{
  "id": 16827481,
  "title": "The most dangerous thing about AI isn’t hallucination. It’s that it stopped showing its work",
  "url": "https://urgent.news/2026/10/28/the-most-dangerous-thing-about-ai-isnt-hallucination-its-that-it",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-28T01:00:28.000Z",
  "source": {
    "name": "e27",
    "slug": "e27",
    "url": "https://e27.co/the-most-dangerous-thing-about-ai-isnt-hallucination-its-that-it-stopped-showing-its-work-20261025/"
  },
  "original_language": "en",
  "account": "The most alarming issue surrounding artificial intelligence is not its propensity for fabricating information, but rather the fact that it often fails to disclose its sources. For the past two years, the primary focus of AI safety discussions has been on hallucinations - instances where AI models generate facts, court cases, or sources that do not exist. However, a more pressing concern is on the rise, and it may prove to be more consequential. AI-generated answers can be accurate yet impossible to verify, and this is no longer an exception but rather the norm. A significant trend is the prevalence of zero-click searches, where users receive answers without visiting any websites. This occurrence has risen from 60% in 2024 to 68% in early 2026, as reported by DesignRush. Global publisher traffic has decreased by 33% compared to the previous year, small publishers have lost 60% of their referral traffic, and Business Insider experienced over half of their search traffic loss, resulting in a 21% staff reduction. This issue is not solely about nostalgia for journalism; it is about the changing incentives. Publishers who once had reasons to produce timestamped, checkable, primary information are losing the traffic that funded their proper execution. AI systems rarely provide the sources for their information, with only 6.8% of ChatGPT responses including external links as of May 2026, a significant increase from roughly 1% a year earlier. As a result, 93 out of 100 AI responses send users to unrelated websites, leaving no opportunity for verification. Automated traffic now constitutes 57.4% of all web requests, surpassing human traffic, and AI agent traffic grew by 45% quarter-on-quarter in the second quarter of 2026 alone. Nevertheless, 97% of the files designed for AI systems to read and credit have received no requests. This situation is not just a search problem; it has become a revenue problem. In the context of consumer decisions, half of US shoppers have made purchases after researching them within AI tools, even if the actual transaction occurred elsewhere. However, in-chat checkout, the critical step in this process that leaves a traceable record, converts three times worse than a normal click-through to a retailer's website. OpenAI has already partially abandoned this approach, and Walmart's Daniel Danker described in-chat checkout as a temporary phenomenon. Consequently, the part of the shopping journey that involves AI is the discovery and reasoning phase, while the critical part that leaves an auditable, timestamped record is being pushed back to conventional websites. This trend is concerning for consumer protection advocates, not just those worried about publishing economics. In the realm of consumer protection, it is crucial to emphasize the importance of showing one's work. AI product design today often prioritizes speed over transparency. Instead of following a sequential approach - checking evidence, understanding the situation, and acting - AI systems prioritize providing instant answers and recommendations, with the supporting evidence often buried or unavailable for verification. The argument presented here is not to slow down AI development or force every chatbot to resemble a search engine, but to establish a narrower, buildable standard. Any AI system that influences real decisions, such as purchases, beliefs, or actions, should be capable of providing clear evidence of the claims it makes, when those claims were made, and the basis for those claims. This requirement is both a design necessity and a moral imperative, as it ensures transparency and trustworthiness in AI systems. Southeast Asia, in particular, cannot afford to wait for others to address this issue. The region is rapidly integrating AI assistants into commerce, banking, and everyday decision-making, often leaving consumers with limited avenues to verify the information presented to them. The thinning layer of verifiable information, combined with the expanding AI interfaces, will create a widening gap between those who can check the claims made by AI systems and those who cannot. This gap will become increasingly problematic as AI interfaces become more user-friendly, leading to a situation where the distinction between verified information and unsubstantiated claims becomes blurred. The solution is straightforward: maintain records of the evidence and present them in a way that allows the decision-makers to easily access them. An unverifiable answer is not intelligence; it is merely a well-crafted rumor.",
  "summary": "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 […] The post The most dangerous thing about AI isn’t hallucination. It’s that it…",
  "key_points": [
    "AI systems often fail to disclose sources for generated information",
    "Zero-click searches rose from 60% in 2024 to 68% in early 2026",
    "93 out of 100 AI responses send users to unrelated websites with no verification path"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}