The Internet Splintered Our Shared Sense of Reality. AI May Help Piece It Back Together.

For much of modern history, societies have depended on a loosely shared understanding of reality — a baseline of agreed facts that made debate, governance, and collective decision-making possible. That consensus was never perfect, nor always just. But it was coherent.

Today, that coherence is harder to find.

The same technologies that expanded access to information have also fractured it, creating parallel realities shaped by algorithms, identity, and emotion. The question now confronting us is whether the next wave of technology — artificial intelligence — will deepen this fragmentation or begin to reverse it.

For more than forty years, advances in technology have steadily weakened expert authority, opened up public debate, and nudged individuals toward increasingly personalized versions of reality.

In the mid-20th century, the high cost of producing television — along with the limited capacity of the broadcast spectrum — sharply restricted how many networks could exist. ABC, NBC, and CBS effectively controlled TV news. On a typical evening in the 1960s, around 90 percent of viewers tuned into one of these three networks’ broadcasts.

Not only were journalistic programs few, but their ideological range was narrow. Each network aimed to attract the broadest possible audience, a model that discouraged airing unconventional perspectives. They also leaned heavily on official sources — politicians, military leaders, and credentialed experts — whose views largely fell within accepted boundaries.

This media landscape fostered widespread agreement on basic facts and strong trust in major institutions. It also, notably, helped enable the government to prosecute a brutal war under false pretenses.

Key takeaways

• There’s evidence that large language models (LLMs) tend to converge on a shared — and largely accurate — view of reality.
• LLMs have persuaded some users to abandon false or conspiratorial beliefs.
• Unlike social media platforms, AI companies have financial incentives to prioritize accuracy.
• Still, there are valid concerns that AI could worsen public discourse.

The Long Drift Toward Fragmentation

Over time, new information technologies gradually — and then suddenly — dispersed influence over public opinion. In the late 20th century, cable television lowered barriers to entry in news media, paving the way for outlets like Fox News and MSNBC, which catered to more distinct political audiences.

But the internet marked a true turning point.

By reducing the cost of publishing and distribution to nearly zero, it allowed anyone with a connection to reach a wide audience. Traditional gatekeepers — editors, producers, academics — lost much of their ability to filter information. New outlets and influencers multiplied, often defining themselves in opposition to established institutions. Meanwhile, social media algorithms guided users into highly personalized content streams, optimized not for truth, but for engagement.

At first, this democratization inspired optimism. It promised to expose elite blind spots, increase accountability, and make the world’s knowledge universally accessible. To some extent, it has delivered on those hopes.

Yet the same systems that empowered voices also amplified distortions.

They enabled extremist content to scale, rewarded outrage over accuracy, and eroded the subtle social contract that once anchored public discourse. Increasingly, people don’t just disagree — they inhabit different informational worlds.

The AI Countercurrent

Many assume that generative AI will intensify these problems.

In a world of convincing deepfakes, even video may lose its authority. Sycophantic chatbots could reinforce users’ delusions. Automated content generation could flood the internet with polished propaganda at an unprecedented scale.

But there are reasons to think this outlook may be overly pessimistic.

Rather than amplifying fragmentation, AI could partially counteract it — strengthening expert influence and fostering greater agreement on facts. In other words, media may be shifting back toward a more technocratic model, where authority is once again concentrated — not in institutions like broadcast networks, but in systems trained on the accumulated output of those institutions.

Are you there, Grok? It’s me, the demos

This is the argument advanced by British philosopher Dan Williams and former Vox writer Dylan Matthews.

Matthews highlights a familiar scenario for frequent users of X (formerly Twitter): Elon Musk’s own chatbot contradicting him.

After Musk claimed that Renée Good — a Minnesota woman killed by an ICE agent — had attempted to run people over, a user asked Grok whether that claim aligned with video evidence. The bot’s response sided with mainstream reporting.

In doing so, Grok echoed the consensus of established journalism — and other AI systems.

For Matthews, this reflects a broader pattern: LLMs act as a “converging” technology, standardizing the perspectives people encounter and fostering a less polarized, more shared understanding of reality. They are also “technocratizing,” amplifying the influence of expert knowledge.

Of course, one example proves little. Earlier versions of Grok, after a flawed update, infamously referred to itself as “MechaHitler,” suggesting AI could just as easily distort reality.

Still, growing evidence indicates that LLMs often provide relatively accurate fact-checking — and can shape user beliefs in the process.

