Information Has Never Been This Easy to Find, Nor This Unreliable

Jake Moffatt’s mother passed away. He needed to fly from Vancouver back to Toronto for the funeral.

He logged into Air Canada’s website, opened the online customer service chat window, and asked a simple question: Is there a bereavement fare discount policy?

The chatbot gave him a clear, complete, and seemingly entirely reasonable answer: First, purchase a full-price ticket, then submit a refund application within 90 days after travel to receive a reimbursement for the bereavement fare difference.

Moffatt bought the ticket, flew, handled the funeral, and submitted the refund application within the specified timeframe. Then it was rejected. Air Canada told him: Our bereavement discount must be requested before ticket issuance; it cannot be applied retroactively after issuance. What the chatbot said? AI-generated, does not represent company policy.

In February 2024, the court ruled against Air Canada, awarding a few hundred Canadian dollars in compensation. But the real cost wasn’t the money—it was the afternoon Moffatt spent arguing with customer service during his most vulnerable time after his mother’s death. He did what any normal person would do: asked a question, received a seemingly official answer, and believed it.

This is a microcosm. You ask AI tools dozens of questions every day, receiving dozens of fluent, confident answers. How many are true? How do you know?

Every “Solution to Information Overload” Digs a Deeper Hole

The history of information channels is not a gradual curve of progress. It’s more like an earthquake—long periods of accumulation, sudden fractures, and new landscapes rebuilt on the ruins.

The first fracture occurred in the mid-1990s. The internet transformed information access from “institutional monopoly” to “accessible to everyone.” Before this, ordinary people had extremely limited channels for non-local information: newspapers, radio, television, libraries. A common feature of these channels was the presence of a gatekeeper—editors, journalists, librarians—acting as a filter layer between you and the information. Gatekeepers limited the volume of information you could access while ensuring a basic quality baseline.

Search engines solved the “can’t find it” problem but created “found too much.” When Google launched in 1998, PageRank was essentially an automated trust assessment—the more high-quality pages linked to a site, the more credible it was. This worked effectively on the early, smaller-scale internet. After SEO industrialization, links became commodities that could be bought and sold, and the trust foundation was hollowed out.

The second fracture occurred between 2006 and 2013. Social media transformed information distribution from “you go looking” to “it comes to you.” Facebook News Feed, Twitter timelines, WeChat Moments—information appeared automatically before your eyes through algorithms and social connections.

What died in this transition was RSS. RSS gave you complete control over what you saw: each website provided a content feed, and you subscribed to the feeds you were interested in using a reader. No algorithmic intervention, no ad insertion. Google Reader peaked with tens of millions of users. Google shut it down in 2013. As The Verge put it: “Google first conquered RSS, then abandoned it.” Users didn’t return to other RSS readers; they returned to the embrace of algorithmic feeds. The active right to choose simply vanished.

The third fracture occurred between 2022 and 2024. The release of ChatGPT marked the shift in information access from “you sift through multiple sources” to “a model synthesizes one answer for you.” From launch to capturing 17% of digital query market share, it took less than three years.

Each fracture followed the same pattern: new technology solved the pain points of the previous generation while digging a deeper hole. Search engines dug “SEO pollution”; social media dug “filter bubbles”; AI aggregation dug “unverifiability.” And each time, the hole became harder to fill. SEO pollution is at least traceable; filter bubbles are perceptible; AI hallucinations are imperceptible—incorrect and correct answers look identical on the surface.

Over thirty years, the technology for information access has advanced ten-thousandfold, while the methodology for information verification has barely moved.

Findable, but Not Trustworthy

Twenty years ago, the problem plaguing ordinary people was: I want to know something, where do I look? Today’s problem is: I’ve found ten answers, which one is true?

The tipping point arrived abruptly. Ahrefs sampled nearly a million new webpages in April 2025, finding that 74.2% contained detectable AI-generated content. Graphite researched over 60,000 English articles, discovering that by the end of 2024, more than half were completed by large language models. Europol’s prediction is even more aggressive—by 2026, 90% of online content could be synthetically generated.

