{"id":8823,"date":"2025-12-11T11:05:04","date_gmt":"2025-12-11T10:05:04","guid":{"rendered":"https:\/\/lorit-consultancy.com\/en\/?p=8823"},"modified":"2026-02-20T09:25:03","modified_gmt":"2026-02-20T08:25:03","slug":"interpretation-and-implementation-of-safety-standards-with-ai","status":"publish","type":"post","link":"https:\/\/lorit-consultancy.com\/en\/2025\/12\/interpretation-and-implementation-of-safety-standards-with-ai\/","title":{"rendered":"When AI Meets the Rulebook: Interpretation and Implementation of Safety Standards in the age of Artificial Intelligence"},"content":{"rendered":"<h2>Opportunities, Risks and Limitations of AI when it comes to reading the Standards<\/h2>\n<p>The rapid development of Large Language Models (LLMs) and AI tools is increasingly changing the way technical documents and regulatory standards are interpreted and applied. What is already possible today \u2013 and where clear limitations remain \u2013 affects key areas of Lorit Consultancy day-to-day work as well: from functional safety according to <a href=\"https:\/\/lorit-consultancy.com\/en\/standards\/automotive\/iso26262\/\" target=\"_blank\" rel=\"noopener\"><strong>ISO 26262<\/strong><\/a>, to medical standards such as <strong><a href=\"https:\/\/lorit-consultancy.com\/en\/standards\/medical-devices\/iec60601\/\" target=\"_blank\" rel=\"noopener\">IEC 60601<\/a>\/ <a href=\"https:\/\/lorit-consultancy.com\/en\/standards\/medical-devices\/iec62304\/\" target=\"_blank\" rel=\"noopener\">62304<\/a><\/strong>, to quality management systems like ISO 9001, <a href=\"https:\/\/lorit-consultancy.com\/en\/standards\/automotive\/iatf16949\/\" target=\"_blank\" rel=\"noopener\"><strong>IATF 16949<\/strong><\/a>, or risk management according to <strong><a href=\"https:\/\/lorit-consultancy.com\/en\/standards\/medical-devices\/iso14971\/\" target=\"_blank\" rel=\"noopener\">ISO 14971<\/a><\/strong>.<\/p>\n<p>This blog draws on practical experience and concrete examples to explore how reliably artificial intelligence interprets normative texts, where typical sources of error may lie, what risks can arise from incorrect interpretations \u2013 and where AI can indeed help.<\/p>\n<h2 id=\"How-reliably-does-AI-interpret-complex-normative-requirements?.1\" data-local-id=\"9cc7eded-353e-4229-8940-44e5322ba07f\" data-renderer-start-pos=\"6907\">How reliably does AI interpret complex normative requirements?<button class=\"css-x4slh0\" type=\"button\" data-testid=\"anchor-button\" aria-hidden=\"true\"><\/button><\/h2>\n<p data-renderer-start-pos=\"6971\" data-local-id=\"dc108091-4e28-4e11-9c0c-3d776b1993f3\">Modern AI models can analyse large amounts of text and present the results in a structured way. When working with standards, this means:<\/p>\n<ul>\n<li data-renderer-start-pos=\"7111\" data-local-id=\"5696afb6-1f20-4fc7-83a6-d9cf231c2216\">AI can logically organise chapters<\/li>\n<li data-renderer-start-pos=\"7111\" data-local-id=\"5696afb6-1f20-4fc7-83a6-d9cf231c2216\">It can summarise complex correlations (though not always in a truly logical way)<\/li>\n<li data-renderer-start-pos=\"7111\" data-local-id=\"5696afb6-1f20-4fc7-83a6-d9cf231c2216\">It can explain terminology and cross-references better than many traditional search tools<\/li>\n<li data-renderer-start-pos=\"7111\" data-local-id=\"5696afb6-1f20-4fc7-83a6-d9cf231c2216\">It provides an initial orientation regarding scope and key requirements<\/li>\n<\/ul>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-8824\" src=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-book-interpretation_adobe-stock-1024x585.jpeg\" alt=\"\" width=\"650\" height=\"371\" srcset=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-book-interpretation_adobe-stock-1024x585.jpeg 