Last updated: 24 April 2026

1. Why we publish this

MagicScreen (a product of TechMagic) uses artificial intelligence to assist recruiters and hiring teams in screening candidates. AI used for hiring is one of the most consequential applications of AI: it affects people's livelihoods, and it operates in an area covered by strict equality, employment, and data protection law.

We believe candidates and customers deserve to understand how our AI works, where it is limited, and what safeguards are in place. This Statement documents our approach. It complements our Privacy Policy, Candidate Privacy Notice, and Data Processing Agreement.

2. What MagicScreen does

MagicScreen is a decision-support tool for recruiters. It helps with:

  • structuring screening interviews (text, audio, or video);

  • evaluating candidate responses against the criteria set by the customer;

  • producing summaries, skill assessments, and recommended rankings;

  • surfacing notable strengths, risks, and follow-up questions for the recruiter.

MagicScreen does not make hiring decisions. Every output is intended for human review. Customers are contractually responsible for ensuring a qualified human reviews AI outputs before taking any action that affects a candidate (see Section 3 of the Terms of Service).

3. How our AI works

3.1 Models we use

MagicScreen uses a combination of:

  • Large Language Models (LLMs) provided by leading third-party model providers, used to generate summaries, answer questions, and draft assessments. The current providers are listed in our Subprocessor List;

  • Speech-to-text for transcribing audio/video responses, where the customer has configured audio or video screening;

  • Classifiers and scoring heuristics developed by MagicScreen for structured evaluation against customer-defined criteria.

The specific providers used at any given time are listed in our Subprocessor List.

3.2 What the AI is given

For each screening, the AI is given:

  • the job role and screening criteria configured by the customer;

  • the candidate's CV or application responses (if provided by the customer);

  • the candidate's responses during the MagicScreen session.

The AI is instructed to evaluate responses against the customer's criteria and to produce a structured assessment. It is not instructed to consider, and is prompted against using, characteristics protected under the UK Equality Act 2010 or equivalent law (including race, sex, age, disability, religion, sexual orientation, marital status, or pregnancy).

3.3 What the AI is not given

  • Inferences about protected characteristics are not requested and are prompted against.

  • Candidate data is not combined across customers.

  • The AI does not have access to the internet and cannot independently look up information about a candidate.

4. Human oversight

AI outputs in MagicScreen are recommendations for a human, not decisions. Our product is designed to enforce this:

  • every AI-generated summary, score, and recommendation is presented alongside the underlying evidence (candidate responses, CV excerpts) so the reviewer can verify it;

  • the platform requires a human action (accept / reject / advance) before any status change is recorded;

  • rejection flows prompt the reviewer to provide a reason, improving auditability;

  • customers can configure confidence thresholds below which AI outputs must be flagged for additional review.

Customers remain the decision-makers and the controllers of the screening process.

5. Data, training, and improvement

  • We do not train our AI models on Customer Data or candidate data. Prompts sent to third-party model providers are configured with training and retention-limiting controls (e.g. zero-data-retention APIs where available).

  • Aggregated, anonymised usage data (which features are used, where errors occur, how long screenings take) is used to improve the Service.

  • Feedback from customer users (thumbs up/down on AI outputs, corrections) is used in aggregate to evaluate model performance but is not used to train the underlying LLMs.

Details are in Section 6 of the DPA.

6. Fairness and bias

6.1 Our approach

AI trained on human-generated text can reflect and amplify societal biases. We take this seriously. Our measures include:

  • Prompt design that instructs the AI to evaluate responses on skills, experience, and behaviours relevant to the role, and to disregard demographic or personal characteristics.

  • Standardised evaluation criteria configured by the customer, applied uniformly to every candidate in a given workflow.

  • Output format constraints that require structured justifications tied to the candidate's own responses.

  • Periodic fairness evaluations using representative test sets to measure differences in scoring and recommendation outcomes across demographic groups (where such groups can be lawfully evaluated).

  • Model provider oversight: we select model providers that publish responsible-AI practices and we monitor for material changes.

6.2 Our limitations

  • AI outputs are probabilistic and can be wrong.

  • Bias testing is a statistical exercise and cannot guarantee individual-level fairness.

  • The customer controls screening criteria, question design, and thresholds — those inputs strongly influence outcomes.

  • Accents, dialects, and languages with less training data may be transcribed or interpreted with lower accuracy.

6.3 What customers should do

  • Calibrate questions and criteria to the role, reviewed by someone with employment and equality law knowledge.

  • Use MagicScreen alongside other assessment methods, not as the sole basis for decisions.

  • Review AI outputs before acting on them.

  • Monitor outcomes over time for signs of adverse impact on any protected group.

  • Provide candidates with a meaningful alternative channel where required.

7. GDPR Article 22 — automated decisions

GDPR Article 22 grants data subjects the right not to be subject to a decision based solely on automated processing (including profiling) that produces legal or similarly significant effects — such as decisions about employment.

MagicScreen is designed to be used with meaningful human review and therefore, as used in accordance with our Terms, does not constitute solely automated decision-making. Specifically:

  • AI outputs are recommendations presented to a recruiter, not decisions;

  • the platform requires a human action before any status change;

  • rejection reasons are captured and auditable;

  • candidates can request human review through the controlling customer (see Candidate Privacy Notice).

Customers are contractually prohibited from configuring the Service to make solely automated decisions with legal or similarly significant effects without implementing appropriate Art. 22 safeguards.

8. Candidate rights

Candidates whose data is processed through MagicScreen can:

  • receive information about the use of AI in the screening they participated in;

  • request human review of any assessment produced with MagicScreen;

  • express their point of view and contest a decision;

  • access, correct, or delete their personal data, subject to the customer's retention obligations.

These rights are exercised with the customer (the controller). MagicScreen will cooperate as the processor.

See the Candidate Privacy Notice for a candidate-facing version.

9. EU AI Act

AI systems used for "recruitment or selection of natural persons" fall within Annex III of the EU AI Act and are classified as high-risk. MagicScreen is developing its compliance posture in line with the Act, including:

  • a quality management system for AI development;

  • risk management, data governance, technical documentation, and logging requirements;

  • transparency and human oversight obligations;

  • post-market monitoring and incident reporting;

  • conformity assessment, where applicable.

A full EU AI Act conformity statement will be published ahead of the relevant provisions taking effect. Customers in the EU acting as deployers of a high-risk AI system have their own obligations under the Act (including human oversight, DPIA input, and candidate information); we support customers in meeting them.

10. Security of AI systems

The security measures applicable to the wider Service (see Annex 2 of the DPA) also apply to AI pipelines. Additionally:

  • prompts and outputs are logged (with access controls) for audit and incident response;

  • we monitor for prompt-injection and adversarial-input patterns;

  • model provider selection considers data-handling commitments and security posture.

11. Changes and contact

We may update this Statement to reflect changes in our AI systems, model providers, or applicable law. Material changes will be summarised at the top of the page.

Questions: hello@techmagic.co (or the DPO address if one is appointed).