AI Systems Disclosure
Last updated: April 2026
PublicMoodTracker is an AI-powered platform. Artificial intelligence is central to how we collect, process, and present political intelligence. This disclosure document describes all AI systems in production use, our human oversight mechanisms, our bias mitigation measures, and your right to explanation — in accordance with the EU AI Act (2024), emerging Kenyan AI guidelines, and our own ethical commitments.
1. AI Systems in Production
| System | Model | Version | Purpose | Risk Level |
|---|---|---|---|---|
| Sentiment Classifier | XLM-RoBERTa-large | Phase 13.1 | Classifies text as POSITIVE / NEGATIVE / NEUTRAL about a named politician | Medium |
| Named Entity Recognition | BERT multilingual (fine-tuned) | v3.2 | Identifies Kenyan politicians, counties, parties in text | Low–Medium |
| Topic Modelling | BERTopic | v0.16 | Discovers emerging political themes from unlabelled text corpus | Low |
| Spike Detection | Statistical (Z-score + IQR) | v2.1 | Flags abnormal mention velocity for real-time alerts | Low |
| Spam / Bot Filter | Logistic regression ensemble | v4.0 | Filters low-quality, duplicate, or bot-generated content before scoring | Low |
| AI Insights Generator | GPT-class LLM (API) | Rotated | Generates plain-language summaries of sentiment trends for admin dashboard | Low (human-reviewed) |
2. What AI Does — and Does Not — Do
2.1 What AI Does
- Reads publicly available news and social media text and assigns a sentiment direction (positive/negative/neutral) for named politicians.
- Identifies which politicians and locations are mentioned in each piece of content.
- Calculates aggregate sentiment scores weighted by model confidence.
- Detects statistical anomalies in mention patterns that may indicate breaking news or coordinated campaigns.
- Suggests thematic labels for trending political topics (e.g., "Finance Bill protests", "Cabinet shuffle").
2.2 What AI Does NOT Do
- Generate quotes: AI never fabricates or invents statements attributed to politicians. All quoted material comes from the original source.
- Predict elections: Sentiment scores are not electoral forecasts. We explicitly prohibit framing them as such.
- Assess truthfulness: AI does not fact-check political claims or assess whether a politician's statements are true.
- Make autonomous decisions: No automated system makes consequential decisions about individuals without human review.
- Profile private individuals: Our NER and sentiment systems only process public figures in their public roles.
3. Human Oversight and Governance
3.1 Editorial Review
A human editorial team reviews AI-flagged anomalies weekly. Any score that deviates by more than ±0.5 from a 7-day baseline for a major politician triggers a manual review to assess whether the shift reflects genuine discourse or a data quality issue.
3.2 Bias Audits
Independent bias audits are conducted quarterly, assessing for:
- Regional bias (does the model systematically score Nairobi-based politicians differently from Turkana-based ones at equivalent media coverage?).
- Gender bias (are female politicians scored differently for similar language used about them?).
- Ethnic/tribal framing (does the model respond differently to ethnically coded language?).
- Source concentration bias (does heavy reliance on one source distort national aggregates?).
Audit findings and remediation actions are published in our Annual AI Transparency Report.
3.3 Model Cards
Each AI model we deploy has a model card documenting: intended use, training data, evaluation metrics, known failure modes, and ethical considerations. Model cards are available to researchers on request at ai@siasaiq.com.
4. Labelling of AI-Generated Content
All AI-generated content on PublicMoodTracker is clearly labelled:
- Sentiment scores display an "AI-computed" badge and link to this disclosure.
- AI Insights (auto-generated summaries) are labelled with the 💡 AI icon and a disclaimer that they are analytical summaries, not factual claims.
- Spike alerts are labelled "System-detected spike" to distinguish them from editorial news alerts.
5. Known Bias and Fairness Issues
| Issue | Severity | Status | Mitigation |
|---|---|---|---|
| Sheng slang under-representation | Medium | Ongoing | Quarterly lexicon updates; Sheng community review panel (planned Q3 2026) |
| Satire misclassification rate ~3% | Low | Ongoing | Satire signal features added to Phase 13.1; target <1.5% by Phase 14 |
| Urban/online source bias | Medium | Ongoing | Community radio transcripts being integrated (planned Q4 2026) |
| Gender bias in Swahili political language | Low | Monitored | Gender-balanced training examples added; quarterly audit in place |
6. Your Right to Explanation
Under Article 22 GDPR and KDPA Section 26, you have the right to request a meaningful explanation of any automated output that affects you or an entity you represent.
For any politician profile, click the 🔍 transparency icon to see:
- The source articles and social posts that contributed to the current score.
- Individual mention scores and confidence values.
- The time window applied.
- The model version used.
For a formal written explanation of a specific score, contact ai@siasaiq.com. We will respond within 10 business days.
7. Responsible AI Commitments
- We will not sell sentiment data for political campaign targeting or micro-targeting.
- We will not allow our platform to be used to suppress political participation or manufacture disinformation.
- We will publish an annual AI Transparency Report with bias audit results, model performance metrics, and policy updates.
- We will engage with Kenyan civil society and academic institutions to improve our models and methodology.
- We will comply with any future Kenyan AI regulatory framework within 12 months of enactment.
8. Contact
AI ethics and transparency: ai@siasaiq.com
Model card requests and research access: research@siasaiq.com