1024w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-book-interpretation_adobe-stock-1920x1097.jpeg 1920w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-book-interpretation_adobe-stock-768x439.jpeg 768w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-book-interpretation_adobe-stock-1536x878.jpeg 1536w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-book-interpretation_adobe-stock-2048x1170.jpeg 2048w\" sizes=\"auto, (max-width: 650px) 100vw, 650px\" \/><\/p>\n<p>However: AI does not \u201cunderstand\u201d standards in a legal-technical sense. It recognizes statistical patterns in wording, not the underlying formal logic. Reliability therefore depends heavily on input quality, context, and the specific version of the standard.<\/p>\n<p>Thus, while AI often delivers good explanatory outputs \u2013 for example on the structure of the ISO 26262 safety lifecycle \u2013 it can make mistakes in precise requirements such as ASIL decomposition, hardware architectural metrics (ISO 26262), or detailed rules in standards like IEC 61010-1.<\/p>\n<h2 id=\"Typical-sources-of-error.1\" data-local-id=\"c11074a6-338c-4871-960c-88dc33cf130c\" data-renderer-start-pos=\"7953\">Typical sources of error<\/h2>\n<h3 data-renderer-start-pos=\"7981\" data-local-id=\"a08fa28d-4370-4b59-9c92-1c6ef18e6c17\"><strong data-renderer-mark=\"true\">1. Incorrect output<\/strong><\/h3>\n<p data-renderer-start-pos=\"8001\" data-local-id=\"ff5dee23-b603-45a4-b48e-69b8c944f4c6\">One recurring issue with common AI tools is that they can generate <strong data-renderer-mark=\"true\">plausible-sounding but incorrect<\/strong> normative requirements, such as:<\/p>\n<ul>\n<li data-renderer-start-pos=\"8137\" data-local-id=\"d3ddb678-6354-414e-bfba-87ed5301ed3f\">fictional chapters<\/li>\n<li data-renderer-start-pos=\"8137\" data-local-id=\"d3ddb678-6354-414e-bfba-87ed5301ed3f\">non-existent tables<\/li>\n<li data-renderer-start-pos=\"8137\" data-local-id=\"d3ddb678-6354-414e-bfba-87ed5301ed3f\">incorrectly interpreted safety limit\/threshold values<\/li>\n<\/ul>\n<p data-renderer-start-pos=\"8239\" data-local-id=\"090478b6-6030-4588-bfd0-e8dbcfccdfb3\">This is particularly risky in safety and medical domains. For example, I recently asked a widely used AI tool to explain a table with electrical clearances and voltages from a standard, and the result was disappointing: the table and values were interpreted completely incorrectly. So, caution is advised here!<\/p>\n<p data-renderer-start-pos=\"8239\" data-local-id=\"090478b6-6030-4588-bfd0-e8dbcfccdfb3\"><img loading=\"lazy\" decoding=\"async\" class=\"wp-image-8826 aligncenter\" src=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-hallucination-mode_adobe-stock-1024x683.jpeg\" alt=\"\" width=\"650\" height=\"433\" srcset=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-hallucination-mode_adobe-stock-1024x683.jpeg 1024w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-hallucination-mode_adobe-stock-1920x1280.jpeg 1920w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-hallucination-mode_adobe-stock-768x512.jpeg 768w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-hallucination-mode_adobe-stock-1536x1024.jpeg 1536w, https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2025\/12\/AI-hallucination-mode_adobe-stock-2048x1365.jpeg 2048w\" sizes=\"auto, (max-width: 650px) 100vw, 650px\" \/><\/p>\n<h3 data-renderer-start-pos=\"8558\" data-local-id=\"f85ef84c-827b-4bb5-b8e9-92a350fd3a3d\"><strong data-renderer-mark=\"true\">2. Outdated or incomplete data sources<\/strong><\/h3>\n<p data-renderer-start-pos=\"8597\" data-local-id=\"5ba35428-3247-4c17-b505-97bad19aa791\">Because many standards are not publicly accessible (except through paid licences), AI models are frequently trained