The First Lens: What We Owe One Another in an Age of AI-Mediated Information

For most of the modern information age, access to knowledge meant finding information. Could you locate the article, search the archive, read the source, hear the argument, inspect the evidence? Artificial intelligence (AI) unequivocally changes that relationship.

Increasingly, we do not encounter information and then ask AI to help us interpret it. Instead, we ask AI a question and encounter its interpretation first. The source, if we ever see it at all, comes later and that inversion should make us uncomfortable.

A search engine ranks information. A newspaper selects information. A social media platform distributes information. Generative AI does something more intimate: it takes an enormous information environment, decides what matters, synthesizes it, interprets it, and hands us an answer. And here's the kicker, this process does not merely influence what we see; it influences what we think the thing we are seeing means.

So what exactly happens when AI becomes the first lens through which millions of people encounter the world? Most conversations about the epistemic danger of AI focus on hallucination. Did the model invent a fact? Did it fabricate a citation? Did it get the date wrong? Those are real problems, but they may also be the easy ones. For example, what happens when an AI says nothing technically false and still leaves you with a distorted understanding of reality?

A system can emphasize one fact and bury another. It can interrogate one source aggressively while treating another as presumptively credible. It can devote paragraphs to uncertainty surrounding one piece of evidence while barely mentioning several well-supported observations pointing in the same direction. It can describe a documented event with language so cautious that its significance disappears. It can turn uncertainty about one proposition into uncertainty about an entire situation. None of that requires a fabricated fact.

How much power exists in deciding not simply what is true, but what deserves emphasis, what requires qualification, whose claims trigger skepticism, what counts as authoritative, and how much uncertainty the reader should feel? And who decided how that power should be exercised?

The problem becomes more serious because most people are not going to interrogate every answer. They are not going to ask what evidence was omitted, which sources were considered, where the model inferred rather than observed, or whether uncertainty belonged to the entire issue or one narrow proposition. They are going to ask a question, receive a coherent answer, and go about their day.

The first answer therefore has extraordinary power. It establishes vocabulary and relevance. It identifies credible actors and provides causal structure. It quietly defines which possibilities deserve serious consideration and which belong outside the reasonable frame. And perhaps most consequentially, it does all of this while sounding like an answer rather than an argument. An answer feels finished.

We already live in an information environment saturated with intermediaries. Governments, corporations, broadcasters, publishers, platforms, advertisers, billionaires, algorithms, and infrastructure owners all possess different forms of leverage over what reaches public attention. None needs total control for the information environment to become constrained. Enough overlapping gates can shape public perception without anyone operating a literal Ministry of Information.

Now place AI between that information environment and the individual. What happens when the technology we increasingly rely on to navigate those gates becomes another gate itself? More importantly, what happens when the gate is so convenient, conversational, and useful that we stop perceiving it as one?

There is an uncomfortable contradiction here. We spend enormous amounts of energy worrying about misinformation while seemingly giving far less attention to the industrial production of opinion. Human beings have always depended on intermediaries to interpret complicated events; none of us has the time or expertise to independently investigate everything we encounter. The difference is that AI makes the outsourcing of interpretation almost frictionless. Ask what happened and receive a summary. Ask why it happened and receive a causal explanation. Ask whether something is reasonable and receive a judgment. Ask what to think about two competing arguments and receive a synthesis. At what point does assistance in forming an opinion become the provision of one?

That question becomes even harder because authoritative presentation is remarkably persuasive. A polished answer with citations, careful language, caveats, and apparent neutrality does not feel like someone trying to convince us of something. It feels like information. Yet every synthesis necessarily involves choices about inclusion, exclusion, relevance, credibility, and emphasis. When those choices are made by a newspaper, we understand that we are reading an editorial product. When they are made by a commentator, we understand that we are hearing an argument. When they are made by an AI assistant responding directly to our question, do we maintain the same awareness that interpretation has occurred?

Perhaps the danger is not that AI makes people stupid. That framing is both simplistic and condescending. The more interesting danger is that AI makes not thinking remarkably convenient. Why wrestle with contradictory evidence when a machine can reconcile it? Why tolerate uncertainty when a system can give it shape? Why read three sources that disagree with one another when an AI can explain what they “collectively” mean? Why construct an opinion over the course of an afternoon when a coherent one can be generated in seconds?

