The Public Web Is Getting Harder to Trust

The public web still matters. It still works. It is still the fastest way to get oriented, compare sources, and surface useful information.

What has changed is something more subtle and more important: the cost of verification.

The core problem is not simply that AI-generated content exists. The bigger issue is that more of the web’s information layer now sits between the user and the original source. AI-mediated search rewrites and summarizes on the user’s behalf. Search results compete with scaled unoriginal content. Syndicated and copied pages can outrank or out-circulate original reporting.

The result is not a dead web. It is a lower-trust, higher-friction research environment.

That distinction matters. This is not an argument for panic, nostalgia, or blanket hostility to AI. It is an argument for a more disciplined standard of evidence. Organizations that keep researching like it is still 2019 will make more avoidable mistakes. Organizations that adapt will keep the advantage.

The problem is not information scarcity. It is source separation.

There is no shortage of information online. If anything, there is more available text than ever. But abundance and reliability are not the same thing.

The old mental model of web research was fairly simple: search, scan the results, open a few sources, compare them, and work back to the original evidence if the claim mattered. That workflow was imperfect, but it made verification relatively natural.

Now, more of the web experience is designed to prevent that extra step. Generative interfaces answer first. Summaries compress context. Citations appear authoritative even when they are incomplete, broken, or attached to the wrong source. Users are nudged toward consuming conclusions instead of inspecting evidence.

That is why the real issue is source separation. The question is no longer just whether an answer looks plausible. It is whether the user is still close enough to the original source to verify it.

AI search is making provenance harder, not easier.

One of the strongest pieces of evidence here comes from the Tow Center for Digital Journalism at Columbia, which tested eight generative search tools across 1,600 queries. Its conclusion was hard to ignore: the tools collectively returned incorrect answers to more than 60% of the queries tested.

These failures were not minor formatting issues. They included:

• fabricated links
• broken URLs
• wrong attributions
• citations pointing to copied or syndicated versions rather than the original source

That matters because provenance is not a cosmetic detail. In a serious research environment, the source trail is part of the answer. If a system gives a confident summary while quietly damaging the citation chain, it makes the user faster only in the least useful sense. It speeds them toward an answer while making it harder to know whether the answer deserves trust.

This is where a lot of casual commentary misses the point. The danger is not just hallucination. The danger is authoritative-looking source confusion. A polished wrong answer with a damaged citation path is more dangerous than a blank screen because it lowers the odds that a user will stop and check.

Users are verifying less.

The behavioral evidence points in the same direction.

A Pew Research Center analysis found that when Google showed an AI summary, users clicked a traditional search result link only 8% of the time (compared to 15% when no summary appeared). Users clicked links inside the AI summary itself in only 1% of visits. Sessions also ended more often when an AI summary was present.

That does not prove every AI summary is harmful. It does show that the old habit loop — query, open sources, compare sources, inspect the original — is weakening.

And that matters because verification is often less a conscious principle than a workflow side effect. When users naturally click through, they expose themselves to source context, bylines, dates, caveats, and competing framings. When they stay inside the summary layer, they get less of that. Even a good summary reduces friction partly by removing steps. The problem is that some of those steps were where trust got earned.

Search platforms signal that scale is a quality problem.

The trust issue is not just an abstract fear about the future of the internet. The platforms themselves are responding to it.

In March 2024, Google announced updated spam policies targeting expired-domain abuse, scaled content abuse, and site reputation abuse. The company explicitly defined scaled content abuse as large amounts of unoriginal content produced primarily to manipulate rankings, whether by automation or by humans.

The signal here is not that AI content is inherently spam. It is that large-scale low-value content pressure became serious enough for the dominant search platform to create sharper policy categories around it.

In other words, even Google is telling the market that origin, incentives, and originality matter. That should temper two bad instincts at once: the naive belief that all machine-assisted content is fine, and the simplistic belief that the tool itself is the whole problem. The real issue is lower-value material being produced at scale and injected into discovery systems that many users still treat as neutral.

People already know the environment is strained.

This is not just a platform problem. It is a public perception problem too.