on secondary sources such as online articles, presentations, or forum discussions. This can lead to:<\/p>\n<ul>\n<li data-renderer-start-pos=\"8817\" data-local-id=\"3771a0f0-e474-433a-8a40-537b83fc31c2\">mixing of old and new versions of standards<\/li>\n<li data-renderer-start-pos=\"8817\" data-local-id=\"3771a0f0-e474-433a-8a40-537b83fc31c2\">interpretations based on online commentary rather than the original text<\/li>\n<li data-renderer-start-pos=\"8817\" data-local-id=\"3771a0f0-e474-433a-8a40-537b83fc31c2\">missing details, particularly those found in annexes<\/li>\n<\/ul>\n<h3 data-renderer-start-pos=\"8998\" data-local-id=\"4308a7ec-f04b-409b-b195-17a424ed890d\"><strong data-renderer-mark=\"true\">3. Human-influenced errors<\/strong><\/h3>\n<p data-renderer-start-pos=\"9025\" data-local-id=\"05dd3255-aea3-42ea-9d7d-96855dce9243\">Inaccurate or leading prompts can unintentionally bias the response. Models also generate answers based on probabilities \u2013 not on absolute truth.<\/p>\n<p data-renderer-start-pos=\"9172\" data-local-id=\"c39c647d-1d4b-4655-91f5-c5ba99dd807e\">Wrong or imprecise prompts may result in:<\/p>\n<ul>\n<li data-renderer-start-pos=\"9217\" data-local-id=\"9915e701-4ed4-4384-8432-73b50b32b9d4\">misinterpretations<\/li>\n<li data-renderer-start-pos=\"9217\" data-local-id=\"9915e701-4ed4-4384-8432-73b50b32b9d4\">overly broad or overly complicated answers<\/li>\n<li data-renderer-start-pos=\"9217\" data-local-id=\"9915e701-4ed4-4384-8432-73b50b32b9d4\">mixing of unrelated topics or standards<\/li>\n<li data-renderer-start-pos=\"9217\" data-local-id=\"9915e701-4ed4-4384-8432-73b50b32b9d4\">\u201cembellished\u201d statements if the user implicitly nudges the model in a certain direction<\/li>\n<\/ul>\n<h3 data-renderer-start-pos=\"9421\" data-local-id=\"2828f30a-d7e9-42af-8f47-8f25a711f99a\"><strong data-renderer-mark=\"true\">4. Technical limitations and bugs<\/strong><\/h3>\n<p data-renderer-start-pos=\"9455\" data-local-id=\"014eb941-e585-484f-99ac-c554931e849e\">Complex relationships, mathematical challenges such as those in FMEDA or SPFM\/LFM calculations, and iterative models like HARA \u2192 FSC \u2192 TSC can exceed the sequential reasoning capabilities of many AI tools. Even modern tools may misinterpret tables, metrics, or calculations.<\/p>\n<h3 data-renderer-start-pos=\"9733\" data-local-id=\"2075fd07-ad40-4ed5-b1ad-f32d8b09aeb6\"><strong data-renderer-mark=\"true\">5. Soft errors<\/strong><\/h3>\n<p data-renderer-start-pos=\"9748\" data-local-id=\"8e118b67-f6f0-41d1-b153-3ee04ebb9da8\">Soft errors are well known in informatics as errors that cause temporary, unintentional changes in logic circuits or memory states (see our blog <a class=\"_ymio1r31 _ypr0glyw _zcxs1o36 _mizu1v1w _1ah3dkaa _ra3xnqa1 _128mdkaa _1cvmnqa1 _4davt94y _4bfu1r31 _1hms8stv _ajmmnqa1 _vchhusvi _kqswh2mm _syaz13af _ect41gqc _1a3b1r31 _4fpr8stv _5goinqa1 _f8pj13af _9oik1r31 _1bnxglyw _jf4cnqa1 _30l313af _1nrm1r31 _c2waglyw _1iohnqa1 _9h8h12zz _10531ra0 _1ien1ra0 _n0fx1ra0 _1vhv17z1\" title=\"https:\/\/lorit-consultancy.com\/en\/2020\/10\/are-we-going-soft-on-errors-part-1\/\" href=\"https:\/\/lorit-consultancy.com\/en\/2020\/10\/are-we-going-soft-on-errors-part-1\/\" target=\"_blank\" rel=\"noopener\" data-renderer-mark=\"true\">Are we going soft on errors?<\/a>).<\/p>\n<p data-renderer-start-pos=\"9748\" data-local-id=\"8e118b67-f6f0-41d1-b153-3ee04ebb9da8\">Such temporary changes can affect system reliability, and AI systems are theoretically not immune to these events either.