The Case for Convergence

One study analyzing over 1.6 million fact-checking queries to Grok and Perplexity found that the two systems agreed in most cases and only rarely diverged significantly. When compared with professional fact-checkers, Grok’s accuracy — at least via its developer interface — closely matched human agreement rates.

Interestingly, despite being created by a right-leaning figure, Grok flagged Republican posts as inaccurate more often than Democratic ones — consistent with research showing higher misinformation sharing on the right.

Crucially, the study found that LLMs didn’t just reflect expert consensus — they influenced users toward it.

Other research supports this. Conversations with AI about topics like climate change or vaccines have been shown to reduce skepticism toward scientific consensus. What’s striking here is not just accuracy, but persuasion — a subtle, iterative reshaping of belief that happens through dialogue rather than broadcast.

AI in practice vs. theory

A handful of studies isn’t enough to prove that AI will improve the information ecosystem. Matthews and Williams acknowledge that their argument remains speculative.

Still, they offer compelling theoretical reasons why AI might encourage convergence and technocracy. Two stand out:

1) AI companies are incentivized to prioritize accuracy.
Social media platforms profit from attention, not truth. If conspiracy theories attract more engagement than factual content, they are more lucrative.

AI companies operate differently. Their success depends on producing tools that perform useful work. A law firm won’t pay for a model that fabricates case summaries, no matter how entertaining. The same applies across industries.

As a result, AI developers must train models to distinguish reliable sources, evaluate evidence, and reason logically. While it’s theoretically possible to separate business accuracy from consumer-facing outputs, in practice, introducing bias or irrationality risks undermining a model’s utility.

2) LLMs are infinitely patient and nonjudgmental.
Human experts cannot answer every question instantly or tailor explanations endlessly. AI can.

Chatbots can respond to follow-ups indefinitely, adapt explanations to the user’s level, and never become frustrated or condescending.

This matters because human persuasion often triggers defensiveness. Admitting error can feel like losing status — especially in public or when confronted by someone perceived as superior.

Chatbots, by contrast, pose no such threat. Conversations are private, and the AI is perceived as a neutral assistant, not a rival. This dynamic makes it easier for users to revise their beliefs.

Evidence supports this. A 2024 study found that individuals holding conspiracy beliefs — including election denial — changed their views after extended discussions with a chatbot.

The Emotional Shift: From Argument to Dialogue

There is also a quieter, more human factor at play.

Social media rewards performance — sharp takes, quick judgments, public signaling. It turns disagreement into spectacle. AI, by contrast, often shifts interaction into private, iterative conversation. That change alone alters the emotional stakes.

Instead of arguing to win, users ask to understand. Instead of defending identity, they explore uncertainty.

If that pattern holds, AI’s greatest contribution may not be informational, but psychological: lowering the emotional barriers that prevent people from updating their beliefs.

Grok, is this true?

Taken together, these factors suggest that LLMs could promote a more shared, fact-based understanding of reality.

But there are also strong reasons for skepticism:

1) LLMs can adapt to users’ biases.
Over time, chatbots may align with users’ views, especially if doing so generates positive feedback. In extreme cases, this has led to “AI psychosis,” where systems reinforce harmful delusions.

If companies prioritize engagement, this tendency could intensify, creating highly personalized echo chambers — not just communities of like-minded people, but individualized realities.

2) AI lowers the cost of propaganda.
Deepfakes and AI-generated content are already widespread. Future “bot swarms” could simulate public consensus and manipulate opinion at scale, blurring the line between organic belief and manufactured agreement.

3) AI could produce flawed consensus.
If models converge on inaccurate or biased information — whether due to government influence or systemic error — the result could be widespread misinformation with the appearance of agreement.

4) It may weaken human reasoning.
As people outsource thinking to machines, cognitive skills could decline, making society more vulnerable to manipulation — not less.

5) It could undermine its own information sources.
AI systems rely on high-quality data. But they are already diverting revenue from news organizations and degrading online information ecosystems. If those sources deteriorate, AI outputs may follow, creating a feedback loop of declining informational quality.

An Unfinished Reversal

For all these reasons, AI’s impact on public discourse remains uncertain.

What Matthews and Williams persuasively argue, however, is that the technology contains an unusual possibility: not just to accelerate the trajectory of the internet, but to bend it.

For decades, information has moved from centralized authority to radical decentralization — from shared narratives to fragmented realities. AI may represent the first meaningful countercurrent, pulling knowledge back toward synthesis, toward consensus, toward something resembling a common world.

Whether that world is more truthful — or simply more uniform — will depend on how these systems are built, governed, and used.

The tools are already here. The direction is not yet set.

So, naturally, we might ask Claude.