{"course":{"slug":"evaluating-ai","title":"Knowing If It Works","url":"https://addaly.com/learn/evaluating-ai","level":"intermediate","track":"build"},"lesson":{"slug":"standards-and-frameworks","title":"What the emerging rules ask for, and what they do not settle","url":"https://addaly.com/learn/evaluating-ai/standards-and-frameworks","position":62,"of":76,"minutes":9,"answer":"Every framework converges on the same practical demands (documented intended use, data provenance, measured per-group performance, logging, human oversight and re-evaluation) while the legal detail differs by country and is still being written.","keyIdea":"Every framework converges on the same practical demands (documented intended use, data provenance, measured per-group performance, logging, human oversight and re-evaluation) while the legal detail differs by country and is still being written."},"body":{"format":"markdown","text":"## A caution before anything else\n\nAI regulation is unsettled, it differs by jurisdiction and sector, and it is changing while you read this. Nothing here is legal advice, and anyone shipping something consequential needs a lawyer in the relevant country. What follows is the shape of the landscape, which is useful for deciding what to measure and what to keep.\n\n## The voluntary frameworks\n\n**NIST's AI Risk Management Framework** (United States, voluntary) organises work into four functions: Govern, Map, Measure, Manage. This entire course lives inside *Measure*, and the framework's contribution is mostly a shared vocabulary for talking to people outside engineering.\n\n**ISO/IEC 42001** specifies a management system for AI and can be certified against, in the same way as the better-known information-security standard. It asks about process (do you have policies, roles, and a review cycle) rather than about whether your model is any good. **ISO/IEC 23894** covers risk management guidance.\n\nNeither tells you what accuracy is acceptable, because no standard can.\n\n## The EU AI Act, in outline\n\nThe Act sorts systems by risk. A small set of practices is prohibited outright. A defined list of **high-risk** uses (including employment, education, essential services, and certain safety components) carries substantive obligations: risk management, data governance, technical documentation, automatic logging, human oversight, and stated levels of accuracy, robustness and cybersecurity. A further tier carries transparency duties, such as telling people they are interacting with a machine or that content is synthetic. General-purpose model providers carry their own obligations.\n\nThe obligations are phased, the harmonised technical standards that operationalise them are still being produced, and national implementation and enforcement practice are still forming. Anybody who tells you exactly what a given clause will require of your product in practice is ahead of the evidence.\n\n## The rules that will bind you sooner\n\nFor most teams, sector-specific and existing law matters more than any AI-specific statute:\n\n- **Medical devices**, where software that supports diagnosis has been regulated for years.\n- **Financial services model risk**, with long-standing supervisory expectations about validation, independent review and documentation.\n- **Employment screening**, where some jurisdictions already require bias audits of automated hiring tools.\n- **Data protection.** In India the Digital Personal Data Protection Act governs personal data; in Europe the GDPR does, including rules about automated decisions. Both bear directly on whether user content may become an evaluation set.\n- **Consumer protection and advertising law**, which is where \"our AI is 99% accurate\" becomes a claim you must be able to evidence.\n\n## What they all converge on\n\nRead across the frameworks and the same practical list appears every time. If you do these, you are most of the way to any of them:\n\n1. Documented intended use and out-of-scope use.\n2. Documented data provenance, including consent basis and dates.\n3. Measured performance, reported per relevant group rather than only in aggregate.\n4. Logging sufficient to reconstruct a decision after the fact.\n5. Human oversight that is specified and, ideally, measured.\n6. Incident response, including a way for affected people to complain and be answered.\n7. Periodic re-evaluation, with results retained.\n\nPoint seven has a practical consequence that surprises teams: you may need to answer \"how did you know it worked\" about a version you shipped eighteen months ago. That means retaining evaluation artefacts (the set as it was, the results, the model version, the report) not just the current ones. Storage is cheap; reconstructing a defence is not.\n\n## What this changes about how you measure\n\nThree things in this course become obligations rather than good practice once any of these regimes applies to you.\n\n**Per-group reporting stops being optional.** An aggregate figure does not answer a question about whether a system works for a protected group, and the fairness arithmetic later in this course is the technical content behind that requirement.\n\n**Accuracy claims must be evidenced.