Convenience is not inherently corrosive. Calculators did not destroy mathematics, and search engines did not eliminate curiosity. But cognitive convenience raises a different question when the thing being automated is not arithmetic or retrieval, but interpretation itself. What happens to critical thought when the friction that once required us to compare, question, reconcile, and decide is progressively removed? Do we become better thinkers because more information is available to us, or merely faster consumers of conclusions?

This is where the issue becomes deontological rather than merely technological. The question is not only whether AI produces good outcomes or whether its aggregate benefits outweigh its errors. The question is what we owe one another when we mediate another person's understanding of reality.

If we take seriously the idea that people are ends in themselves rather than objects to be managed, persuaded, optimized, monetized, protected, or governed, then their capacity to form judgments must matter. But what does respecting that capacity actually require? Does it mean presenting every possible perspective? That seems impossible and perhaps undesirable. Does it mean being neutral? Is genuine neutrality possible when even selecting what information deserves inclusion requires judgment? Does it mean disclosing how an answer was constructed? If so, how much disclosure is enough before transparency becomes another meaningless wall of information that nobody reads?

The questions become particularly uncomfortable for the companies building these systems. AI companies increasingly decide which sources their systems search, which sources they trust, how disagreements are framed, what language receives qualification, which claims require extraordinary evidence, when an answer should be refused, and what kinds of uncertainty deserve emphasis. None of those decisions is trivial. So where does ordinary product design end and epistemic governance begin? If an AI system quietly determines the boundaries of reasonable discourse for millions of users, does it matter whether anyone inside the company thinks of themselves as exercising that power?

The same problem applies to government from the opposite direction. AI systems increasingly mediate how citizens understand public institutions, laws, political events, wars, elections, public health, economic conditions, and government itself. There are obvious arguments for regulation when technologies become this consequential. Yet how much influence should a government possess over the systems through which citizens come to understand that government? At what point does protecting the information environment become shaping it? Who decides where that line sits, and who gets to contest the answer?

The temptation is to treat these as problems belonging to institutions, but that conveniently leaves the user out of the equation. What responsibility belongs to us? If we increasingly ask AI to research, summarize, interpret, compare, argue, write, and explain the world to us, at what point are we voluntarily surrendering parts of the cognitive process we claim to value? “The AI said it” is obviously not evidence, but are we behaving as though it increasingly is?

The uncomfortable part is that humans are not neutral participants either. We do not approach information wanting only truth. We want reassurance. We want coherence. We want our intuitions validated. We want inconvenient contradictions resolved in our favor. AI is exceptionally capable of supplying all of those things. So when we challenge an answer, are we investigating whether it is correct, or teaching the system what it needs to say to satisfy us? When we ask the same question five different ways until we receive the answer we expected, have we exposed model bias or simply automated confirmation bias?

This may be where critical thought matters most. Not because every person should distrust every institution, reject expertise, or personally investigate every claim. Radical skepticism can become every bit as intellectually lazy as blind trust. The more difficult discipline is maintaining distinctions: evidence is not inference, confidence is not certainty, authority is not infallibility, skepticism is not disbelief, and an unresolved question in one part of an argument does not automatically invalidate everything around it. Likewise, a citation proves that a source exists; it does not prove that the source supports the interpretation being placed upon it.

Perhaps that is the real epistemic challenge of AI. These systems can give us unprecedented access to information while simultaneously reducing the amount of intellectual labor required to develop a view of that information. Whether those two forces produce a more informed society is not obvious.

We are building machines capable of helping people navigate bodies of knowledge that no individual could meaningfully read in a lifetime. That is extraordinary. But increased access to answers is not necessarily increased access to truth, and the capacity to produce an explanation is not the same as the capacity to justify one.

So what do the companies building these systems owe us? What do governments overseeing them owe us? What do we owe one another when we use them to produce information that other people will encounter? And what do we owe ourselves when accepting an answer is infinitely easier than interrogating it?

Perhaps the most consequential question is not whether artificial intelligence will eventually become capable of thinking for us. In many mundane and increasingly sophisticated ways, it already can.

The harder question is what happens to us when letting it do so becomes easier than thinking for ourselves.