{
  "id": 12216524,
  "title": "Hallucination blocked before human",
  "url": "https://urgent.news/2026/10/05/alucinacao-bloqueada-antes-do-humano",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-05T19:40:00.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/beanstechbr/alucinacao-bloqueada-antes-do-humano-27ae"
  },
  "original_language": "pt",
  "account": "IBM has developed a principle of engineering called \"blocked hallucination before human\" which prioritizes the detection and suppression of false information generated by AI before it is displayed to users. This approach, used in IBM Granite, reduces hallucinations by up to 73% compared to models without prior verification. It involves validating against trusted sources and can reduce the cognitive load on human auditors by 60% in operational scenarios. The approach is compatible with Brazil's Artificial Intelligence Legal Framework and is formalized in the IBM Granite Guardrails Framework, version 2.1.",
  "summary": "The concept of \"Alucinação bloqueada antes do humano\" refers to a pre-filtering principle in AI engineering that prioritizes the detection and suppression of false, contradictory, or unsupported outputs generated by AI before they are presented to the user, without relying on real-time human intervention. This approach employs guardrails at the stage of generation or immediate post-processing, blocking hallucinations with subsecond latency. Solutions like IBM Granite, which incorporate retrieval-augmented verification, have been shown to reduce hallucinations by up to 73% compared to models without prior verification (IBM Research, 2024). The key to this approach is validating against reliable sources, such as validated legal or clinical documents, rather than just statistical classification. It does not replace human auditing in critical scenarios like healthcare or justice but can reduce cognitive load by 60% in operational cases (internal IBM Brazil study, 2023). Implementing \"block before human\" involves inserting technical verification layers between token generation and delivery, focusing on inference pipeline validation rather than post-generation analysis. This real-time, streaming-aware validation interrupts generation as soon as a hallucination indicator is detected with ≥92% confidence, requiring native integration between the model, retriever, and verifier. Limitations include the inability to eliminate hallucinations in deep knowledge gaps or resolve intentional ambiguities that require human interpretative context. Effectiveness hinges on the quality and coverage of RAG sources and calibration of the confidence threshold, which must be auditable and documented.",
  "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."
}