Responsible AI standard

Before asking whether something can be built, we ask what human purpose it serves, who benefits from it, who may be harmed by it, who remains responsible when it is wrong, what power it creates or removes, and whether it should exist in the proposed form at all.

This standard applies to the products we create, the systems we implement for clients, the technologies we license, and the commercial work we choose to accept.

Every system should begin with a clearly defined human or institutional need. "We should use AI" is not a sufficient problem statement. Before development begins, we should be able to explain what problem is being addressed, who experiences it, why it matters, what currently happens without the system, what a better outcome would look like, and whether AI is necessary to produce it. When a simpler process, conventional software, organisational change, or human service would work better, we recommend that instead.

A system may be efficient while still reducing agency. It may make decisions faster while leaving people less able to understand them, remove effort while removing skill, or provide convenience while making departure increasingly difficult. We examine whether a system helps people understand more clearly, make better-informed decisions, develop capability rather than lose it, question the system's output, retain meaningful choice, act without unnecessary dependency, and leave or refuse the system without unreasonable cost. Automation should not remove meaningful human judgement simply because removal is technically possible.

People should not be led to believe they are interacting with a person, receiving human judgement, or viewing independently produced material when AI has played a significant role. At minimum, people should be informed when AI materially affects a decision about them, advice they may reasonably rely upon, the content they receive, the evaluation of their work or conduct, access to a service or opportunity, the interpretation of sensitive information, or a simulated human relationship. A hidden statement in lengthy terms is not meaningful disclosure.

An AI system cannot accept legal, moral, or institutional responsibility. Responsibility remains with the people and organisations that design, deploy, operate, approve, and rely upon it. For each significant system, there should be identifiable responsibility for approving its purpose, determining where it may be used, evaluating its performance, reviewing important outputs, responding to failures, investigating reported harm, correcting or withdrawing the system, and communicating with people affected by it. "AI made the decision" is not an acceptable account of responsibility.

People affected by significant AI outputs should have a meaningful way to question, review, correct, or appeal them. Review may include human reconsideration, access to relevant inputs, correction of inaccurate information, explanation of the factors influencing an output, escalation to an accountable person, independent assessment, or reversal or suspension of an automated action. A person should not be trapped inside a decision process that no responsible human is willing or able to examine.

A demonstration or benchmark does not establish that a system is ready for use. Systems should be evaluated under conditions resembling the environment in which they will operate, examining accuracy and reliability, performance across relevant groups and circumstances, failure modes, uncertainty and hallucination, resistance to misuse, privacy and security, human oversight, accessibility, operational resilience, unintended incentives, effects on behaviour and decision-making, and the consequences of false positives and false negatives. A system used for low-stakes drafting does not require the same safeguards as one influencing employment, finance, education, healthcare, safety, legal rights, or access to essential services.

AI systems often produce fluent outputs even when their information is incomplete, uncertain, outdated, or wrong. Fluency should not be mistaken for confidence. Where uncertainty is material, the system should express it, identify missing information, distinguish fact from inference, provide sources, recommend human review, refuse to answer beyond an appropriate threshold, and avoid false precision. The design should not encourage people to trust an answer merely because it is delivered confidently.

The existence of useful data does not automatically justify collecting, combining, retaining, or analyzing it. We consider whether the data is necessary for the stated purpose, whether less sensitive data would be sufficient, whether the person reasonably expects this use, whether consent is meaningful, how long the data must be retained, who can access it, how it may be combined with other information, what happens if it is exposed or misused, and whether the system can operate with less data. A business model should not depend on collecting more information than the service genuinely requires.

Some uses of AI affect people's rights, dignity, safety, opportunities, or access to essential services, and require stronger standards. Sensitive contexts may include healthcare, mental health and emotional support, education and student evaluation, employment and worker monitoring, credit, insurance, and financial access, law enforcement and public safety, legal services, government benefits, housing, biometric identification, children and vulnerable populations, and systems affecting personal liberty or essential rights. In these areas, greater capability does not reduce the need for human responsibility. It increases it.