<\/p>\n<p data-renderer-start-pos=\"9748\" data-local-id=\"8e118b67-f6f0-41d1-b153-3ee04ebb9da8\"><\/div><\/div><\/div><div class=\"content_section blue_bg blog_trenner_section\"><div class=\"row align-center medium-align-spaced\"><div class=\"columns border_solid_square post_thumbnail small-10 medium-5 large-3\"><div  data-ratio=\"1.2783505154639\" class=\"\"><picture><source media=\"(min-width:1024px)\" srcset=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2021\/01\/dijaz-maric-1.webp\" width=\"873\" height=\"1116\" type=\"image\/webp\" ><source media=\"(min-width:640px)\" srcset=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2021\/01\/dijaz-maric-1-801x1024.webp\" width=\"801\" height=\"1024\" type=\"image\/webp\" ><img decoding=\"async\" src=\"https:\/\/lorit-consultancy.com\/wp-content\/uploads\/2021\/01\/dijaz-maric-1-640x640.webp\" alt=\"\" loading=\"lazy\" width=\"640\" height=\"640\" type=\"image\/webp\" ><\/picture><svg version=\"1.1\" id=\"svg_border_solid_square\" class=\"svg_border_solid_square\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" xmlns:xlink=\"http:\/\/www.w3.org\/1999\/xlink\" x=\"0px\" y=\"0px\"\n\t viewBox=\"0 0 337 411.2\" style=\"enable-background:new 0 0 337 411.2;\" xml:space=\"preserve\">\n<polygon id=\"bg\" class=\"bg\" points=\"65,332.6 337,332.6 336.9,411.2 3.6,411.2 \"\/>\n<path id=\"border\" class=\"border\" d=\"M334,329.6V5.9c0-1.6-1.3-2.9-2.9-2.9l0,0L5.9,3.1C4.3,3.1,3,4.3,3,6v400.8c0,1.6,0.8,1.9,1.8,0.6\n\tl59.6-74.7l266.6-0.1C332.7,332.5,333.9,331.2,334,329.6z\"\/>\n<\/svg><\/div><div class=\"image_bottom_content\"><p>Dijaz Maric, Quality &amp; Safety Consultant<\/p>\n<\/div><\/div><div class=\"columns post_content small-12 medium-6 large-7\"><p>If this topic resonates with challenges in your organisation, feel free to get in touch. We\u2019re here to help translate standards, risks and technology into workable solutions.<\/p>\n<p><a href=\"https:\/\/lorit-consultancy.com\/en\/contact-us\/#wpcf7-f4221-p4082-o1\">Contact us<\/a> for bespoke consultancy or join one of our upcoming <a href=\"https:\/\/lorit-consultancy.com\/en\/training\/\" target=\"_blank\" rel=\"noopener\">online courses<\/a>.<\/p>\n<a class=\"add_logo_border\" target=\"_blank\" href=\"https:\/\/lorit-consultancy.com\/en\/consultancy\/\"><span>Learn more<\/span><\/a><\/div><\/div><\/div><\/div><div class=\"single_content_section single_post_section content_section\"><div class=\"row\"><div class=\"post_content columns\"><\/p>\n<h2>Risks arising from incorrect interpretation of technical and safety-critical standards<\/h2>\n<p>Incorrect AI-generated outputs are not merely a quality issue \u2013 they can create real and tangible risks:<\/p>\n<h3 data-renderer-start-pos=\"10249\" data-local-id=\"4161da9b-87fd-466b-90cd-d03a70a63fe1\"><strong data-renderer-mark=\"true\">1. Safety risks<\/strong><\/h3>\n<p data-renderer-start-pos=\"10265\" data-local-id=\"4f82cf60-82ba-42e4-95c9-025bec6d8a4a\">If an incorrect AI assumption influences a safety concept (e.g., a wrong ASIL allocation), it may lead to insufficient safety measures.<\/p>\n<h3 data-renderer-start-pos=\"10404\" data-local-id=\"4b86b78d-d7c2-408f-9d8c-0a04693f5438\"><strong data-renderer-mark=\"true\">2. Contractual and audit risks<\/strong><\/h3>\n<p data-renderer-start-pos=\"10435\" data-local-id=\"672304a6-36dc-43f3-a109-e01a80f63be6\">Standards are often part of contractual requirements. Misinterpretations can lead to non-compliance, audit deviations, or even liability issues.<\/p>\n<h3 data-renderer-start-pos=\"10583\" data-local-id=\"b961b15e-1b1d-47e3-8051-4dfd802b2b5c\"><strong data-renderer-mark=\"true\">3. Incorrect risk assessments<\/strong><\/h3>\n<p data-renderer-start-pos=\"10613\" data-local-id=\"69e835dc-14fb-40c6-b01c-6d8a3d7c7281\">In the medical device industry, AI-generated risk assessments according to ISO 14971 may be misleading if essential parameters are missing or incorrectly linked.