As artificial intelligence continues to reshape the digital landscape, a new challenge is emerging for content creators globally, and potentially with unique implications for Kenyans: perfectly human-written content being incorrectly flagged as AI-generated. This misidentification can lead to a host of problems, from damaged credibility to platform penalties. But why is this happening, and what does it mean for the vibrant and diverse content coming out of Kenya?
The concern isn’t just theoretical. Across the internet, writers and creators report their original work being mislabeled by AI detection tools. While these tools aim to uphold authenticity, their current limitations can inadvertently penalize genuine human expression, especially when it doesn’t fit the mold of their training data.
Kenyan English is too funny because you could genuinely write something right now, with a pen, then type it and it would be flagged as AI generated https://t.co/QPtJSHvNWL
— Theo (@Theo_mwangi) May 6, 2025
The Kenyan Hand in Training Early AI: The OpenAI/Sama Engagement
A significant chapter in this story involves OpenAI, the creators of ChatGPT. To make their early models less prone to generating harmful, biased, or toxic output, OpenAI contracted Sama, a San Francisco-based firm with operations in Kenya. Starting around late 2021, Sama employed Kenyan workers to label and filter vast quantities of text data. This data, reportedly sourced from the “darkest recesses of the internet,” included graphic descriptions of violence, hate speech, sexual abuse, and other disturbing content. The task of these Kenyan workers was to identify and categorize this content, effectively teaching the AI what to avoid.
However, investigations, notably a detailed exposé by TIME magazine in early 2023, revealed troubling aspects of this arrangement. Kenyan data labelers were reportedly paid very low wages, with figures around $1.32 to $2 per hour cited (though Sama contested the exact figures, suggesting higher potential earnings for some). More critically, workers described the profound psychological toll of being constantly exposed to horrific content, often with what they felt was inadequate mental health support. The work was so traumatic that Sama reportedly ended its contract with OpenAI for this specific type of content moderation in February 2022, earlier than planned.
While direct court cases against OpenAI specifically for this Kenyan data labeling work haven’t been as high-profile as those involving other tech giants, the situation highlighted the ethical considerations and labor practices within the global AI development pipeline. It’s important to note that Sama has faced legal challenges in Kenya concerning its work with other major tech companies like Meta (Facebook), with content moderators suing over alleged poor working conditions, unfair dismissal, and inadequate psychological support. These cases, though distinct, underscore the broader scrutiny on the treatment of content moderators and data labelers in the region.
The “Unintended Consequence”: AI Detectors vs. Kenyan English
Now, could the involvement of Kenyan data labelers in training initial AI models have inadvertently led to AI detection tools now misidentifying authentic Kenyan content?
This hypothesis suggests that if AI models were significantly exposed to linguistic patterns (even if the task was filtering rather than style generation) from a large pool of Kenyan workers, and if AI detection tools are not adequately trained on diverse global Englishes, including Kenyan English, a mismatch could occur.
It’s crucial to clarify that the primary task of the Kenyan labelers for OpenAI was to filter harmful content, not necessarily to teach the AI to write in a specific Kenyan style. However, the broader issue of algorithmic bias in AI detection tools is well-documented. These detectors are predominantly trained on massive datasets of text, which often overrepresent standard American or British English.
Kenyan English, with its unique syntax, vocabulary, idiomatic expressions, and influences from Sheng and various local languages, presents distinct linguistic characteristics. Research, like the Liang et al. (2023) paper cited by UC Berkeley’s D-Lab, has shown that AI detectors are more prone to misclassifying text written by non-native English speakers (NNES) as AI-generated. This is often because the linguistic structures used by NNES, while perfectly correct and natural within their context, may appear less “predictable” or fit less neatly into the patterns the AI detector expects from its training data, leading to a false positive.
Therefore, while the direct impact of the Sama/OpenAI data labeling on ChatGPT’s writing style making it “Kenyan” is debatable it is also likely not the primary cause, the situation does highlight a critical point: Kenyans were involved in the AI supply chain, yet the tools now used to police AI-generated content may not be adequately trained to recognize the authenticity of their own diverse linguistic expressions. The problem likely lies more with the narrow training data of the AI detection tools rather than the generative AI models learning to “write like Kenyans” from the filtering task.
Worldcoin and Regulatory Scrutiny: A Parallel
While distinct from AI content flagging, the recent legal challenges and High Court ruling against Worldcoin in Kenya offer a relevant parallel. Worldcoin’s project, involving iris scans for a digital ID, faced intense scrutiny over data privacy, informed consent, and regulatory compliance, culminating in a court order to halt its data collection and delete previously gathered biometric data pending a proper Data Protection Impact Assessment (DPIA).
This case demonstrates Kenya’s increasing assertiveness in regulating new technologies and safeguarding digital rights. It sets a precedent for how the country might approach other emerging tech issues, including the potential harms caused by flawed AI systems, such as biased content detectors that unfairly penalize local creators.
The journey of AI development is intertwined with global human effort. The experiences in Kenya, from data labeling for foundational models to the challenges creators now face with detection tools, highlight the urgent need for more inclusive, equitable, and linguistically aware AI systems. Ensuring that the technologies designed to foster authenticity do not inadvertently penalize or erase the diverse voices of Kenyans and others globally is a critical task for the entire tech ecosystem.
Discover more from Techspace Africa
Subscribe to get the latest posts sent to your email.