** Whatever level you state in your documentation is a level somebody may hold you to, which argues for stating a measured figure with its interval and its set, rather than a round number that sounded good.\n\n**Logging becomes a design constraint.** \"Sufficient to reconstruct a decision\" is a demanding specification, and it is much cheaper to build at the start than to retrofit after a request arrives.\n\n## The honest limit\n\nCompliance is not safety. A system can satisfy every documentation requirement and still hurt people, and a careful system can be non-conformant on paperwork. The frameworks are floors and record-keeping disciplines, and they are worth meeting; they are not a substitute for the measurement judgement this course has been about, and treating them as one is how organisations end up with excellent binders and unexamined systems."},"check":{"question":"What is the most practical reason to retain evaluation artefacts (the set as it was, results, model version, report) for old releases?","options":[{"text":"Regulations require a fixed retention period for all AI evaluation data.","correct":false,"tempting":"States a universal rule that does not exist; retention duties vary by jurisdiction, sector and data type."},{"text":"You may have to answer how you knew a version shipped long ago was fit for use, and that answer cannot be reconstructed later.","correct":true},{"text":"Old evaluation sets are needed to demonstrate improvement over time to customers.","correct":false,"tempting":"A marketing use that does not require the full artefact trail and is not what the record exists for."},{"text":"Retained sets can be re-used to retrain the model when performance degrades.","correct":false,"tempting":"Confuses evaluation data with training data, and using a held-out set for training destroys the only thing it was for."}]},"prev":{"slug":"independent-evaluation","title":"Why the builder's own numbers are weak evidence","url":"https://addaly.com/learn/evaluating-ai/independent-evaluation"},"next":{"slug":"evaluating-for-harm","title":"Measuring the worst case, not the average one","url":"https://addaly.com/learn/evaluating-ai/evaluating-for-harm"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'What the emerging rules ask for, and what they do not settle', in Knowing If It Works. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/evaluating-ai/standards-and-frameworks (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What the emerging rules ask for, and what they do not settle","courseTitle":"Knowing If It Works","module":{"index":7,"slug":"assurance","title":"Evidence other people can act on"},"answer":null,"keyIdea":"Every framework converges on the same practical demands (documented intended use, data provenance, measured per-group performance, logging, human oversight and re-evaluation) while the legal detail differs by country and is still being written.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"a-caution-before-anything-else","runs":[{"text":"A caution before anything else"}]},{"t":"p","runs":[{"text":"AI regulation is unsettled, it differs by jurisdiction and sector, and it is changing while you read this. Nothing here is legal advice, and anyone shipping something consequential needs a lawyer in the relevant country. What follows is the shape of the landscape, which is useful for deciding what to measure and what to keep."}]},{"t":"h2","id":"the-voluntary-frameworks","runs":[{"text":"The voluntary frameworks"}]},{"t":"p","runs":[{"text":"NIST’s AI Risk Management Framework","b":true},{"text":" (United States, voluntary) organises work into four functions: Govern, Map, Measure, Manage. This entire course lives inside "},{"text":"Measure","i":true},{"text":", and the framework’s contribution is mostly a shared vocabulary for talking to people outside engineering."}]},{"t":"p","runs":[{"text":"ISO/IEC 42001","b":true},{"text":" specifies a management system for AI and can be certified against, in the same way as the better-known information-security standard. It asks about process (do you have policies, roles, and a review cycle) rather than about whether your model is any good. "},{"text":"ISO/IEC 23894","b":true},{"text":" covers risk management guidance."}]},{"t":"p","runs":[{"text":"Neither tells you what accuracy is acceptable, because no standard can."}]},{"t":"h2","id":"the-eu-ai-act-in-outline","runs":[{"text":"The EU AI Act, in outline"}]},{"t":"p","runs":[{"text":"The Act sorts systems by risk. A small set of practices is prohibited outright. A defined list of "},{"text":"high-risk","b":true},{"text":" uses (including employment, education, essential services, and certain safety components) carries substantive obligations: risk management, data governance, technical documentation, automatic logging, human oversight, and stated levels of accuracy, robustness and cybersecurity. A further tier carries transparency duties, such as telling people they are interacting with a machine or that content is synthetic. General-purpose model providers carry their own obligations."}]},{"t":"p","runs":[{"text":"The obligations are phased, the harmonised technical standards that operationalise them are still being produced, and national implementation and enforcement practice are still forming. Anybody who tells you exactly what a given clause will require of your product in practice is ahead of the evidence."