put their phones away and return to paper books. But that probably misunderstands the problem too.
Calculators didn’t make mathematics useless. Search engines didn’t make knowledge useless. Spell-checkers didn’t eliminate the value of writing. Technology tends to automate parts of a skill rather than eliminate the entire skill.
AI may automate enormous amounts of information retrieval and processing. It may even make memorizing many facts less economically valuable. But the skills surrounding that automation can become more important. Reading carefully. Comparing claims. Recognizing uncertainty. Following an argument. Detecting contradictions. Understanding context.
These are not particularly flashy skills. They certainly don’t sound as futuristic as prompt engineering or AI literacy. Yet AI literacy without ordinary literacy would be rather strange.
If machines increasingly handle the work of producing answers, humans may need to become better at judging them. That may ultimately be the irony behind the PISA 2025 results. Reading performance is declining at precisely the moment when technology appears capable of doing more of our reading for us.
It is tempting to conclude that this makes reading less important. The opposite may be true.
In a world where information was difficult to obtain, the advantage belonged to people who could find the answer. In a world where everyone can generate an answer in seconds, the advantage may belong to people who can tell whether the answer is any good.
