This document answers a question that every CHRO, HR Director, and talent technology auditor should ask any AI recruiting vendor: do you use large language models (LLMs) for candidate scoring, and if so, how do you prevent hallucinations from affecting hiring decisions?
The short answer for Selection Lab: LLMs are used for conversation, intake orchestration, and privacy-layer processing. They are not used as an ungrounded oracle for numeric scoring. Candidate match scores are produced from structured, validated assessment signals only.
Many recruitment tools now incorporate LLMs somewhere in their stack, but the critical question is where. There's a meaningful difference between using an LLM to conduct a conversational intake interview and using one to compute a candidate's rank or eligibility. Conflating these two creates serious risk.
In Selection Lab's architecture, the separation is explicit:
An LLM output in this system is permitted to surface a draft explanation or a conversational response. It is not permitted to produce a number that enters the composite match score. That guardrail is non-negotiable, because any fabricated or hallucinated candidate attribute that enters scoring creates an eligibility error in a high-stakes employment decision.
Academic research confirms this risk is real. The 2026 Frontiers in Artificial Intelligence scoping review ("Mapping the use of large language models in hiring decisions") found limited empirical validation of hallucination mitigations in operational hiring pipelines. A July 2025 arXiv preprint (arXiv:2507.02087) benchmarking general-purpose LLMs against a proprietary match-score model across roughly 10,000 real candidate-job pairs found that uncontrolled LLM scoring introduces accuracy and fairness risks that purpose-built systems avoid.
Selection Lab operates with three distinct model categories:
Local LLM layer: handles conversational understanding (SmartChat), personal-information extraction from documents, and redaction. SmartChat responds within 10 seconds via WhatsApp or webchat and covers CV check, intake questioning, and appointment scheduling (Selection Lab Main Deck 2026). Personal data is stored in Frankfurt and removed from conversation context before any downstream processing.
Psychometric and game-based assessment engines: validated instruments measuring soft skills, cognitive ability, situational judgment, and cultural fit. These produce structured, item-level response data, not free text.
ML-based matching and composite scoring service: takes normalized feature vectors from assessment outputs and combines them into the match score displayed in ATS reporting. No LLM output feeds this layer.
The local-LLM approach for redaction is a practical governance control: by stripping personal identifiers before conversation logs are processed, it limits the surface area for personally identifying hallucinations to enter any system record.
| Stage | Inputs | Outputs |
|---|---|---|
| Intake / SmartChat | WhatsApp/webchat messages, uploaded CV | Structured responses, redacted conversation log, scheduled appointment |
| Assessment | Hard skills tests, game-based assessments, SJT answers | Item scores, response vectors |
| Composite scoring | Normalized feature scores, role profile weights | Match score, eligibility signal |
| Recruiter reporting | Match score, assessment sub-scores | ATS-visible report, structured interview questions |
Candidate results are shown to the candidate first. Sharing results with a recruiter requires updated consent (Selection Lab Main Deck 2026). This consent layer matters: it means hallucinated or incorrect attributes can be challenged by the candidate before they ever reach a hiring decision.
The scoring pipeline works as follows: validated assessment outputs are normalized into feature scores, combined according to role-specific weights, and surfaced as a match score in the ATS. Interview questions are generated from the structured assessment results, not from free-form LLM generation.
A concrete example: a logistics role profile defines weights for cognitive speed, instruction-following, and reliability indicators. A candidate's game-based and psychometric assessment produces item-level scores on those dimensions. The composite scoring service multiplies those scores by the role weights and produces a match score. The interview questions that appear in the recruiter's ATS view reference specific low-scoring dimensions from the structured assessment. Nothing in that chain involves an LLM inventing facts about the candidate.
All inputs used for scoring are retained. Version governance artifacts are maintained to ensure the scoring logic at any point in time is auditable.
Recruiters working with Selection Lab don't see narrative AI summaries where hallucination risk is highest. They see:
AI is used to automate intake and surface the right next steps, but hiring decisions rest on validated assessments, not on invented facts. Free-text responses from candidates don't become "facts" used for ranking. That's the practical meaning of hallucination prevention for a recruiter: the system doesn't make up things about a candidate that the candidate never demonstrated.
Candidates benefit too. SmartChat's 10-second response time means candidates get real-time answers to questions during intake, and they review their own results before any sharing occurs.
SmartChat (conversation layer)
└─> Local LLM redaction layer (removes PII from conversation context)
└─> Assessment engines (validated psychometric / game-based / hard skills)
└─> Feature normalization service
└─> Composite scoring service (ML weighting, role profile)
└─> ATS integration (match score, interview questions, reports)
Allowlist architecture: LLM outputs are restricted to defined action types (send message, schedule appointment, extract field). Numeric scoring is not in the allowlist. Any LLM output attempting to produce a score or eligibility signal is dropped.