}]},{"t":"h2","id":"the-rules-that-will-bind-you-sooner","runs":[{"text":"The rules that will bind you sooner"}]},{"t":"p","runs":[{"text":"For most teams, sector-specific and existing law matters more than any AI-specific statute:"}]},{"t":"ul","items":[[{"text":"Medical devices","b":true},{"text":", where software that supports diagnosis has been regulated for years."}],[{"text":"Financial services model risk","b":true},{"text":", with long-standing supervisory expectations about validation, independent review and documentation."}],[{"text":"Employment screening","b":true},{"text":", where some jurisdictions already require bias audits of automated hiring tools."}],[{"text":"Data protection.","b":true},{"text":" In India the Digital Personal Data Protection Act governs personal data; in Europe the GDPR does, including rules about automated decisions. Both bear directly on whether user content may become an evaluation set."}],[{"text":"Consumer protection and advertising law","b":true},{"text":", which is where “our AI is 99% accurate” becomes a claim you must be able to evidence."}]]},{"t":"h2","id":"what-they-all-converge-on","runs":[{"text":"What they all converge on"}]},{"t":"p","runs":[{"text":"Read across the frameworks and the same practical list appears every time. If you do these, you are most of the way to any of them:"}]},{"t":"ol","start":1,"items":[[{"text":"Documented intended use and out-of-scope use."}],[{"text":"Documented data provenance, including consent basis and dates."}],[{"text":"Measured performance, reported per relevant group rather than only in aggregate."}],[{"text":"Logging sufficient to reconstruct a decision after the fact."}],[{"text":"Human oversight that is specified and, ideally, measured."}],[{"text":"Incident response, including a way for affected people to complain and be answered."}],[{"text":"Periodic re-evaluation, with results retained."}]]},{"t":"p","runs":[{"text":"Point seven has a practical consequence that surprises teams: you may need to answer “how did you know it worked” about a version you shipped eighteen months ago. That means retaining evaluation artefacts (the set as it was, the results, the model version, the report) not just the current ones. Storage is cheap; reconstructing a defence is not."}]},{"t":"h2","id":"what-this-changes-about-how-you-measure","runs":[{"text":"What this changes about how you measure"}]},{"t":"p","runs":[{"text":"Three things in this course become obligations rather than good practice once any of these regimes applies to you."}]},{"t":"p","runs":[{"text":"Per-group reporting stops being optional.","b":true},{"text":" An aggregate figure does not answer a question about whether a system works for a protected group, and the fairness arithmetic later in this course is the technical content behind that requirement."}]},{"t":"p","runs":[{"text":"Accuracy claims must be evidenced.","b":true},{"text":" Whatever level you state in your documentation is a level somebody may hold you to, which argues for stating a measured figure with its interval and its set, rather than a round number that sounded good."}]},{"t":"p","runs":[{"text":"Logging becomes a design constraint.","b":true},{"text":" “Sufficient to reconstruct a decision” is a demanding specification, and it is much cheaper to build at the start than to retrofit after a request arrives."}]},{"t":"h2","id":"the-honest-limit","runs":[{"text":"The honest limit"}],"role":"limit"},{"t":"p","runs":[{"text":"Compliance is not safety. A system can satisfy every documentation requirement and still hurt people, and a careful system can be non-conformant on paperwork. The frameworks are floors and record-keeping disciplines, and they are worth meeting; they are not a substitute for the measurement judgement this course has been about, and treating them as one is how organisations end up with excellent binders and unexamined systems."}]}],"check":{"question":[{"text":"What is the most practical reason to retain evaluation artefacts (the set as it was, results, model version, report) for old releases?"}],"options":[{"text":[{"text":"Regulations require a fixed retention period for all AI evaluation data."}],"correct":false,"tempting":[{"text":"States a universal rule that does not exist; retention duties vary by jurisdiction, sector and data type."}]},{"text":[{"text":"You may have to answer how you knew a version shipped long ago was fit for use, and that answer cannot be reconstructed later."}],"correct":true,"tempting":null},{"text":[{"text":"Old evaluation sets are needed to demonstrate improvement over time to customers."}],"correct":false,"tempting":[{"text":"A marketing use that does not require the full artefact trail and is not what the record exists for."}]},{"text":[{"text":"Retained sets can be re-used to retrain the model when performance degrades."}],"correct":false,"tempting":[{"text":"Confuses evaluation data with training data, and using a held-out set for training destroys the only thing it was for."}]}],"order":[3,2,0,1],"why":null},"prev":{"slug":"independent-evaluation","title":"Why the builder’s own numbers are weak evidence","minutes":8},"next":{"slug":"evaluating-for-harm","title":"Measuring the worst case, not the average one","minutes":9}}}