Everyone is worried about the wrong AI failure.

For two years, the loudest AI safety conversation has centred on hallucination: the model that invents a fact, a court case, a source that never existed. That risk is real and well documented.

But a quieter problem is spreading underneath it, and it may end up mattering more. AI answers can be completely accurate and still be impossible to check. Increasingly, that isn’t an edge case. It is the default experience.

The click is disappearing

Zero-click searches, where a person gets an answer and never visits a website, reached 68 per cent in early 2026, up from 60 per cent in 2024, according to data reported by DesignRush. Global publisher traffic fell 33 per cent year-on-year, small publishers lost 60 per cent of their referral traffic, and Business Insider lost more than half its search traffic and cut 21 per cent of its staff.

This isn’t really a story about journalism nostalgia. It’s a story about incentives. The people who once had a reason to produce timestamped, checkable, primary information — the reviews, price histories, safety recalls, and on-the-ground reporting that both humans and AI models quietly depend on — are losing the traffic that funded doing it properly.

The citation is disappearing faster than the click

Even when an AI system answers using real, current information, it rarely tells you where that information came from. Similarweb data reported by Search Engine Journal found that only 6.8 per cent of ChatGPT answers included an external link as of May 2026, up from roughly 1 per cent a year earlier. That still means 93 out of 100 responses send the user nowhere to verify anything.

Meanwhile, the machines are reading more of the web than ever. Digiday reports that automated traffic now accounts for 57.4 per cent of all web requests, more than actual humans, and AI agent traffic grew 45 per cent quarter-on-quarter in the second quarter of 2026 alone. Yet 97 per cent of the files publishers built specifically so AI systems could read and credit them received zero requests.

Put plainly: the machines are consuming the open web faster than ever, crediting almost none of it, and the people who built that web are being paid for it even less than before.

Also Read: Ecosystem Roundup: Taiwan blocks Grab’s US$600M foodpanda deal over Uber’s stake

This stopped being just a search problem. It’s now a money problem

The same pattern is showing up at the exact moment it matters most: when people decide what to buy. Digital Applied reports that half of US shoppers say they have made a purchase after researching it inside an AI tool, even when the transaction itself happened elsewhere. But in-chat checkout — the one step in that journey that leaves an auditable, timestamped record — converts roughly three times worse than a normal click-through to a retailer’s own site. OpenAI has already partly retreated from it, and Walmart’s Daniel Danker described in-chat checkout as “a very temporary moment in time.”

So the part of the shopping journey moving fastest into AI is discovery and reasoning, the part with no receipt. The part that still leaves a paper trail is the part being pushed back to ordinary websites. That should concern anyone who thinks about consumer protection, not only people who worry about publishing economics.

Why I keep returning to this

Before I built anything in software, I worked as a paramedic. The habit that job drilled into me was never cleverness. It was sequence: check the evidence, understand the situation, then act. You don’t get to skip the first step because the second one feels urgent, and in an emergency, skipping it is how people get hurt.

A great deal of AI product design today runs in the opposite order. The action — the answer, the recommendation, the purchase — arrives instantly. The evidence behind it, if it exists at all, is buried, unlinked, or already gone by the time anyone would think to check it.

Also Read: How AI became my sparring partner

What “showing your work” should actually mean

None of this is an argument for slowing AI down or forcing every chatbot to behave like a search engine with extra steps. It is an argument for a narrower, buildable standard: any AI system that influences a real decision — what to buy, what to believe, what to do next — should be able to show, on request, what claim it made, when it made it, and what that claim was based on.

That is a design requirement, not a moral lecture. Timestamps and traceable sourcing are inexpensive to build in from day one and expensive to bolt on after the fact. Most teams currently treat them as optional, if they think about them at all.

Southeast Asia does not have the luxury of waiting for someone else to solve this first. The region is folding AI assistants into commerce, banking and everyday decisions at speed, often for consumers who have far fewer alternative ways to double-check a claim than someone with a research team or a market data terminal. If the underlying layer of verifiable information keeps thinning out while AI interfaces keep expanding on top of it, the gap between people who can check what they’re told and people who can’t will not close. It will widen quietly, behind a much friendlier-looking interface.

The fix is simple to describe, even if it takes real engineering discipline to build: keep the receipts, and put them where the person making the decision can actually see them.

An answer you cannot check is not intelligence. It is just a well-formatted rumour.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post The most dangerous thing about AI isn’t hallucination. It’s that it stopped showing its work appeared first on e27.