{
  "id": 476349,
  "title": "AI’s Growing Hallucination Problem Puts Enterprise Adoption to the Test",
  "url": "https://urgent.news/2026/08/10/ais-growing-hallucination-problem-puts-enterprise-adoption-to-the-test",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-10T14:12:53.000Z",
  "source": {
    "name": "Newsweek",
    "slug": "newsweek",
    "url": "https://www.newsweek.com/ais-growing-hallucination-problem-puts-enterprise-adoption-to-the-test-12296603"
  },
  "original_language": "en",
  "account": "Artificial intelligence has been advancing at an exponential pace, outstripping the safeguards designed to manage its capabilities. Enterprises are now integrating AI into their core operations, yet the technology still struggles with producing confident errors, a phenomenon that can transform a productivity aid into a potential liability. OpenAI has observed an eightfold surge in ChatGPT Enterprise message usage over the past year, while the issue of hallucinations persists despite the growing sophistication of models. The implications are escalating from harmless chatbot responses to posing genuine professional risks. A recent case involved the Connecticut Supreme Court expressing concern over lawyers using AI-generated case citations in legal documents. Earlier, the Financial Times revealed that certain PwC Middle East reports contained fabricated footnotes and other inaccuracies stemming from AI hallucinations. These instances highlight a challenge that enterprises cannot overcome by merely acquiring access to the latest model: the surrounding system must ensure the output remains dependable. As AI becomes more entrenched within corporate structures, the challenge of maintaining security grows increasingly difficult to ignore. Obsidian Security, a company focused on securing AI agents that access sensitive business information, recently raised $85 million in funding, boasting a valuation of $1.1 billion. Nearly 70 percent of its clients now permit AI agents to interact with company data. Industry estimates suggest that AI hallucinations may cost businesses tens of billions of dollars annually, with organizations experiencing losses reaching millions of dollars when errors go undetected. As enterprises grant AI increased access to corporate systems, the reliability of the data and information AI can access becomes crucial to determining trustworthiness. Ryan Taylor, Chief Technology Officer at OSIBytes, contends that the industry has prioritized measuring AI by what it can generate, rather than scrutinizing the process by which it arrives at an answer. His criticism stems from the underlying architecture of AI systems, emphasizing the need for mechanisms that control data access, protect proprietary information, and validate the credibility of responses. Taylor's observations are rooted in his extensive experience with enterprise infrastructure, a field he explored further during his master's studies in 2024, where he focused on the impact of AI hallucinations. He posits that the mathematical formula underlying a model's structure significantly influences the reliability of its outputs. According to Taylor, conventional AI systems often suffer from a failure pattern where models extract incomplete context from a limited data slice, subsequently generating answers based on that deficient information. He illustrates this issue with an example of his own, recounting an experience where an AI system he used for website development cost him about $5,000 but ultimately deleted the entire project due to the system's inability to determine how to handle the data file. Taylor attributes this to the hallucinations that occur when an AI system operates within its limitations, such as insufficient memory to handle complex tasks. His solution involves a hyper-agent structure, comprised of 40 agents responsible for gathering data from distinct sources before other agents can reconcile the information and determine the most coherent outcome. This approach is likened to a library where multiple agents retrieve information from specialized sections, enabling the system to work with more controlled and relevant data pools. Additionally, Taylor addresses a concern that may be even more sensitive than hallucinations: intellectual property. His private AI and agentic code-to-cloud platform, PyBlox, is designed to cater to individual clients and keep their data separate from other systems to prevent cross-contamination. This boundary is seen as a prerequisite for businesses considering the integration of AI into their operations, as it safeguards proprietary material from unauthorized access. Taylor argues that as AI penetration in corporate environments deepens, the importance of robust infrastructure will only increase. Each new capability granted to AI introduces another potential access point that must be rigorously controlled. Taylor emphasizes that useful AI relies heavily on the surrounding architectural framework, advocating for partnerships with businesses and venture capital firms that can help test the platform's capacity to meet enterprise needs. The industry seems to be moving in this direction, as AI becomes more accessible for deployment, the critical business question remains whether companies can integrate AI without compromising control. Ryan Taylor's cautionary stance highlights the need for human oversight to verify AI outputs, as impressive demonstrations of AI capabilities do not necessarily translate into reliable, business-ready solutions.",
  "summary": "As AI scales rapidly, businesses can face a key challenge: granting access without losing control or oversight.",
  "key_points": [],
  "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."
}