Under-representation of local languages in AI

Executive summary

Language shapes thought, behavior and access to services. Kenya is multilingual (dozens of living indigenous languages alongside Kiswahili and English), yet most modern AI/NLP systems are trained on high-resource languages. That mismatch risks excluding large population groups from the benefits of AI-enabled promotive and preventive health tools (e.g., SMS nudges, symptom checkers, health-wallet reminders, behaviour-change messages). Community health systems (CHVs, CHUs) are central to Kenya’s preventative care strategy — bridging AI systems to communities will only work if those systems operate reliably in local languages and cultural contexts. Ethnologue (Free All)+1

---

1. The Kenyan language landscape (why this matters)

Kenya is highly multilingual: Ethnologue and national surveys report dozens of living indigenous languages (over 60), with Kiswahili and English as official lingua francas but many rural populations more comfortable in their mother tongues. This linguistic diversity is an asset for culturally appropriate health promotion — but it also creates practical barriers for one-size-fits-all digital tools. Ethnologue (Free All)

---

2. How mainstream AI models under-represent local languages

Large language models (LLMs) and many commercial NLP systems are overwhelmingly trained on web-scale and English-heavy corpora. Critics and researchers have documented (a) skewed training data, (b) lack of documentation and curation for low-resource languages, and (c) risks (biases, hallucinations) when models are used outside their well-documented domains. Community-driven initiatives (e.g., Masakhane) exist to remedy this gap, but progress remains limited relative to need. Dr Alan D. Thompson – LifeArchitect.ai+1

---

3. Why language under-representation harms promotive & preventive health in Kenya

1. Miscommunication & lower uptake. Health advice that’s poorly translated or only in English/Swahili can be misunderstood or ignored — undermining preventive campaigns (vaccination reminders, ANC attendance, sanitation messages). Local studies and program reviews in Kenya document language-related communication breakdowns that reduce quality and effectiveness. Semantic Scholar+1
2. Inequitable access to AI services. If symptom checkers, triage chatbots, or SMS-based behaviour-change tools only work well in English/Swahili, speakers of other mother tongues are excluded. This deepens urban–rural and socio-economic gaps. CHW Central
3. Poor data quality and feedback loops. Digital health systems (e.g., M-TIBA) can capture rich usage data, but if inputs (free-text, voice, choices) are mismatched to users’ language, analytics and automated recommendations will be biased or noisy. That reduces the value of digital health wallets, reminders and decision-support tools for preventive services. PMC

---

4. Kenyan digital health: a quick reality check

Kenya’s Community Health Strategy 2020–2025 prioritizes promotive and preventive services delivered via Community Health Units (CHUs) and Community Health Volunteers (CHVs). Digital platforms have been piloted successfully (e.g., M-TIBA’s health wallet and real-time claims data in Kisumu) and show how mobile tools can support UHC and preventive interventions — but they work best when aligned with local language and community workflows. CHW Central+1

---

5. Concrete examples of the problem (illustrative)

• SMS behaviour-change campaigns: Generic English messages often have lower comprehension and behaviour change in communities where local language is primary; poorer outcomes are reported where communication is not localized. IR Library
• CHV-supported digital enrolment (e.g., M-TIBA enrolment drives): CHVs often translate, interpret, or rewrite messages during enrolment — a human workaround that increases cost and introduces inconsistency. Scaling requires automated multilingual support that preserves meaning and cultural nuance. PMC

---

6. Practical recommendations (policy + technical + programmatic)

Policy & governance

1. National guidance: Include multilingual standards in Kenya’s digital health regulatory framework and community health strategy — require language accessibility as part of procurement and certification. (Ministry of Health digital health guidance & draft regulations already call for stronger governance; add explicit language/accessibility clauses.) CHW Central+1

Data & capacity
2. Fund community-led data collection. Support grassroots projects (e.g., Masakhane, language documentation efforts) to create open, ethically sourced corpora for Kenyan languages — including voice datasets for low-literacy users. arXiv
3. Document provenance & consent. Ensure that language data collection follows privacy, consent and benefit-sharing rules; tie datasets to community review boards.

Technology
4. Localize AI pipelines. For promotive/preventive tools (SMS, IVR, chatbots, voice reminders), implement a layered approach:

• Off-the-shelf: Kiswahili + English baseline.
• Add targeted mother-tongue modules where needed (Kikuyu, Luo, Kalenjin, Kamba, etc.).
• Human-in-the-loop: CHVs validate automated translations before wide deployment.

5. Use lightweight models at the edge. Deploy compact translation and intent-classification models tuned for local languages (cheaper and easier to audit than large LLMs). Community datasets + transfer learning (from Masakhane–style work) accelerate progress. arXiv

Programmatic
6. Integrate CHVs as co-designers. Train and pay CHVs to curate messages, test local phrasing, and moderate AI outputs. This reduces harm and increases cultural appropriateness. Kenya’s CHS explicitly positions CHVs as central to promotive/preventive care. CHW Central

Risk mitigation
7. Audit & evaluation. Monitor model outputs for bias, hallucinations and unsafe recommendations (the literature on LLM risks highlights the need for careful audit and documentation). Implement feedback loops that let communities flag harmful or confusing messages. Dr Alan D. Thompson – LifeArchitect.ai

---

7. Implementation roadmap (practical, phased)

• 0–6 months: pilot multilingual SMS/IVR for one county (use CHV co-design; collect parallel human translations); include monitoring metrics for comprehension and uptake.
• 6–18 months: expand to 3–5 counties; develop open datasets (text + voice) and fine-tune small models; integrate with M-TIBA/DHIS2 reporting where relevant. PMC+1
• 18–36 months: national rollout of multilingual modules for priority preventive campaigns (immunization, ANC, malaria prevention), accompanied by a regulatory standard and an independent audit mechanism.

---

8. Short conclusion

If AI is to strengthen promotive and preventive healthcare in Kenya, it must speak the languages of the communities it serves — technically and culturally. Investments in community-led data, multilingual models, CHV partnerships and regulatory safeguards are not optional: they are central to equity, safety and the effectiveness of digital health at scale.

---

Selected reputable sources

Note: I also included links and evidence above; the five most load-bearing claims in this note are supported by the citations shown inline.

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21). https://doi.org/10.1145/3442188.3445922. Dr Alan D. Thompson – LifeArchitect.ai

Orife, I., Kreutzer, J., Sibanda, B., Whitenack, D., Siminyu, K., Martinus, L., … Bashir, A. (2020). Masakhane — Machine Translation for Africa (AfricaNLP Workshop, ICLR 2020). arXiv. https://arxiv.org/abs/2003.11529. arXiv

Ministry of Health (Kenya). (2020). Kenya Community Health Strategy 2020–2025. Government of Kenya. (PDF). Retrieved from the Ministry of Health repository. CHW Central

Huisman, L., van Duijn, S. M. C., et al. (2022). A digital mobile health platform increasing efficiency and transparency towards universal health coverage in low- and middle-income countries. Digit Health. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9005819/ (Case/pilot evidence: M-TIBA in Kisumu, Kenya). PMC

Ethnologue. (2024). Kenya — Languages. SIL International. Retrieved from https://www.ethnologue.com/country/KE (summary of Kenya’s living languages). Ethnologue (Free All)

Obel, G. A., Mwangi, D. P. M., & Richard, D. (2020). An investigation of the effect of language barrier on effective communication for the provision of quality healthcare: a case of Kericho County health facilities, Kenya. Scientific Research Journal. (Local empirical evidence on language barriers in Kenyan health settings). scirj.org