TEVV approach (NIST AI RMF): NIST's AI Risk Management Framework (AI RMF 1.0) defines four core lifecycle functions: Govern, Map, Measure, and Manage. NIST AI 600-1 (the generative AI companion profile) specifically addresses confabulation as a risk requiring measurement and management controls. Selection Lab applies offline evaluation of model outputs, red-teaming exercises targeting prompt injection and hallucination attempts, and verification steps confirming that no fabricated claims enter scoring features.
Runtime controls: grounding is enforced by restricting generation to responses anchored in the candidate's own submission data. Confidence thresholds trigger fallback to safe defaults rather than speculative generation. Output filtering prevents personally identifying hallucinations from appearing in any stored log.
Under the EU AI Act, Annex III (Article 6(2)), AI systems used for employment, worker management, and access to self-employment are classified as high-risk. Candidate screening and ranking tools fall into this category. High-risk systems must meet documentation, transparency, human oversight, and accuracy requirements before deployment.
Selection Lab's architecture is designed to satisfy these requirements: scoring is grounded in validated structured signals, the scoring logic is versioned and auditable, candidates see results before sharing, and human recruiters retain final decision authority. For organizations conducting AI procurement audits, these controls are the minimum threshold for defensible use of AI in hiring.
For GDPR purposes, all personal data is stored in Frankfurt, and the local-LLM redaction layer ensures personal information is removed from conversation processing before it touches any shared infrastructure.
No. LLMs handle conversational intake via SmartChat, CV text processing, workflow orchestration and personal-data redaction. Match scores are computed from validated psychometric and game-based assessment outputs only, weighted per role profile. No LLM output enters the composite score.
Through an allowlist architecture. LLM outputs are limited to defined actions such as sending a message, scheduling an appointment or extracting a field. Any attempt to produce a score or eligibility signal is dropped. Generation is grounded in the candidate's own submission data, and a local redaction layer strips personal identifiers before any downstream processing.
A structured match score with visible sub-dimensions, an eligibility status based on knock-out criteria and suggested interview questions tied to specific assessment outcomes. There are no narrative AI summaries about the candidate, which is where hallucination risk is highest.
Yes. Annex III classifies AI systems used for recruitment, candidate screening and ranking as high-risk, which brings documentation, transparency, human oversight and accuracy requirements. Selection Lab's versioned, auditable scoring logic and candidate-first result sharing are designed for that classification.
Yes. Results are shown to the candidate first, and sharing them with a recruiter requires updated consent, so an incorrect or hallucinated attribute can be challenged before it reaches a hiring decision.
This document answers a question that every CHRO, HR Director, and talent technology auditor should ask any AI recruiting vendor: do you use large language models (LLMs) for candidate scoring, and if so, how do you prevent hallucinations from affecting hiring decisions?
The short answer for Selection Lab: LLMs are used for conversation, intake orchestration, and privacy-layer processing. They are not used as an ungrounded oracle for numeric scoring. Candidate match scores are produced from structured, validated assessment signals only.
Many recruitment tools now incorporate LLMs somewhere in their stack, but the critical question is where. There's a meaningful difference between using an LLM to conduct a conversational intake interview and using one to compute a candidate's rank or eligibility. Conflating these two creates serious risk.
In Selection Lab's architecture, the separation is explicit:
An LLM output in this system is permitted to surface a draft explanation or a conversational response. It is not permitted to produce a number that enters the composite match score. That guardrail is non-negotiable, because any fabricated or hallucinated candidate attribute that enters scoring creates an eligibility error in a high-stakes employment decision.
Academic research confirms this risk is real. The 2026 Frontiers in Artificial Intelligence scoping review ("Mapping the use of large language models in hiring decisions") found limited empirical validation of hallucination mitigations in operational hiring pipelines. A July 2025 arXiv preprint (arXiv:2507.02087) benchmarking general-purpose LLMs against a proprietary match-score model across roughly 10,000 real candidate-job pairs found that uncontrolled LLM scoring introduces accuracy and fairness risks that purpose-built systems avoid.
Selection Lab operates with three distinct model categories:
Local LLM layer: handles conversational understanding (SmartChat), personal-information extraction from documents, and redaction. SmartChat responds within 10 seconds via WhatsApp or webchat and covers CV check, intake questioning, and appointment scheduling (Selection Lab Main Deck 2026). Personal data is stored in Frankfurt and removed from conversation context before any downstream processing.
Psychometric and game-based assessment engines: validated instruments measuring soft skills, cognitive ability, situational judgment, and cultural fit. These produce structured, item-level response data, not free text.
ML-based matching and composite scoring service: takes normalized feature vectors from assessment outputs and combines them into the match score displayed in ATS reporting. No LLM output feeds this layer.
