AI transparency

Why this page exists

We build products that use artificial intelligence. We think anyone affected by one should be able to find out, in plain language, what the technology is doing, what it is not allowed to decide, and how to challenge it.

This statement covers our own products and the principles we apply to everything we build for clients. It is written to be readable by the people using our software, not only by their lawyers.

Our commitments

• Disclosure at the point of use. When you are interacting with AI or reading AI-generated content, the interface tells you there and then. Not once, buried in a policy.
• No impersonation of a person. We do not build systems designed to make someone believe a machine is a human being.
• A human decides what matters. Where an output affects money, safety, employment, learning outcomes or access to a service, a person reviews it and can overturn it.
• Explainability. If a system scores, ranks or matches you, it shows the basis in terms you can understand.
• Contestability. There is always a route to a human being who can look again.
• Your content is not training data. We do not allow client or end-user content to be used to train third-party models.
• Measured before it ships. AI features carry evaluation sets, documented failure modes and production monitoring.

Where AI is used in our products

Riff

Riff uses models to match people on their answers, values and communication style, and to produce a compatibility score with a breakdown of why it was reached. It also offers AI companions — conversational personas available while real matches are found. These are labelled as AI throughout, and they are not presented as real users. Safety systems use automated detection to flag harmful behaviour; enforcement decisions that affect a person's account involve human review.

TeachWise AI

TeachWise generates and sequences learning material to build a curriculum around a stated goal. Generated content is checked for coherence and prerequisites before a learner sees it. Learning material can contain errors; it supplements rather than replaces qualified teaching, and learners are told the material is AI-generated.

Project controls (in development)

Our project controls application uses AI to read progress documents, reconcile them against a baseline, highlight variance and draft reporting narrative. It does not approve a forecast, a change or a payment. Every figure remains traceable to its source document and every report is signed off by a named person.

In our own work

We use AI assistance in writing code, drafting documentation and analysis. A human reviews everything before it reaches a client or production, and confidential material is never placed in tools outside an agreed processing arrangement.

How these systems work

Our products use large language models and related machine learning systems provided by established third parties, accessed through their commercial APIs under terms that prohibit training on our data. We add retrieval over relevant content, structured prompting, tool use and validation of outputs before anything is shown to a user.

We select models on capability, safety record and contractual terms, and we can change provider. Where the choice of provider materially affects how personal data is handled, that is disclosed in the relevant product privacy notice.

Limitations you should know about

These systems have real and well-documented weaknesses. Being clear about them is part of using them responsibly.

• They can be confidently wrong. A fluent, well-structured answer is not evidence of a correct one.
• They can reflect bias present in training data, which can produce unequal outcomes across groups. We test for this and monitor it, and we do not claim it is solved.
• They are not deterministic. The same input can produce different output.
• They have knowledge limits and may be unaware of recent events unless given current information.
• They are not a professional. Nothing our products generate is legal, medical, financial or safety-critical advice.

Human oversight

Oversight is designed in rather than promised. For each AI feature we record what it does, the risk if it is wrong, who reviews the output, what that person can see in order to review it meaningfully, and what happens when they disagree.

Reviewers can override an output, and overrides are logged and fed back into evaluation. A feature that cannot be meaningfully overseen is not a feature we ship in a consequential context.

How we govern changes

Prompts, models, retrieval sources and guardrails are version-controlled and change-managed like any other code. A change to an AI component runs against the evaluation set before release, and a regression blocks the release.

Each AI feature has documentation covering its intended use, the data it draws on, its known limitations, its evaluation results and its oversight design. Clients receive this for features we build for them.

We follow the direction of travel in UK regulatory guidance and the EU AI Act — risk-based classification, transparency for systems people interact with, human oversight for high-risk uses, and technical documentation — and we assess each product against the obligations that apply to it.

Your choices

You can ask whether AI was involved in an output that affects you, ask for the reasoning behind it, ask a human to review it, and object to processing as described in our GDPR statement.

To do any of that, or to report an output you believe is wrong or harmful, email Contact@Riff-Apps.com. Reports of harmful output are treated as defects and investigated.