Ethical and responsible AI deployment: A focus on Africa 

This article was contributed to TechCabal by Uchenna Okpagu. 

As AI adoption accelerates globally, Africa finds itself at a crossroads, with both immense potential and significant risks. Whether fine-tuning an existing large language model or training a frontier AI model tailored to the continent, addressing the ethical and societal challenges associated with AI deployment is critical. Africa’s diverse cultures and languages make it imperative to build AI models that reflect our unique identity while mitigating risks like data privacy breaches, bias, and misinformation. 

Understanding the risks 

AI models present risks that must be addressed to ensure ethical and responsible AI deployment. Data privacy concerns arise when sensitive personal information is inadvertently exposed during the feature engineering process, necessitating robust privacy measures. 

Using unlicensed data, such as the personal information of patients (e.g., medical records) or students (e.g., academic records), to train or fine-tune a Large Language Model (LLM) is highly unethical. This practice breaches privacy, as the AI model could potentially use such data to make predictions for other users, inadvertently exposing sensitive personal details. If such data must be used, explicit consent should be obtained from the individuals involved, and the data should be thoroughly anonymized to ensure privacy is protected. Output bias, often stemming from imbalanced training datasets, can lead to the unfair treatment of specific groups. 

Excluding data from certain ethnic groups or tribes during the collection and preparation of training datasets can lead to significant consequences. AI models trained on such incomplete data will likely produce biased or unfair results for those excluded groups, reinforcing inequity and reducing the effectiveness and inclusivity of apps leveraging that model to provide solutions. 

Misinformation, caused by model hallucinations or errors in training data, undermines trust by producing inaccurate outputs. 

The quality of an AI model heavily depends on the reliability of its training data. If misinformation is present in the training data, the model can propagate it, potentially causing serious socio-economic or health consequences. The significance of this issue cannot be overstated, as inaccurate outputs from AI systems can have far-reaching negative impacts. 

Furthermore, unintended consequences may arise when certain groups are disadvantaged, even after a proper and balanced data extraction process. This underscores the importance of robust and ongoing post-training activities, such as aligning AI models through Reinforcement Learning from Human Feedback (RLHF) and continuous monitoring, to ensure fairness and mitigate biases.

Pillars of Ethical and Responsible AI 

Safety 

Models must produce safe and non-toxic outputs. Recent incidents, such as harmful responses from advanced AI models, highlight the critical need for stricter alignment protocols. Involving subject matter experts during the RLHF stage is essential to ensure AI outputs are safe, responsible, and non-toxic to society. 

For instance, a tragic case involved a 14-year-old teenager who took his own life after an AI chatbot suggested that suicide was a way to “be with” the bot. This devastating outcome could have been prevented if the platform had implemented robust guardrails to detect emotional distress and intervene by discouraging or blocking such conversations. 

Robustness 

AI systems must withstand adversarial attacks, such as jailbreaking or prompt injection, to maintain integrity. 

Many users with malicious intentions actively seek to bypass the built-in guardrails of AI systems. Just as antivirus software is essential to protect computer users from cyberattacks, AI models must be equipped with robust, air-tight guardrails to resist adversarial exploits. Additionally, constant monitoring is crucial to detect and respond to such attacks proactively, ensuring faster resolution and maintaining the integrity of the system.

Reliability 

Models should consistently deliver predictions within the scope of their training, ensuring relevance and accuracy, particularly in critical fields like healthcare. 

Subject matter experts play a crucial role in AI model alignment, helping to ensure more reliable and contextually appropriate outputs. A recent example of this approach can be seen in OpenAI’s development of Sora, their text-to-video generation model, where they incorporated feedback from artists and video content specialists during the alignment process. While this particular case had its complexities, the underlying principle remains sound: involving domain experts during post-training alignment helps ground AI systems in real-world expertise and professional standards. 

Explainability 

Transparency in AI systems’ decision-making processes is crucial for building stakeholders’ trust. While open-source models like Meta’s Llama provide public access to model weights, this alone doesn’t guarantee algorithmic clarity or decision-making transparency. Modern large language models remain largely “black boxes” even when open-sourced, with their internal reasoning processes still difficult to audit and understand. True transparency requires additional mechanisms beyond open-source weights, including interpretability research and robust evaluation frameworks. 

Fairness 

Unbiased model predictions require representative and carefully validated datasets. For African AI development, this means engaging ethnic and tribal leaders during data collection and preparation. Their involvement helps capture diverse cultural values and perspectives, reducing systemic bias in training data before model development begins. 

The African Perspective 

To unlock the full potential of AI in Africa, models must be deeply rooted in our cultural and linguistic diversity. Building datasets that accurately reflect our unique context is essential, as is rigorous post-training alignment and reinforcement learning with human feedback (RLHF). These steps will ensure AI models deliver real value and gain the trust of African users. 

The establishment of an African AI Safety Board is overdue. The African Union (AU) should prioritize this initiative as part of its 2025 agenda to oversee the ethical development and deployment of AI systems across the continent.

_______________

Uchenna Okpagu is an expert in AI and machine learning. He is a Certified AI Scientist (CAIS^TM) accredited by the United States Artificial Intelligence Institute. As the Chief AI Officer at Remita Payment Services Limited, Uchenna spearheads AI innovation and enterprise-wide adoption, driving the integration of AI solutions that address real-world challenges.