AI Transparency

Last updated: 21 October 2026

1. Where we use AI

All AI features use Claude Haiku 4.5, a large language model made by Anthropic, called through Anthropic's commercial API. We don't train or fine-tune our own models.

2. Where we don't use AI

• Reading files. Text is pulled from PDF and Word files with standard parsing libraries, not AI. Scanned PDFs with no text layer can't be read. They're listed as unreadable, not silently dropped.
• Fallback scoring. If the AI service is unavailable, CVs are scored with a keyword method (TF-IDF similarity plus a check for the job's key terms). These scores are rougher. Every score in your results is labelled AI score or Keyword score, so you can tell which is which. Keyword scores aren't saved, so running the match again retries them with the AI.
• The shortlist cut-off. You set the minimum score. Candidates below it aren't hidden: they're listed separately with their scores so you can review them.

3. What the AI doesn't do

• It doesn't decide who is hired, interviewed, or rejected. Those decisions belong to the recruiter.
• It doesn't email, message, or contact candidates.
• It doesn't learn from your data. Each score is a fresh request, and no model is trained on your CVs or job descriptions.
• It isn't asked to judge personality, emotion, or other personal traits. It's asked only how well the CV's skills, experience, and qualifications fit the role.
• It's told to ignore the candidate's name, gender, age or date of birth, photo, ethnicity, nationality, religion, marital or family status, and disability. It's also told not to guess age from graduation or job dates and not to mark down employment gaps. Work rights, location, and licences count only when the job description asks for them.

4. Known limitations

AI scores are a useful first sort, not a verdict. Keep these in mind:

• Very long CVs are cut off. Anything after the first 12,000 characters isn't read, so skills listed late in a very long CV can be missed.
• The model sees the whole CV. We don't strip personal details out before scoring, so the model still reads the candidate's name and anything else they included. It's told to ignore those details (see section 3), but language models can carry biases from their training data, and an instruction reduces that risk without removing it. Treat the score as one input and read the CV before you rule anyone out.
• Wording matters. A candidate who describes a skill in unusual terms, or whose experience comes from another industry, can score lower than their real fit.
• Scores can shift. Rewording a job description can move a CV's score by a few points. Comparing candidates within one role is more reliable than comparing scores across roles.
• Mistakes happen. The model can list a skill as matched or missing in error. The skill lists are there so you can check its reasoning quickly.
• Other languages. CVs and job descriptions in other languages are scored on meaning, but scoring has been tested most in English.

5. How AI data is handled

• CV and job text is sent to Anthropic over an encrypted connection only to produce the score. Under Anthropic's commercial terms, API data isn't used to train their models. See Anthropic's Privacy Center.
• We store each score with its skill lists, so the same CV isn't sent to the AI again for the same role. Stored scores are deleted when you delete the CV, the job, or your account.
• If we change the model, the scoring instructions, or how much text is read, stored scores are treated as out of date and recalculated.
• Each agency's CVs, jobs, and scores are kept separate and are never used to score another agency's candidates.

For everything we collect and how long we keep it, see our Privacy Policy.

6. Your responsibilities as a recruiter

You make the hiring decisions and remain responsible for them. We recommend you:

• Read the CV, not just the score, before you shortlist or reject someone, especially near your cut-off.
• Tell candidates you use AI-assisted screening, for example in your job ads or privacy notice. Some places require this. Under the EU AI Act, for example, AI used to screen job applicants is classed as high-risk.
• Only upload CVs you have a lawful basis to process. See section 6 of our Privacy Policy.
• Give candidates a way to ask for a human review of a decision and to have their data deleted.

7. If you're a candidate

If a recruitment agency screened your CV with CV Matcher, that agency controls your data and makes the decisions. Contact them first about access, correction, deletion, or a human review. If you can't reach them, or you have a concern about how CV Matcher's AI works, email [email protected] and we'll help where we can.

8. Changes and contact

We update this page whenever we change the AI model, what it reads, or the features that use it. The date at the top shows the latest revision.

Current setup: model claude-haiku-4-5 · scoring version 2 · CV limit 12,000 characters · job limit 8,000 characters.

Questions about how we use AI: [email protected]