AI Ethics & Responsible Use

[TODO: A plain-language description of the actual evaluation mechanism — what the model receives, what it compares a response against (behavioural indicators / evidence markers), and what it outputs. Written so a non-technical HR reader understands it and a technical reviewer cannot call it hand-waving. Equally important: what the model does not receive — e.g. whether demographic attributes, names, or photographs are part of the input at all. Source: product + engineering.]

[TODO: State where and how a candidate is told they are interacting with an AI system rather than a person — the exact placement and wording of the in-session disclosure, not just its presence on this website.]

[OPEN QUESTION — flagged for the product owner, needs answering before this page ships: EU AI Act Article 50 transparency obligations became enforceable 2 August 2026 and apply to chatbot-style interactions. ApriCOT’s AI Role-Play is exactly that pattern. Confirm whether the product currently surfaces this disclosure in-session, at the start of a role-play, or only here on the marketing site. If it is website-only today, that is a product gap to close, not a copy gap to write around.]

[TODO: requires verification with whoever ran the studies before publishing. content-faqs.md states that pilot studies compared ApriCOT’s AI-based evaluations against traditional human-led assessments and showed “a high degree of alignment.” This is the most valuable claim in the entire content set and currently the least substantiated. Before it appears here it needs: sample size, population and roles assessed, the comparison method, the alignment measure actually used with its value (correlation, agreement rate, or similar), who conducted the study, and when. A claim of this weight without methodology is the single biggest liability on this page — publish the numbers or do not publish the claim. Source: whoever ran the pilot studies, plus legal review.]

[TODO: How the system is bias-tested and monitored for adverse impact — the methodology, which protected characteristics are examined, the statistical test applied (e.g. four-fifths rule / impact ratio), how often testing runs, whether it is conducted internally or by an independent third party, and whether results are made available to clients. Phrase findings in terms of “tested for” and “monitored for adverse impact” — do not describe the product as bias-free or unbiased anywhere on this page; see the language note in this file’s header comment. Source: data science + legal.]

[TODO: Be specific about where a human is in the loop and where one is not — the two are equally important and buyers assume the worst when only the first is stated. content-faqs.md says ApriCOT “can be used in a hybrid mode, combining AI-driven assessments with human oversight where needed.” Needs: what hybrid mode concretely means, whether it is default or opt-in, who the human reviewer is (client-side assessor or ApriCOT-side), what they can see, and whether they can override a generated rating. Also state plainly what happens in non-hybrid mode — i.e. where no human reviews the output before it reaches a decision-maker. Source: product + delivery.]

[TODO: This section is a feature, not a disclaimer — stating limits builds more trust with an enterprise buyer than claiming none. content-faqs.md already establishes that ApriCOT complements rather than replaces psychometric tools and human assessors; start there. Needs the full list of things the product does not claim to do: e.g. it does not make hiring or termination decisions on its own, it does not infer protected characteristics, it does not diagnose or make any clinical or psychological determination, it does not assess candidates outside the competency framework it is configured against. Confirm each item is actually true of the product before listing it. Source: product + legal.]

[TODO: Which models are used and whether they are third-party or in-house; crucially, whether client or candidate data is ever used to train or fine-tune any model, and if so under what consent. “Your data is not used for training” is one of the most-asked questions in enterprise AI procurement — answer it directly. Also cover: model version control, the change-management process when a model is updated, and whether clients are notified of changes that could affect scoring. Source: engineering + legal.]

[TODO: What a candidate is entitled to and how they exercise it — notice that AI is being used, an explanation of how their result was reached, a route to appeal or contest it, and a route to request human review. State which of these ApriCOT provides directly and which are the client organisation’s responsibility to operate, since in most deployments the candidate’s relationship is with the client, not with ApriCOT. Cross-reference the Privacy Policy for data rights, which are distinct from these evaluation rights. Source: legal + product.]

[TODO: legal review — do not assert compliance with any regime the company has not actually been assessed against. Describing an obligation is safe; claiming to meet it is not. The regimes that matter for this product:]

• EU AI Act — employment-related assessment is classified high-risk; those obligations apply from 2 December 2027. Article 50 transparency duties for chatbot-style interaction are already enforceable (2 August 2026) — see section 2.
• NYC Local Law 144 — requires an independent annual bias audit of automated employment decision tools, publication of the audit summary, and advance notice to candidates. [TODO: confirm whether ApriCOT is used for NYC candidates at all; if so this is an operational requirement, not a future one.]
• EEOC guidance (US) — adverse impact analysis expectations for selection procedures. Ties directly to section 4.
• [TODO: confirm whether any other regime applies given actual client geography — e.g. UK, EEA member-state employment law, or Indian equal-opportunity obligations.]

[TODO: Who owns AI ethics internally — a named role or committee, its composition, what it reviews and how often, and what authority it has to block a release. A buyer is checking whether accountability sits with an identifiable person or with nobody. If no formal structure exists yet, that is a decision to make before this page ships rather than something to describe vaguely. Source: leadership.]

[TODO: How a candidate, client, or researcher raises a concern about an assessment outcome or the system’s behaviour — the contact route, what happens after a report is received, the acknowledgement and resolution timelines, and whether anonymous reports are accepted. Keep this distinct from the DPDP grievance route, which covers data handling rather than evaluation fairness; link across rather than merging them. Source: support + legal.]