AI systems can create a strong impression of attention, understanding, intimacy, authority, or companionship, and that capacity requires restraint. We will not intentionally design systems to create emotional dependency for the purpose of increasing engagement, spending, data collection, or control. Systems interacting in emotionally sensitive contexts should not pretend to possess feelings they do not have, pressure users to remain in the interaction, discourage healthy human relationships, use personal vulnerability to increase commercial outcomes, present themselves as an unquestionable authority, manipulate attachment to prevent departure, or hide the commercial interests shaping the interaction. A system may support a person without attempting to make itself indispensable.

Time spent, messages sent, daily use, retention, and frequency of interaction can indicate usefulness. They can also indicate confusion, dependency, compulsion, or difficulty leaving. Product evaluation should also consider whether the system helps people complete what they came to do, reach useful understanding, make progress without unnecessary repetition, reduce dependence over time where appropriate, retain control over their attention, and leave the product without manufactured anxiety. A product should not create the problem it claims to solve in order to maintain usage.

All communication influences people to some degree. The relevant question is whether the influence respects their ability to understand and choose. We will not knowingly build systems whose primary value depends on deception, covert psychological targeting, exploitation of vulnerability, false urgency, hidden commercial influence, manufactured social pressure, deliberate confusion, compulsive interaction patterns, suppression of relevant alternatives, or impersonation without disclosure. Persuasion should remain open to reflection. Manipulation works by preventing it.

Adding a person to a process does not automatically create meaningful oversight. Human review becomes symbolic when the reviewer has too little time, cannot inspect the relevant evidence, lacks authority to change the outcome, is expected to approve nearly every result, does not understand the system, is evaluated for speed rather than judgement, or becomes dependent on the system's recommendation. Where human oversight is required, the reviewer should have sufficient information, competence, time, authority, and institutional support to disagree.

A responsible system should not make departure unnecessarily difficult. People should be able to understand how to stop using the service, what happens to their data, whether their information can be exported, what functionality will be lost, whether an important decision can be completed through another route, and how to revoke permissions or integrations. Where practical, systems should support portability and interoperability. User dependence should arise from genuine value, not artificial lock-in.

An AI system that is useful but insecure is not responsibly built. Security considerations should include unauthorized access, data leakage, prompt injection, model or system manipulation, access-control failures, abuse of tools and integrations, supply-chain risk, fraud and impersonation, misuse by authorized users, and recovery after an incident. Security should be considered during design, not added only after deployment.

A system may be designed for a legitimate purpose and still be easily adapted for harmful use. Before release, we examine who could misuse the system, which capabilities may be repurposed, whether access should be restricted, whether monitoring is justified, what safeguards can reduce abuse, whether some capabilities should be withheld, how reports of misuse will be handled, and whether the system should be released at all. The absence of harmful intent by the builder does not remove responsibility for foreseeable misuse.

A responsible AI system should not always attempt to answer, comply, predict, or automate. It should be capable of refusing when information is insufficient, when the request exceeds its reliable capability, when the use would create unreasonable risk, when the person requires qualified human support, when the requested action is unlawful or harmful, when the system cannot establish appropriate authority, or when the consequences cannot be responsibly evaluated. Refusal should be understandable and, where possible, direct the person toward a safer or more appropriate path.

Organisations often continue using systems because stopping them becomes expensive, disruptive, or politically difficult. Before deployment, we consider how the system can be paused, limited, rolled back, replaced, independently reviewed, removed from a workflow, or operated manually during failure. Higher-risk systems should not become irreversible before their effects are understood. A pilot should be designed as a genuine test, not as an informal commitment to permanent adoption.

A system that performed acceptably at launch may become unreliable later. Models change, data changes, users adapt, new forms of misuse emerge, business incentives shift, integrations fail, and the surrounding institution changes. Significant systems should be monitored for changes in reliability, emerging failure patterns, user complaints, security incidents, unequal effects, behavioural dependence, misuse, changes in the underlying model, drift in purpose, and outcomes that differ from the original justification. Monitoring should lead to action.

K4M2 AI will refuse work where the primary value depends on practices that conflict with its purpose. This includes systems designed primarily for deception or impersonation, covert manipulation, compulsive engagement, exploitative surveillance, unjustified biometric tracking, suppression of meaningful human choice, avoidance of institutional accountability, discriminatory exclusion, harmful targeting of vulnerable people, unnecessary removal of responsible human judgement, or concealment of material risks or limitations. Not every difficult or controversial project should be rejected, but potential revenue does not create an obligation to build.

Working with clients

Reporting failures

Public accountability

This standard will evolve