<\/p>\n<h3 data-renderer-start-pos=\"10778\" data-local-id=\"bd089330-aaf1-4fbc-b2cb-921b03a7d0ef\"><strong data-renderer-mark=\"true\">4. Quality degradation<\/strong><\/h3>\n<p data-renderer-start-pos=\"10801\" data-local-id=\"2bb9ef32-6add-470e-9218-54b5475af182\">Wrong AI outputs in quality management (e.g., incorrect process requirements) can undermine the consistency of a quality management system.<\/p>\n<h2 data-renderer-start-pos=\"10801\" data-local-id=\"2bb9ef32-6add-470e-9218-54b5475af182\">Where AI makes sense \u2013 and where it is better left out<\/h2>\n<table>\n<thead>\n<tr>\n<td><strong>Use Case<\/strong><\/td>\n<td><strong>Assessment<\/strong><\/td>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Summarising sections and providing initial orientation<\/td>\n<td><strong>Yes<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Explaining technically complex concepts<\/td>\n<td><strong>Yes, with caution<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Reformatting long texts<\/td>\n<td><strong>Yes, but risk of misinterpretation remains<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Deciding on conformity\/compliance<\/td>\n<td><strong>No \u2013 human judgement required<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Detailed safety interpretations and risk analyses (e.g., FMEA)<\/td>\n<td><strong>No \u2013 clear limitations<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Legally binding interpretations (standards often have legal impact)<\/td>\n<td><strong>No \u2013 requires human expertise<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Handling unstructured or sensitive data<\/td>\n<td><strong>Rather no \u2013 high risk of misinterpretation<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>Conclusion<\/h2>\n<p>Public AI tools can support certain aspects of standards interpretation, especially when a quick overview is required. However, when it comes to detailed, safety-critical or compliance-relevant topics, AI does not automatically increase efficiency.<\/p>\n<p>In the end, expertise determines the correctness and reliability of normative implementation.<\/p>\n<p>By <a href=\"https:\/\/lorit-consultancy.com\/en\/about-us\/#dijaz-maric\">Dijaz Maric<\/a>, Quality Management &amp; Reliability Engineering Consultant<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Opportunities, Risks and Limitations of AI when it comes to reading the Standards The rapid development of Large Language Models (LLMs) and AI tools is increasingly changing the way technical documents and regulatory standards are interpreted and applied. What is already possible today \u2013 and where clear limitations remain \u2013 affects key areas of Lorit [&hellip;]<\/p>\n","protected":false},"author":11,"featured_media":8828,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[9,90,91,96],"tags":[],"class_list":["post-8823","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized","category-reliability","category-safety","category-quality-management"],"acf":[],"_links":{"self":[{"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/posts\/8823","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/comments?post=8823"}],"version-history":[{"count":9,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/posts\/8823\/revisions"}],"predecessor-version":[{"id":8949,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/posts\/8823\/revisions\/8949"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/media\/8828"}],"wp:attachment":[{"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/media?parent=8823"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/categories?post=8823"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lorit-consultancy.com\/en\/wp-json\/wp\/v2\/tags?post=8823"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}