Intelligence,
handled carefully.

1. Why this page exists

Machine-learning systems can accelerate useful work — and they can quietly cause harm at scale. We do not treat “AI” as a marketing decoration. Where we use it, we do so with intent, with named principles, and with a clear line of human accountability. This page is our public commitment to those principles.

We align our practice with widely-recognised frameworks such as the OECD AI Principles, the NIST AI Risk Management Framework (AI RMF 1.0), the UNESCO Recommendation on the Ethics of Artificial Intelligence, the EU AI Act, and, in India, the emerging obligations under the DPDP Act 2023 and the guidance being developed by MeitY. Where these frameworks require more of us than the strict letter of the law, we apply the higher bar.

2. Our five principles

2.1 Human oversight

Every consequential decision made by or with a machine-learning system inside our products is subject to human oversight. In practice:

• Model outputs that materially affect a person’s access to a service, safety, earnings, or reputation are gated by a human review step, not shipped straight to the user.
• Where a fully-automated decision is unavoidable, we describe it plainly to the user, and we provide a route to human review under Article 22 GDPR, the DPDP Rules, and equivalent laws.
• We name an internal owner for each production model — a person, not a team — who is answerable for its behaviour.

2.2 Privacy by construction

We treat training data and inference data with the same discipline we apply to any other personal information.

• We only train or fine-tune on data we have a lawful basis to use, minimising personal data where possible and preferring synthetic or licensed corpora.
• User conversations, generated content, or model prompts are not used to train third-party foundation models by default. Where a user opts in to help us improve a feature, the opt-in is explicit and revocable.
• We prefer providers whose data-processing terms guarantee that our inference traffic will not be used to train their models.
• Where a feature can run against a local, on-device model without loss of quality, we prefer that.

2.3 Transparency

Users have the right to know when they are talking to a machine.

• Any product surface where an AI system authors, ranks, or moderates content is labelled clearly, in-product, in plain English.
• Where a suggestion, translation, or summary comes from a model, we disclose that fact next to the output — not buried in a policy page.
• We publish an AI Use Register inside each product describing which features rely on which model families, at what layer of the stack.
• We do not simulate a human identity to gain a user’s trust.

2.4 Fairness

A model that only works well for the majority of users is a broken model. Our commitments:

• We evaluate models against the actual population of users a product will reach — including language, gender, region, and accessibility axes — and we do not ship a system that materially under-performs for a protected group without a documented reason and a mitigation plan.
• We test models against known bias benchmarks appropriate to the modality and log the results.
• Where a moderation model is involved, we make the reasoning visible to users on request, and we provide a clear appeal route.
• We do not build automated ranking or matchmaking that uses caste, religion, or sensitive personal attributes as an input.

2.5 Security

Machine-learning systems introduce new attack surfaces — prompt injection, data extraction, model theft. We treat those as first-class security concerns.

• We validate and sanitise inputs to model calls, and we constrain what a model is permitted to do downstream (no unbounded tool access, no unreviewed code execution).
• Model outputs that trigger side effects (sending a message, making a payment, changing a setting) are subject to authorisation checks that do not depend on the model itself.
• We monitor for anomalous prompts, data-exfiltration patterns, and known prompt-injection attack shapes.
• Our responsible-disclosure programme — see the Security page — covers our AI systems as well as our conventional surfaces.

3. Things we will not do

• We will not use models to make decisions that materially affect a person’s liberty, livelihood, health, or safety without a human accountable for that decision.
• We will not build systems that identify individuals in public spaces without their consent, or that generate biometric inferences (emotion, sexuality, political opinion) from voice, video, or text.
• We will not use synthetic voices or faces to impersonate a real person without their explicit, revocable consent.
• We will not generate content designed to deceive users about the source of the content (deep-fake defamation, synthetic evidence).
• We will not use dark patterns, engagement optimisation, or personalisation that we would not be willing to describe openly to the affected user.

4. Accountability

• The founder personally holds this policy. If any commitment on this page is not being met, the founder is answerable.
• We keep an internal register of our production models: purpose, data sources, evaluation results, and named owner. Extracts from that register are published inside the product that uses the model.
• We aim to publish an annual AI transparency note as our deployment matures — short, honest, comparable year to year.

5. Reporting a concern

If you believe one of our AI-driven features has behaved unfairly, unsafely, or in a way inconsistent with this page:

• Write to connect@nxtspacelabs.com with the subject line “Responsible AI”.
• Include the product, the feature, the input that produced the output, and what the correct behaviour would have looked like.
• We aim to acknowledge these reports within three business days and to investigate them in good faith.