AI Content Policy

ARTIFICIAL INTELLIGENCE USAGE & CONTENT GOVERNANCE

THANDI AI (PTY) LTD | Registration No: 2025/939429/07 | POPIA Reg: 2025-068149

Information Officer: Seelan Govender | hello@thandi.online

AI Ethics Officer: Seelan Govender (interim)

Version: 1.2 | Effective Date: 26 May 2026

---

1. PURPOSE & SCOPE

This policy governs the development, deployment, and usage of Artificial Intelligence systems within Thandi, specifically for career guidance, educational recommendations, and learner support services.

Applies to:

All AI models powering career recommendations

Natural language generation in reports

Chatbot interactions (future feature)

Image/text analysis capabilities

Automated decision-making systems

---

2. AI SYSTEMS IN USE

2.1 Career Matching Engine

Type: Supervised machine learning (random forest + gradient boosting)

Purpose: Match learners to careers based on academic performance, interests, and labor market data

Training data: 50,000+ historical learner records (anonymized), SAQA occupational database, DHET labor market surveys

2.2 University Admission Predictor

Type: Logistic regression + neural network

Purpose: Estimate probability of admission to specific programs

Training data: Historical admission data from 26 SA universities (2020-2024)

2.3 Bursary Eligibility Classifier

Type: Rule-based system + NLP

Purpose: Match learners to bursary opportunities

Training data: 200+ bursary databases, NSFAS guidelines, corporate bursary criteria

2.4 Content Generation (Reports)

Type: Template-based generation with GPT-4 fine-tuning

Purpose: Create personalized career guidance reports

Training data: Career counseling best practices, educational psychology literature

---

3. PLATFORM-SPECIFIC AI USAGE POLICIES

3.1 Learner-Facing AI

Principles:

Empowerment, not replacement: AI supports human decision-making, does not replace counselors

Transparency: Learners informed they're interacting with AI

Age-appropriate: Content tailored for 14-18 year old comprehension

Safety-first: No advice that could cause physical, emotional, or financial harm

Prohibited Uses:

❌ No medical/psychological diagnosis

❌ No financial advice beyond bursary identification

❌ No encouragement of risky or illegal activities

❌ No collection of sensitive data (sexual orientation, political views, religious beliefs)

❌ No creation of addictive engagement loops

3.2 School-Facing AI (Analytics Dashboard)

Principles:

Aggregated only: Individual learner data never shown to schools

Anonymization minimum: 10+ learners required for any school-level statistic

Purpose limitation: School data only for institutional planning, not learner surveillance

Prohibited Uses:

❌ No individual learner tracking by schools

❌ No teacher performance evaluation based on AI recommendations

❌ No automated disciplinary action based on predicted outcomes

❌ No sharing school data with third parties without explicit consent

3.3 Research & Development AI

Principles:

Anonymization: All research data stripped of identifiers

Consent: Opt-in required for research participation (separate from service consent)

Transparency: Publish research findings openly (where appropriate)

Prohibited Uses:

❌ No using learner data to train general-purpose AI models

❌ No selling anonymized data to third-party AI companies

❌ No psychological experiments without IRB approval

---

4. CONTENT GENERATION GUIDELINES

4.1 Report Content Standards

All AI-generated career reports must:

✅ Include:

Clear statement that recommendations are AI-generated and probabilistic

Confidence level indicators (e.g., "High confidence: 85% match")

Explicit verification requirements ("⚠️ Verify with university")

Multiple alternative pathways (not single recommendation)

Resources for human support (school counselor contact)

✅ Tone & Language:

Encouraging and supportive (not deterministic: "You WILL succeed")

Culturally sensitive (South African context-aware)

Age-appropriate reading level (Grade 9 comprehension minimum)

Gender-neutral career descriptions (unless specific program requirements exist)

Multi-language support (primary: English, secondary: isiZulu, Afrikaans in development)

❌ Avoid:

Definitive statements: "You are guaranteed admission"

Discouraging language: "You are not good enough for this career"

Gender stereotyping: "This career is better for boys/girls"

Socio-economic bias: Assuming financial resources beyond stated constraints

Racial or cultural assumptions: All recommendations based on merit and interest only

4.2 Bursary Content Accuracy

Verification Protocol:

All bursary information cross-checked against provider website (automated monthly)

Human verification of top 50 bursaries quarterly

"Last verified" date displayed on every bursary listing

User flagging system for outdated information

Non-inclusion Criteria:

Bursaries requiring payment to apply (likely scams)

Bursaries with opaque selection criteria (unfair practices)

Bursaries from unregistered organizations

Bursaries with discriminatory requirements (unless legally exempted, e.g., certain transformation bursaries)

4.3 Career Description Content

Source Standards:

SAQA occupational descriptions (authoritative)

DHET career guidance materials (government-endorsed)

O*NET database (international, adapted for SA context)

Industry body input (Engineering Council of SA, HPCSA, etc.)

