How we use AI
KallosSim uses AI to generate training cases and assess conversations. This page explains exactly what AI generates at runtime, what humans design in advance, and the limitations you should know about as a practitioner. There is no pre-approved scenario library in the live app.
- Training cases Background, situation, and persona narrative for each session. Generated fresh from your card-deck selections, sector knowledge packs, and documented system prompts — not from a fixed catalogue served to every user.
- Conversation turns The persona's responses during dialogue practice, constrained by prompt rules (persona integrity, equity, statutory framing).
- Reflection scores Principle-based feedback across dimensions such as cultural sensitivity, participation, and professional authority. Each score is grounded in retrieved statutory guidance — the relevant source is cited alongside the feedback.
- Evaluation pipeline Scores are produced by a four-step process: evidence extraction, parallel criterion scoring, cross-validator review, and narrative synthesis. This reduces hallucination and improves consistency.
- Perception analysis How specific statements may have been experienced by the simulated person.
- Reflection journal A structured SCIE-style supervision summary generated after each completed session.
- Card-deck taxonomies Practitioner roles, practice dimensions, training focuses, process stages, and cascade filters — the parameters you select before generation. These are maintained by sector practitioners; they are not individual pre-written cases.
- Assessment rubrics Criteria, definitions, and 0–3 indicators used to score each conversation. Grounded in UK statutory frameworks. A four-axis behavioural summary (Legal Compliance / Relational Engagement / Autonomy Support / Professional Curiosity) sits alongside criterion scores.
- System prompts & guards Documented instructions for case generation, persona play, evaluation, PII handling, and case adjustment — including mandatory equity rules and evidence-first scoring requirements.
- Knowledge retrieval Professional literature (e.g. Working Together 2026, Children Act 1989, Care Act 2014, Equality Act 2010, ACAS) retrieved to ground generation and feedback — when the knowledge index is populated for your environment.
- Post-session oversight (sampled) Editorial QA reviews flagged evaluations; harm reports and user feedback trigger prompt or policy updates. This is not a human read-through of every case before you practise.
How we prevent stereotyping
AI models reflect patterns in their training data. Without explicit constraints, they reproduce stereotypes — including linking a person's name, ethnicity, or cultural background to specific problem types.
We address this at three levels:
Prompt-level equity rules
Every generative prompt includes a mandatory rule: a persona's name or cultural background must not explain or determine their situation. If swapping their identity would change the nature of their problem, the output must be rewritten.
Selection-driven context
Cultural and structural factors enter a case only through your card-deck choices and the generation prompt — not because the AI infers ethnicity or background from a name alone. Prompts forbid using protected characteristics as causal explanations.
Reporting & QA (not pre-approval)
We do not human-review every generated case before you see it. Users can report problematic output; editorial QA samples flagged evaluations. Rubrics and prompts are human-maintained and versioned when models change.
Despite these measures, AI can still produce outputs that reflect bias. If you notice a problematic output, please report it via the feedback option in the app or email hello@kallossim.com.
Limitations
AI scores are not clinical assessments
Reflection scores are pedagogical tools — they surface patterns for reflection, not verdicts on your competence. They should always be interpreted with a supervisor or educator.
Every session is AI-generated
Cases and personas are synthetic constructs created at runtime. They are useful for practice but can be unrealistic, inconsistent, or occasionally inappropriate despite guardrails. Treat them as simulations — not replicas of real people or cases.
Simulated personas are not real people
Personas simplify the complexity of real human responses. Practice builds skill — it does not replace experience with real service users.
AI responses may be inconsistent
The same input can produce different responses across sessions. This mirrors real-world unpredictability, but it also means scores and feedback are not perfectly reproducible.
Professional judgment takes precedence
KallosSim supports learning — it does not replace statutory guidance, supervision, or your professional registration requirements. Always apply your own judgment and your employer's policies.
Data and privacy
Session transcripts are stored linked to your anonymised user ID — not your name or email. We do not use your practice conversations to train AI models. You can export or delete all your data at any time from your account settings.
Reflection scores and session history are visible only to you and, if you are part of an organisation account, to your designated team lead. They are never shared with third parties or used for employment decisions.
For full detail, see our Privacy Notice. For GDPR enquiries contact dpo@kallossim.com.
Which AI models do we use?
KallosSim uses OpenAI's GPT-4.1 and GPT-4.1-mini models via the OpenAI API. Case generation and persona conversation use GPT-4.1. Lighter tasks — scoring, perception analysis — may use GPT-4.1-mini to reduce latency.
We do not use AI image generation for personas. Persona illustrations are SVG composites built from human-authored data fields, not generated faces or photographs.
We review our AI provider choices regularly and will update this page when we make changes.
Questions or concerns?
If you have questions about how AI is used, want to report a problematic output, or have concerns about equity and representation in the platform, we want to hear from you.
hello@kallossim.com