PByarugaba/Luganda-Lusoga-AI-failures-PByarugaba

Luganda & Lusoga AI Failure Cases — Native-Speaker Annotated

A small, annotated set of cases in which current AI systems fail on Luganda and Lusoga — two Ugandan Bantu languages virtually absent from global training data. Each case pairs the original sentence with the machine output, the correct meaning, and a native-speaker analysis of exactly where and why the failure occurs.

The value of this set is not its size. It is that each failure is explained from the inside — by a native-fluent speaker of Luganda, Lusoga, and Lunyoro who is also a 25-year legal and commercial practitioner, and who can therefore say not only that a system is wrong but why, and what correct understanding requires.

Why this exists

These languages are not poorly served by AI because the technology is not sophisticated enough. They are poorly served because the data does not exist. The meaning, cultural knowledge, and contextual intelligence the languages carry have never been written into a corpus a model could learn from.

This set demonstrates the gap rather than asserting it. It is intended as evidence for researchers, funders, and developers building — or evaluating — AI for the roughly one billion people who speak languages in this family, and as a reference point for what native-speaker, domain-expert annotation contributes that general crowdsourcing cannot.

The three cases

The cases are chosen to cover three distinct kinds of failure, not three instances of the same one.

1. Lexical / idiomatic (Luganda). An everyday sentence about a tyre bursting on a drive. Five consecutive errors in a single breath: an idiom read literally, a guessed noun, a landmark tree mistaken for a car part, a polysemous word resolved wrongly, and an onomatopoeia treated as vocabulary. The kind of failure more data might eventually fix.

2. Structural / pragmatic / cultural (Lusoga). A proverb about money in which the subject is dropped after the opening phrase, the meaning is carried by grammatical structure rather than words, and an alliterative pattern distracts a non-speaker from a meaning a native speaker never loses. The kind of failure more vocabulary data will not fix — it is a cognitive-architecture problem.

3. Pragmatic / register / temporal (Luganda). One sentence — "mashed or unmashed?" — that reads as a neutral question, a withheld act of care between spouses, or, in 2006 Kampala street vernacular, "a punch or a slap?". The street register is now obsolete: it was a deliberately constructed, intentionally exclusive, constantly evolving system. This case demonstrates why a "train once, deploy, walk away" approach cannot work for a living language.

Failure taxonomy

Category	Case	What it shows
Lexical / idiomatic	Tyre burst	Literal selection over idiom; missing cultural reference
Structural / pragmatic / cultural	Money proverb	Ellipsis, alliterative distraction, philosophy encoded in grammar
Pragmatic / register / temporal	Mashed or unmashed	Implicature, coded register, deliberate language evolution

Methodology

Each case is annotated by a single native-fluent speaker of Luganda, Lusoga, and Lunyoro. The analysis is not a back-translation exercise: it states the machine output, the literal gloss where one is informative, the correct meaning, and a token-by-token or layer-by-layer account of the failure, with the cultural knowledge each failure requires made explicit.

This reflects the central argument of the work: for low-resource living languages, native-speaker, domain-expert annotation is not a one-time data collection task. It is an ongoing relationship that must be sustained for as long as the language is spoken, because the language does not stop changing when a model finishes training.

Data fields

Each record in data/ai_failures.jsonl contains:

• id — stable identifier
• language_iso639_3, language, region, register
• failure_category
• source_text — the original sentence
• machine_translation — the AI output(s)
• literal_word_for_word — literal gloss where informative (else null)
• expert_correct_meaning — the native-speaker reading
• failure_points — list of {token, machine_reading, correct_reading, failure_type, note}
• cultural_knowledge_required
• stakes — why the failure matters
• annotation_method, annotator_languages, source

Intended uses

• Evaluating Luganda / Lusoga handling in translation and instruction-following models
• A reference for the failure categories low-resource African-language AI must address
• A worked example of native-speaker, domain-expert annotation methodology
• Supporting evidence for funded language-data and cultural-accuracy work

Limitations and scope

This is a small, illustrative set, not a benchmark. It demonstrates failure categories; it does not measure system performance at scale. The cases are drawn from everyday speech, a proverb, and a historical street register — they are not a representative sample of either language.

Orthography for these languages is not fully standardised, and several tokens in the source sentences are written as they appear in the source article; spacing and spelling reflect that source and should be read as one speaker's rendering.

Citation

Byarugaba, P. (2026). Why AI Is Failing a Billion Africans. LinkedIn. Annotations and dataset compilation by the author.

If you use this dataset, please cite it and retain attribution.

Maintainer

Peter Byarugaba — native-fluent in Luganda, Lusoga, and Lunyoro; 25-year legal and commercial practitioner across Uganda and New Zealand. Based in Wellington, New Zealand.

These languages are alive and changing. This set is a starting point, not a finished artifact — which is precisely the point it is meant to make.