These numbers mean that the old method of “cross-verification” is failing. Verifying a piece of information used to be simple: if three independent sources said it, it was probably true. The premise was that each source had an independent collection process—someone went to the scene, someone made a call. But when the content of three websites all comes from the output of the same model, “multiple sources” becomes “multiple copies.” You think you’re doing triangulation, but you’re actually looking at three angles of the same mirror.

Google has felt the pressure. In March 2024, it rolled out a massive core update aimed at reducing 40% of low-quality, useless content—for the first time in the search engine’s thirty-year history, reducing the index became more urgent than expanding it. The effect was limited. Google’s traffic share fell to 28.1% in April 2025, a year-over-year drop of nearly 10 percentage points. Users are voting with their feet, turning to AI search.

But AI search itself is unreliable. The Columbia Journalism School’s Tow Center systematically tested 8 mainstream AI search tools: selecting 200 articles from 20 publishers, extracting original text fragments and asking “which article does this passage come from?” Using traditional Google Search for the same fragments, the original source could be found within the top three results. The collective performance of the AI tools: over 60% of queries returned incorrect answers.

Even more ironic is the paid versions. Perplexity Pro and Grok 3 have slightly higher accuracy rates than their free counterparts, but also higher error rates—because they almost never say “I don’t know.” ChatGPT incorrectly identified 134 articles in 200 queries, expressed uncertainty only 15 times, and never once refused to answer.

This is the structural contradiction: the machine that produces information and the machine that verifies information are the same machine. The marginal cost of generating AI articles approaches zero, while verifying whether an article is credible still requires human effort, time, and professional judgment. The production side is exponential; the verification side is linear. When the gap widens to a critical point, the default trust in the entire information ecosystem will flip—from “published is credible unless proven false” to “all information is not credible unless proven true.”

Five Paths, Each with Pitfalls

Facing this situation, the tech industry and content sector are pursuing five paths.

Conversational Retrieval. The promise of Perplexity, ChatGPT Search, etc.: No more scrolling through ten pages of search results; just ask a question and get an integrated answer. Like switching from grocery shopping to food delivery—it saves time, but you don’t know how the delivery person selected the items or if they skipped certain aisles. ChatGPT search market share is about 17%, and AI platforms overall are growing 3-5 times faster than Google. ChatGPT directs users to external websites at a rate even more than double that of Google. But 60% of its news citations are incorrect.

Proactive Agent Push. An evolved version of conversational retrieval—it’s not you asking it, but it coming to you. AI assistants push information to you before you need it, based on your interests and schedule. Like the return of magazine subscriptions, but with algorithms replacing editors. Traditional magazine editors face reputational pressure; the feedback loop for AI agents optimizes for “clicks/dwell time,” not “accuracy/usefulness.” If you dwell on a piece of fake news for 30 seconds (due to shock), the system may push more similar content. Filter bubbles in the agent era will be more insidious—information appears in the guise of a “prepared daily briefing for you,” looking like an objective intelligence summary, but actually filtered and processed in layers.

Decentralized Verification. The C2PA standard adds a “birth certificate” to digital content—recording who created it, with what tool, when, and whether it was modified. Similar to food traceability chains. Over 5,000 institutions have joined, with the US Department of Defense and the Library of Congress starting to advance it in 2025. The technology is mature; the difficulty lies in adoption—it requires full-chain support from production to distribution to consumption. Photos taken on phones without C2PA modules are like food without traceability codes. Moreover, C2PA only addresses “where it comes from,” not “whether it’s correct.”

Return of Community Curation. Substack paid subscriptions grew from 2 million to 5 million in two years, with over 50 creators earning over a million dollars annually. The logic: when machines are unreliable, people return to trusting specific individuals. Subscribing to an expert’s newsletter is saying “I trust your judgment.” The limitations are obvious: it doesn’t scale. The entire Substack paid user base might not exceed the population of a second-tier city. It’s essentially an elite path—requiring the ability to identify who is trustworthy, the ability to pay, and the time to digest in-depth content.