Update Frequency:

Annual review of all career descriptions

Quarterly updates for high-growth sectors (tech, renewable energy)

Real-time updates for discontinued programs or qualifications

---

5. AI BIAS MITIGATION MEASURES

5.1 Identified Bias Risks

Historical Bias Risk: Training data reflects historical inequalities in education system

Mitigation:

Oversampling: Augment training data with records from quintile 1-3 schools

Fairness constraints: Explicitly penalize model for disparate impact based on school type

Benchmarking: Test model performance across different school quintiles before deployment

Socio-economic Bias Risk: AI may favor careers requiring resources unavailable to low-income learners

Mitigation:

Bursary integration: prominently display financial aid for high-cost pathways

Cost filtering: Allow learners to filter careers by "low-cost entry"

Barrier flagging: Identify careers with expensive requirements (e.g., medical school fees)

Gender Bias Risk: Stereotypical career recommendations based on gender (if provided)

Mitigation:

Optional gender field: Not required for core recommendations

Blind testing: Evaluate model outputs with and without gender data

Gender-neutral language: All career descriptions reviewed for bias

Counter-stereotypical examples: Actively promote careers in non-traditional fields

Geographic Bias Risk: Bias toward urban, Western Cape/Gauteng opportunities

Mitigation:

Provincial filters: Learners select home province for local opportunities

Rural opportunities: Highlight careers viable in rural areas (agriculture, teaching, telecommuting)

Distance learning: Include UNISA and other distance options prominently

Racial Bias Risk: Training data may reflect apartheid-era educational disparities

Mitigation:

Blind processing: AI does not consider race in core recommendation algorithms

Transformation bursary matching: Race considered ONLY for explicit transformation bursaries with consent

Equity audits: Quarterly bias audits across demographic groups

B-BBEE transparency: We are 100% black-owned, committed to transformation

5.2 Bias Testing & Auditing

Quarterly Bias Audits:

Disparate impact analysis: Compare recommendation distributions across school quintiles

Counterfactual testing: Same learner profile, different school location/socio-economic status

Gender parity: Ensure recommendation variance by gender <5%

Rural/urban parity: No systematic preference for urban careers

External Auditing:

Annual third-party algorithmic audit (planned for 2026)

B-BBEE verification includes data transformation practices

Open to Information Regulator audits at any time

User Reporting:

"Bias Report" button in every recommendation

Anonymous flagging of potentially biased content

All reports reviewed within 5 business days

Pattern analysis of bias reports quarterly

5.3 Model Update & Retraining Policy

Retraining Frequency:

Major retrain: Annually with new admission cycle data

Minor updates: Quarterly for new programs/courses

Hotfixes: Immediate for identified bias issues

Data Drift Monitoring:

Track model performance metrics weekly

Alert if recommendation patterns shift significantly

Investigate any demographic group performance degradation

Version Control:

All model versions documented and archived

Ability to rollback to previous version if bias detected

A/B testing for major model changes (limited group first)

---

6. ETHICAL AI PRINCIPLES

6.1 Human-Centered AI

Human in the loop: School counselors/parents must review critical recommendations

Override capability: Learners/parents can manually adjust recommendations

Explanation requirement: All recommendations include reasoning ("Why this career?")

6.2 Beneficence

Learner welfare: All features must demonstrably benefit learners

Harm prevention: Proactive identification and mitigation of potential harms

Positive impact: Track downstream outcomes (university admission rates, learner satisfaction)

6.3 Justice & Fairness

Equitable access: Free tier available for all South African learners

No discrimination: Race, gender, disability, religion never used to deny opportunities

Affirmative action: Actively promote transformation bursaries and historically disadvantaged access programs

6.4 Transparency

Open source: Core recommendation algorithms may be open-sourced in future releases

Explainability: White-box models preferred over black-box where possible

Public documentation: This policy and model cards publicly available

6.5 Accountability

Named responsibility: Seelan Govender accountable for AI ethics

Impact assessments: Algorithmic impact assessments before major model changes

Stakeholder input: Learner, parent, and teacher advisory panels for major features

---

7. PROHIBITED CONTENT & USE CASES

We will NOT generate or allow:

❌ Content promoting illegal activities (drug trafficking, fraud)

❌ Explicit sexual content or sexual career advice

❌ Hate speech or discriminatory content

❌ Violence or self-harm encouragement

❌ Medical diagnoses or treatment recommendations

❌ Financial investment advice (beyond bursary identification)

