Technology and Artificial Intelligence

Identify product, model, version, function, integration, user group, geographical scope, and period.

Document the problem the system is intended to address, why technology or AI was chosen, and which less intrusive alternatives exist.

Map developer, data provider, model provider, integrator, purchaser, operator, decision-maker, and control body.

Identify who is analysed, classified, prioritised, affected, or excluded, including people who are not direct users.

Establish what the output may trigger, which rights, access, resources, or actions may change, and how serious the consequence is.

Document design, testing, deployment, update, monitoring, deviation, withdrawal, and retirement.

Record technical, professional, legal, and ethical competence, financial interests, and need for independent control.

Identify who owns purpose, data, model, deployment, individual case, appeal, correction, and final decision.

Document origin, collection method, time, licence, access, and changes for relevant data.

Test whether the data are suitable for the concrete purpose and relevant population or situation.

Map who and what is represented, underrepresented, overrepresented, or entirely absent.

Document who labelled data, which definitions were used, disagreement, error rate, and changes.

Make filtering, deduplication, normalisation, synthesis, imputation, and derived variables traceable.

Test errors, missing values, copies, common source origin, temporal leakage, and mixing between training and test.

Record privacy, copyright, consent, confidentiality, security, terms of use, and rights to deletion or access.

Establish how errors are corrected in source, derived datasets, model, output, and systems that received the data.

Document model type, version, parameters, training process, libraries, provider, and known limitations.

Make visible which signal, loss function, or business objective the model optimises and which human purpose it represents only indirectly.

Use measures that answer to use and error consequence, and do not report averages alone where subgroups or rare events are decisive.

Distinguish internal testing, external testing, prospective testing, robustness, stress testing, and actual effect after deployment.

Document how scores and probabilities correspond to actual frequencies and how uncertainty is shown to users.

Establish which populations, languages, environments, tasks, and periods were tested and when the model must not be used.

Test relevant alternative models, thresholds, feature choices, and assumptions where these may change the conclusion.

Monitor changing data, behaviour, error rates, misuse, updates, and differences between test and actual use.

Document whether the system informs, recommends, prioritises, rejects, approves, or independently triggers action.

Make thresholds, exceptions, manual rules, and consequences of different outputs visible to responsible roles.

Test whether the human has time, competence, information, and authority to depart from output without unreasonable cost or sanction.

Establish whether the case contains information the model cannot represent and how it is assessed before decision.

Distinguish technical explanation of output from the professional, legal, or normative reasons for the decision.

Identify a person or function that owns the final decision, contradiction, and correction.

Provide a real route through which facts, category, model output, rule use, and decision can be examined again.

Trace how scores and categories are stored, shared, and reused, and how errors are removed from later decisions.

Make visible when material is wholly or partly generated, distinguishing output from registered source, simulation, and human assessment.

Verify quotations, names, numbers, links, legal rules, professional claims, and other testable elements against actual sources.

Document relevant system instruction, user input, retrieved documents, tool use, and limits where needed to understand the output.

Test whether fluency, confident tone, or detail conceals absent grounds, contradiction, or fabricated sources.

Assess rights and safeguards concerning input, training material, style imitation, personal data, voice, image, and output.

Prevent generated portraits, voices, quotations, or roles from being presented as actual persons or events without clear labelling and grounds.

Identify who selected, edited, published, and took responsibility for the concrete use.

Make material error visible, update derived documents, and record which output version was used in a decision or publication.

Identify harm, misuse, adversaries, unintended use, dependencies, and persons or systems that may be affected.

Test normal, extreme, rare, and hostile conditions relevant to the use.

Limit data, model, tools, administrator rights, and dangerous functions by necessary role and traceable use.

Define safe fallback, human override, isolation, notification, and ability to stop or roll back rapidly.

Record errors, near misses, misuse, breaches, harm, and unexpected effects with responsibility and timely follow-up.

Track performance, changing data, subgroups, security breaches, complaints, overrides, and actual effects.

Map third-party models, libraries, cloud services, updates, dependencies, and vulnerability notification.

Use external or separated control where consequence, self-interest, complexity, or lack of access makes internal assessment insufficient.

Collect and use only data relevant and necessary for the documented purpose.

Give affected persons understandable information about system use, data categories, consequences, and responsible actor.

Establish which data, output, logs, rules, and reasons the affected person can access within lawful limits.

Enable correction of source data, category, attribution, and derived consequences throughout the chain.

Test error rates, access, and consequences for relevant groups without converting group analysis into a total judgment about an individual.

Assess language, disability, digital access, competence, and alternative channels for those the system does not serve.

Protect questions, reservation, complaint, and request for human review from adverse classifications lacking independent grounds.

Ensure a competent actor can change the decision, correct data, and stop further use rather than merely explain the system.

Correct the error where it arose and trace consequences through data, model, system, decision, records, and affected persons.

Stop, limit, or add human oversight where uncertainty or consequence makes continued use indefensible.

Document version, reason for change, testing, approval, deployment, and which prior results require reassessment.

Correct data, category, decision, access, and other concrete effects for persons affected by documented error.

Allocate responsibility between provider and user organisation without allowing contract or technical complexity to empty the responsibility chain.

Use independent control for high risk, major asymmetry, limited access, or serious consequence.

Test whether correction actually reduces the error without creating new harm, exclusion, safety failure, or goal displacement.

Define conditions for closure, reassessment, rollback, retirement, data deletion, and preservation of necessary audit trails.