Multimodal Native Content. Podcasts, long-form videos, live streams—the current cost for AI to synthesize these is much higher than for text. Faking an article requires only one API call; faking a two-hour natural conversation podcast (with hesitations, interruptions, tangents) remains difficult. But the window is closing. Moreover, multimodal content cannot be quickly scanned—you can’t judge whether a podcast is worth listening to in thirty seconds like you can with an article.

All five paths share a common structural problem: they all attempt to solve trust “before information reaches you.” Conversational retrieval helps you filter, Agents help you choose, C2PA helps you verify, community curation helps you find reliable people, and multimodal content raises the barrier for fakes. But none of these paths truly eliminate the verification cost—either transferring it (to AI, platforms, or people you trust) or delaying it (until technology becomes widespread).

Information Is Becoming a Luxury Good Again

The technological narrative tells a story: AI has democratized information. Reality tells the opposite story.

The information landscape of 2025 is splitting into two markets. The first: free, massive, AI-generated content pool, wide-ranging, low reliability. The second: paid, limited, human-expert-curated content pool, narrow-ranging, high reliability. Bloomberg Terminal’s annual fee of $25,000 isn’t selling information itself (the same data is available elsewhere); it’s selling “verified, real-time updated, source traceable” information.

In the future, what you pay for won’t be exclusive content, but verification itself.

The nature of the information divide is changing. It used to be at the access level: some people could get online, some couldn’t. Mobile internet gave 5 billion people smartphones worldwide, significantly narrowing this gap. The new divide is at the verification level: everyone can access information, but only some people have the ability (economic, cognitive, temporal) to verify reliability. The urban white-collar worker paying $20 a month for three professional newsletters and the average user relying on free AI search inhabit two completely different quality worlds of information.

It’s like being in a market full of counterfeit money—having a lot of cash doesn’t mean you’re rich; it depends on whether you can distinguish which bills are real.

What to Do: Start with Four Questions

If you wake up tomorrow facing the same information environment, what can you do immediately?

Ask yourself four questions when acquiring any important information:

What is the timeliness requirement? For low-risk, high-timeliness information (what time is the meeting, tomorrow’s weather), AI tools are sufficient. For information used to make important decisions (medication side effects, company finances, historical facts), additional verification steps are needed.

Is there a traceable primary source? Click the links attached to AI answers to see. If they point to another AI summary, an aggregator site, or are inaccessible, credibility is reduced. The standard for a primary source: someone who went to the scene, conducted the experiment, reviewed the archives, examined the reports.

Are multiple sources truly “independent”? If three websites provide the same answer, check if the wording is highly similar. Similarity suggests a common source. Truly independent sources have different phrasing, different emphases, even different minor errors.

What is the incentive structure of the information channel? Platforms that rely on ads tend to produce click-baiting content; authors who rely on subscriptions tend to protect their reputation. The incentive structure determines the ceiling of information quality.

For long-term habits: create a shortlist of 3-5 verified sources for each area you care about; use AI as a starting point (helping you draw the map) not an endpoint (making decisions for you); pay for key information—the fact that someone bears the verification cost is itself a valuable service.

In the next three years, information access channels won’t converge into one. They will stabilize into a three-tier structure: the bottom layer is AI aggregation, solving “what exists,” with the widest coverage and lowest reliability; the middle layer is community curation and professional publications, solving “what to trust”; the top layer is primary data sources and content authentication systems, solving “provable,” with the narrowest coverage and highest reliability.

Most daily needs will move between the bottom and middle layers. For important decisions—health, investments, law, academia—you need to move up one or two layers. Knowing when to move up is the core of information literacy in this era.

If Moffatt had spent an extra five minutes that day, opening Air Canada’s official policy page to read it himself, he wouldn’t have been misled by the chatbot’s hallucination. Five minutes, not long. But he didn’t know he needed to spend those five minutes.

This is perhaps the most unsettling part of the whole affair: you don’t know what you don’t know.