Reuters Institute’s 2025 Digital News Report found that 58% of respondents remain concerned about their ability to tell what is true from what is false online. It also found that when people want to verify whether something is true, they still lean heavily on trusted news brands and official government sources.

That behavior is revealing. It suggests users already understand, at least intuitively, that the web’s default information layer has become noisier. They may not describe the issue as provenance failure or citation drift, but their behavior points in the same direction: when the stakes rise, they look for stronger anchors.

That is also why this article should not end in despair. A harder-to-trust web is not the same as an unusable web. Trusted institutions, original documents, and official sources still exist. The challenge is that getting to them now takes more intention.

Synthetic content is now common enough to weaken old assumptions.

Another shift is that public-facing writing itself is becoming harder to interpret at a glance.

An empirical study hosted by NIH’s PubMed Central found substantial signs of LLM-assisted writing across multiple domains by late 2024, including corporate press releases, financial consumer complaints, job postings, and UN press releases. That does not support the lazy claim that “most of the web is AI-generated.” It does support a narrower and more useful conclusion: naive assumptions about authorship and provenance are weaker than they used to be.

For researchers, that means style is no longer a reliable proxy for credibility. Smooth writing does not tell you much. Formal tone does not tell you much. Even apparent institutional polish does not tell you enough. What matters is the chain underneath the text: who published it, what it cites, whether it points back to original evidence, and whether that evidence survives contact with scrutiny.

The counterargument: The web still works.

A credibility-first argument has to include the counterevidence.

Google has argued that AI in search is driving more queries and higher-quality clicks, and that overall organic click volume remains relatively stable. Pew’s own data also shows that AI summaries did not appear on every search, only on 18% of the searches in its dataset. And Google is correct that AI use alone is not evidence of spam or low quality.

Those caveats matter. They improve the argument rather than weaken it.

The right conclusion is not that the web has collapsed. The right conclusion is that reliable web research now depends more heavily on the user’s method. The web still works for disciplined people. What has degraded is the safety margin for lazy research.

Verification used to be something competent users often did almost by default. Now it is increasingly something they must choose to do on purpose.

What serious researchers should do differently now

The practical response is not to stop using the public web. It is to use it with better operating rules.

1. Prefer original documents over summaries. If a system gives you a neat answer, treat that as orientation, not proof.
2. Verify the URL itself, not just the quoted sentence. A citation that sounds right is not enough if it points to a copied page, a broken link, or a syndicated version stripped of context.
3. Distinguish between discovery tools and evidence sources. AI systems can be useful for routing, framing, and narrowing the field. They are far weaker as final authorities.
4. Privilege trusted institutions and official sources. That does not mean blind trust. It means starting from places with stronger incentives, clearer accountability, and better provenance.
5. Triangulate non-obvious claims. If a fact matters, look for at least two independent sources and check whether either of them points back to a primary source.
6. Actively test counterevidence. The fastest way to get fooled online is to search only for support. The strongest researchers now are the ones who can find the best argument against their own conclusion before someone else does.

The new advantage is verification, not speed.

For years, the web rewarded speed. The person who could find information fastest often won.

That is no longer enough.

In a world of AI summaries, scaled unoriginal content, weak attribution, and increasingly blurred source chains, the better advantage is not retrieving information quickly. It is verifying it correctly.

The public web is still indispensable. But it is less trustworthy by default than many users assume. That means the organizations that keep their standards high — checking provenance, preferring originals, and resisting the temptation to stop at the first smooth answer — will outperform the ones that confuse convenience with truth.

That is not a crisis. It is a new operating condition.

And the teams that understand it early will make fewer expensive mistakes.

Selected sources

• Tow Center for Digital Journalism / Columbia Journalism Review, AI Search Has a Citation Problem
• Pew Research Center, Google users are less likely to click links when an AI summary appears in results
• Google Search Central, March 2024 core update and new spam policies
• Reuters Institute, Digital News Report 2025 executive summary
• NIH PubMed Central, The widespread adoption of large language model-assisted writing across society