AI Use and Transparency Policy

1. Purpose and Scope

This policy establishes how Qualitect Ltd uses artificial intelligence technologies within the Qualitect platform. It sets out our commitment to transparent, responsible and accountable AI use in the design and development of qualifications. This policy applies to all users of the Qualitect platform and covers interactions with third-party AI service providers used by the platform.

2. Definitions

AI-Generated Content means any text, data or recommendations produced by artificial intelligence models deployed within the Qualitect platform.

Evidence Retrieval means the automated process of identifying and ranking relevant evidence using vector embeddings and machine learning techniques.

Generation Pipeline means the end-to-end process that combines AI models, prompt engineering, evidence retrieval and compliance validation to produce qualification content.

Human-in-the-Loop means the mandatory requirement for human review and explicit acceptance of all AI-generated content before it can be saved or submitted.

Prompt Hash means a cryptographic fingerprint of the instructions sent to AI models, used for audit trail purposes without storing the full prompt content.

Syllabus means a document that sets out the content, structure and requirements of a qualification. Sometimes also known as a specification or curriculum.

Qualification means a formal recognition that a learner has achieved a defined standard of knowledge, skill or competence.

RAG Pipeline means Retrieval-Augmented Generation, the process that combines relevant evidence documents with AI generation to produce contextually informed qualification content.

Vector Embeddings means numerical representations of text that capture meaning and context, enabling the platform to find content that is meaningfully similar rather than relying on keyword matching alone.

3. AI Technologies Used

3.1 Primary AI Systems

Qualification generation uses large language models to produce draft learning outcomes, assessment criteria and indicative content. Evidence retrieval uses vector embeddings to identify relevant documents. Compliance validation support combines AI with deterministic rules engines. An AI assistant feature provides contextual help.

3.2 AI Model Providers

Primary Provider: Anthropic Claude serves as our primary large language model through the Anthropic API.

Fallback Provider: OpenAI serves as our fallback large language model provider, engaged automatically when the primary provider experiences availability issues.

Embedding Provider: OpenAI's text-embedding-3-small model is used to generate vector embeddings of uploaded evidence files. Embedding generation is a one-shot transform per chunk and produces numerical representations that cannot be reversed to the original text.

Reranker Provider: Cohere Rerank is used to improve the relevance ordering of retrieved evidence chunks before they are passed to the primary or fallback LLM. The reranker operates on short text snippets only and does not retain the inputs.

Our AI providers do not use Qualitect data to train their models. Data processed for generation is retained by providers only for a limited period for abuse monitoring under each provider's terms and is then deleted, except where retention is required by law. The same oversight and validation requirements apply to all AI providers used by the platform. Each provider is engaged under a Data Processing Agreement.

3.3 Technical Architecture

All AI processing occurs through encrypted connections using TLS 1.3. The platform maintains an append-only audit trail of every AI interaction including model selection, generation parameters and prompt hashes.

4. Human Oversight and Control

4.1 Mandatory Human Review

Every piece of AI-generated content must undergo human review before it becomes part of a qualification. The platform enforces this at the API level, preventing automated acceptance. Users must explicitly accept or reject each generated element.

4.2 Override and Escalation

Users can override compliance validation findings with documented justification. All override decisions are recorded in the audit trail. Where Platform Users identify AI-generated content that may pose compliance risks, they should report these through the Complaints Procedure.

4.3 Quality Monitoring

The platform continuously monitors quality of AI-generated content. Bias detection operates through multiple mechanisms including verb-level matrix enforcement.

5. Data Handling in AI Processing

5.1 Data Sent to AI Providers

The data sent to AI providers is limited to qualification parameters, prompts and extracts from uploaded evidence. Platform Users should minimise the personal data included in uploaded material; where personal data is included, it is processed as described in this policy and, for customer organisations, the Data Processing Agreement.

5.2 Data Not Sent to AI Providers

User personal data, confidential commercial information and complete evidence files are not sent. Learner data is never transmitted under any circumstances.

5.3 Data Processing Agreements

Qualitect maintains comprehensive data processing agreements with all AI service providers specifying data handling requirements and security obligations.

6. Transparency Measures

6.1 Content Labelling

All AI-generated content is clearly labelled. Generation indicators show the generating model, timestamp and evidence sources.

6.2 Audit Trail Transparency

Comprehensive audit trails of all AI interactions are maintained. Prompt hash recording enables audit capabilities without storing sensitive prompt content.

6.3 Explainability

Users can access detailed explanations of how AI-generated content relates to the evidence base.

7. Bias Prevention and Monitoring

7.1 Structural Bias Controls

Verb-level matrix enforcement ensures appropriate cognitive demand. Subject benchmark alignment requires content to align with established frameworks.

7.2 Content Review for Bias

Human reviewers must consider potential bias including demographic assumptions, cultural bias and accessibility barriers.

7.3 Ongoing Monitoring

Regular analysis of generated content identifies potential bias patterns. Users can report concerns through feedback mechanisms.

8. Limitations and Risks

8.1 Content Accuracy

AI-generated content may contain errors despite compliance validation. Users must exercise professional judgement in reviewing and accepting generated content.

8.2 Compliance Support

The platform provides tools to support compliance but does not guarantee regulatory approval. Generated content requires thorough human review before submission to any relevant regulatory body.

8.3 Technical Risk Management

Fallback systems, error handling and monitoring capabilities manage technical risks.

9. Accountability

9.1 Qualitect Responsibilities

Qualitect accepts responsibility for the design, implementation and operation of AI systems within the platform.

9.2 User Responsibilities

Users retain responsibility for qualification content regardless of whether AI assistance was used. Professional expertise in reviewing AI-generated content remains essential.

10. Alignment with UK AI Principles

10.1 Safety, Security and Robustness

Comprehensive security measures including encryption, access controls and monitoring.

10.2 Transparency and Explainability

Clear information about how AI systems operate, what data is processed and how content relates to evidence sources.

10.3 Fairness and Non-Discrimination

Multiple measures to prevent bias through technical controls, human review and ongoing monitoring.

10.4 Accountability and Governance

Clear accountability frameworks for AI system operation and content quality. Our AI governance framework is designed to align with the requirements of ISO 42001, the international standard for AI management systems and with the broader requirements of ISO 9001 and ISO 21001 as they apply to the quality and educational integrity of AI-generated qualification content. Qualitect is working towards formal accreditation against ISO standards relevant to its operations. References to ISO alignment describe the framework adopted and do not indicate current certification.

10.5 Contestability and Redress

Users can challenge AI-generated content and system decisions through the Complaints Procedure.

11. Roles and Responsibilities

Platform Users are responsible for reviewing and explicitly accepting all AI-generated content and applying professional judgement. Platform Users with quality assurance obligations must include AI-generated content within their quality assurance processes. The Compliance and Accreditation Lead has day-to-day responsibility for this policy. The Development Team is responsible for maintaining the technical AI infrastructure.

12. Related Policies

Privacy Policy, Data Protection Policy, Data and Content Sourcing Policy, Quality Assurance Policy, Complaints Procedure.

13. Contact Information

General: hello@qualitect.co.uk
Data protection contact: dataprotection@qualitect.co.uk
Complaints: complaints@qualitect.co.uk
Registered office: Qualitect Limited, 32 Willoughby Road, London, N8 0JG
Trading address: 8a Stafford Street, London, W1S 4RU
Website: qualitect.co.uk