❌ Political campaigning or partisan content

❌ Religious proselytizing

❌ Content encouraging academic dishonesty

❌ Misinformation about COVID-19, vaccines, or public health

We will NOT permit AI use for:

❌ Surveillance of learners or teachers

❌ Automated grading or disciplinary decisions

❌ Predicting learner failure or "risk" scoring

❌ Commercial advertising targeting minors

❌ Creating behavioral profiles for third parties

❌ Law enforcement investigations

❌ Immigration decisions

❌ Credit scoring or financial services

---

8. USER RESPONSIBILITIES

As a learner/parent using Thandi:

✅ Provide accurate information (garbage in, garbage out)

✅ Use AI recommendations as one input among many

✅ Consult with human counselors, teachers, parents

✅ Verify information independently

✅ Report suspicious or biased recommendations

✅ Understand AI limitations (not infallible)

As a school/educator:

✅ Ensure proper consent obtained

✅ Supervise appropriate use (not during unauthorized times)

✅ Teach AI literacy and critical thinking

✅ Do NOT use AI to replace professional counseling

✅ Protect learner login credentials

✅ Report AI errors or biases promptly

---

9. AI ACCIDENT & FAILURE RESPONSE

If AI produces harmful recommendation:

Immediate: User reports via "Flag Issue" button

Within 2 hours: AI recommendation disabled pending review

Within 24 hours: Human review by qualified counselor

Within 48 hours: Corrected recommendation issued if error confirmed

Within 7 days: Root cause analysis and model update if systematic issue

Communication: Affected user(s) notified with explanation

If AI bias detected:

Immediate investigation by AI Ethics Officer

Bias audit within 5 business days

Model adjustment if bias confirmed

User notification if recommendations may have been affected

Report to regulator if systematic discrimination occurred

---

10. POLICY ENFORCEMENT

Monitoring:

Automated content filtering (profanity, hate speech detection)

Human review of flagged recommendations on an ongoing basis

Monthly AI ethics committee review (Seelan Govender + external advisor)

Consequences of Violation:

User violations (attempting to misuse AI):

Warning for first minor offense

Suspension for repeated violations

Permanent ban for serious violations (hate speech, illegal activities)

System violations (AI produces prohibited content):

Immediate model update

Feature temporarily disabled

Root cause analysis

Public incident report if widespread

---

11. POLICY UPDATES

Version history:

v1.2 (26 May 2026): Aligned with POPIA Framework v1.2; removed beta language

v1.0 (21 Dec 2025): Initial release

Update frequency: Quarterly review, or on any material platform change

Stakeholder input: Policy changes reviewed by learner, parent, and teacher advisory panels

Notification: Material changes posted on thandi.online/legal with 14-day notice

---

12. CONTACT & GOVERNANCE

AI Ethics Officer:

Seelan Govender (interim)

hello@thandi.online

0781298701

Responsibilities:

Oversee AI bias audits

Review flagged content

Approve major model changes

Liaise with Information Regulator on AI matters

Advisory Panel (Planned):

2 Learner representatives (Grade 11-12)

2 Parent representatives

2 School counselor representatives

1 AI ethics academic

1 Disability rights advocate

---

13. COMPLIANCE & STANDARDS

We commit to:

✅ POPIA compliance: All AI processing POPIA-compliant

✅ Ubuntu AI Principles: African-centered AI ethics framework

✅ UNESCO AI Ethics: Following UNESCO AI ethics recommendations

✅ South African AI Blueprint: Aligned with DTIC AI strategy

✅ IEEE Ethically Aligned Design: Technical standards for AI systems

---

14. CURRENT LIMITATIONS

Known limitations:

AI models have limited training data in some career fields

Bias testing is ongoing; not all metrics validated

Human review of recommendations is not yet fully automated

Model explainability features under development

Limited language support (primarily English; isiZulu and Afrikaans in development)

Planned improvements:

Expanded training data

Full bias audit completion

Enhanced human validation workflows

Full isiZulu and Afrikaans support

Enhanced explainability features

---

15. ACKNOWLEDGMENT

By using Thandi, you acknowledge:

AI is a tool, not a crystal ball

Recommendations contain uncertainty

Bias may exist despite our mitigation efforts

Human judgment is essential

You will verify all critical decisions independently

---

Document Version: 1.2

Last Updated: 26 May 2026

Next Review: October 2026

THANDI AI (PTY) LTD

170 Innes Road, Morningside, Durban, Kwa-Zulu Natal, 4001

www.thandi.online | hello@thandi.online | privacy@thandi.online | 0781298701

B-BBEE Level 1 Contributor | 100% Black-Owned | POPIA Reg: 2025-068149