African language models have an input economy

GSMA notes that Africa is home to more than 30% of the world's languages while leading AI models remain concentrated in English and other high-resource languages. “Add local languages” can then appear on a product roadmap as if language were a software toggle.

Language performance comes from inputs: permitted text and speech, people who transcribe and label it, compute, evaluation and a process for repairing errors. Each input has an owner and a price. The result is an economy before it is a model feature.

The work is paid.

Rights precede collection

Publicly accessible text is not automatically free for model training. Copyright, database rights, website terms, privacy and cultural expectations can apply. A company needs a documented basis for collecting and using each dataset.

Speech introduces identifiable voices and context. Consent for a research recording may not cover commercial training. A participant should understand purpose, retention, sharing and withdrawal under the applicable process.

Community archives can contain material governed by institutions or customs beyond individual consent. Engagement needs time and the possibility that access will be refused. Scarcity is not permission.

Create a dataset register with source, rights, language, region, time period, collection method and restrictions. If a licence ends, the company should know which model or evaluation set used the data.

Collection has a sampling problem

A language name can cover dialects, locations, age groups, professions and code-switching. Collecting one urban broadcast archive may produce clean audio and narrow representation.

Define the use case before the sample. A call-centre model needs phone audio, product vocabulary and frustrated customers. An agricultural assistant needs local crop and weather terms. A general corpus may support pretraining without proving task performance.

Device and connectivity shape who contributes. An app-based collection campaign can exclude people without smartphones or cheap data. Field collection broadens reach and adds training, travel, equipment and safeguarding.

Record who is missing. Do not describe one dataset as national speech because it contains several thousand clips from one channel.

Annotation is paid judgment

Transcription converts audio into text under a spelling convention. Annotation can add intent, entities, sentiment, safety or preferred answers. Ambiguous examples need adjudication.

Workers require language and subject knowledge. Compensation should include training, difficult cases, quality review and the time needed to use secure tools. Paying per item can reward speed over accuracy unless the design accounts for complexity.

Annotator disagreement is information. It may reveal an unclear task, dialect difference or a question with no single answer. Forcing consensus can erase that signal.

Protect workers from harmful content and surveillance. Limit personal data, provide support for distressing material and avoid collecting productivity metrics without need.

Compute turns data into a recurring bill

Training and fine-tuning require processors, storage, data movement and engineering. Currency and network cost matter when infrastructure is bought abroad. Local compute can reduce some transfer and latency issues while depending on power and available hardware.

A smaller model adapted to a defined task can cost less than training a foundation model. Retrieval from approved documents may solve a knowledge problem without changing model weights. Architecture follows the job and risk.

Inference becomes the operating bill after launch. Text, speech and image workloads consume different resources. Quantisation, caching and batching can reduce cost and change quality or latency.

Report compute boundary and method before making energy or cost claims. A model card with no hardware, run or dataset description cannot support comparison.

Evaluation belongs to each language and task

Translate an English benchmark and it may test translation choices rather than natural use. Build evaluation examples from the target task with independent rights and review.

Measure more than average accuracy. Break results by dialect, code-switching, audio quality and critical intent. A high overall score can hide dangerous performance on account closure, medicine dosage or credit terms.

Human evaluation needs a rubric and multiple reviewers where judgment matters. Report disagreement and sample size. Do not tune repeatedly on the final test set and continue calling it independent.

A product also needs operational metrics: abandonment, escalation, correction, complaint and harm. Laboratory quality is one gate.

Correction keeps the supply chain open

Language changes, products change and users find errors. Give them a reporting path that captures context with consent. A correction should enter triage, root-cause analysis and retesting.

Some failures come from missing knowledge, others from speech recognition, intent routing or a poor policy. Retraining the model is not the universal repair. The owner needs to identify the layer.

Publish known limitations in language users understand. If a service works only for a subset, say so. Silent fallback to English can exclude the person at the moment support is needed.

Model updates can regress a language while improving another. Keep versioned evaluation and the ability to roll back.

Put the ledger beside the launch

The input ledger should show rights and collection cost, contributor compensation, annotation, quality review, compute, evaluation, hosting and correction. Separate one-time build from recurring operation.

Benefits need equal specificity. Which task improves, for whom, against which baseline and at what error cost? Reach in another language is useful when the service completes the job.

GSMA's work identifies data, compute, talent and policy as foundations for African-led AI. Those nouns turn into contracts and labour. A language model can carry local value only if the people and institutions supplying its language are visible in the economics.

Sources

• GSMA, 16 June 2026: The Mobile Economy Africa 2026
• GSMA, July 2024: AI for Africa: Use cases delivering impact