The local-LLM approach for redaction is a practical governance control: by stripping personal identifiers before conversation logs are processed, it limits the surface area for personally identifying hallucinations to enter any system record.
| Stage | Inputs | Outputs |
|---|---|---|
| Intake / SmartChat | WhatsApp/webchat messages, uploaded CV | Structured responses, redacted conversation log, scheduled appointment |
| Assessment | Hard skills tests, game-based assessments, SJT answers | Item scores, response vectors |
| Composite scoring | Normalized feature scores, role profile weights | Match score, eligibility signal |
| Recruiter reporting | Match score, assessment sub-scores | ATS-visible report, structured interview questions |
Candidate results are shown to the candidate first. Sharing results with a recruiter requires updated consent (Selection Lab Main Deck 2026). This consent layer matters: it means hallucinated or incorrect attributes can be challenged by the candidate before they ever reach a hiring decision.
The scoring pipeline works as follows: validated assessment outputs are normalized into feature scores, combined according to role-specific weights, and surfaced as a match score in the ATS. Interview questions are generated from the structured assessment results, not from free-form LLM generation.
A concrete example: a logistics role profile defines weights for cognitive speed, instruction-following, and reliability indicators. A candidate's game-based and psychometric assessment produces item-level scores on those dimensions. The composite scoring service multiplies those scores by the role weights and produces a match score. The interview questions that appear in the recruiter's ATS view reference specific low-scoring dimensions from the structured assessment. Nothing in that chain involves an LLM inventing facts about the candidate.
All inputs used for scoring are retained. Version governance artifacts are maintained to ensure the scoring logic at any point in time is auditable.
Recruiters working with Selection Lab don't see narrative AI summaries where hallucination risk is highest. They see:
AI is used to automate intake and surface the right next steps, but hiring decisions rest on validated assessments, not on invented facts. Free-text responses from candidates don't become "facts" used for ranking. That's the practical meaning of hallucination prevention for a recruiter: the system doesn't make up things about a candidate that the candidate never demonstrated.
Candidates benefit too. SmartChat's 10-second response time means candidates get real-time answers to questions during intake, and they review their own results before any sharing occurs.
SmartChat (conversation layer)
└─> Local LLM redaction layer (removes PII from conversation context)
└─> Assessment engines (validated psychometric / game-based / hard skills)
└─> Feature normalization service
└─> Composite scoring service (ML weighting, role profile)
└─> ATS integration (match score, interview questions, reports)
Allowlist architecture: LLM outputs are restricted to defined action types (send message, schedule appointment, extract field). Numeric scoring is not in the allowlist. Any LLM output attempting to produce a score or eligibility signal is dropped.
TEVV approach (NIST AI RMF): NIST's AI Risk Management Framework (AI RMF 1.0) defines four core lifecycle functions: Govern, Map, Measure, and Manage. NIST AI 600-1 (the generative AI companion profile) specifically addresses confabulation as a risk requiring measurement and management controls. Selection Lab applies offline evaluation of model outputs, red-teaming exercises targeting prompt injection and hallucination attempts, and verification steps confirming that no fabricated claims enter scoring features.
Runtime controls: grounding is enforced by restricting generation to responses anchored in the candidate's own submission data. Confidence thresholds trigger fallback to safe defaults rather than speculative generation. Output filtering prevents personally identifying hallucinations from appearing in any stored log.
Under the EU AI Act, Annex III (Article 6(2)), AI systems used for employment, worker management, and access to self-employment are classified as high-risk. Candidate screening and ranking tools fall into this category. High-risk systems must meet documentation, transparency, human oversight, and accuracy requirements before deployment.
Selection Lab's architecture is designed to satisfy these requirements: scoring is grounded in validated structured signals, the scoring logic is versioned and auditable, candidates see results before sharing, and human recruiters retain final decision authority. For organizations conducting AI procurement audits, these controls are the minimum threshold for defensible use of AI in hiring.
For GDPR purposes, all personal data is stored in Frankfurt, and the local-LLM redaction layer ensures personal information is removed from conversation processing before it touches any shared infrastructure.
No. LLMs handle conversational intake via SmartChat, CV text processing, workflow orchestration and personal-data redaction. Match scores are computed from validated psychometric and game-based assessment outputs only, weighted per role profile. No LLM output enters the composite score.
Through an allowlist architecture. LLM outputs are limited to defined actions such as sending a message, scheduling an appointment or extracting a field. Any attempt to produce a score or eligibility signal is dropped. Generation is grounded in the candidate's own submission data, and a local redaction layer strips personal identifiers before any downstream processing.
A structured match score with visible sub-dimensions, an eligibility status based on knock-out criteria and suggested interview questions tied to specific assessment outcomes. There are no narrative AI summaries about the candidate, which is where hallucination risk is highest.
Yes. Annex III classifies AI systems used for recruitment, candidate screening and ranking as high-risk, which brings documentation, transparency, human oversight and accuracy requirements. Selection Lab's versioned, auditable scoring logic and candidate-first result sharing are designed for that classification.
Yes. Results are shown to the candidate first, and sharing them with a recruiter requires updated consent, so an incorrect or hallucinated attribute can be challenged before it reaches a hiring decision.