Beyond the Eurocentric reporting and data imbalances we just unravelled, AI models face a deep structural array of algorithmic biases, categorised by how they enter the pipeline. There are basically two types: System or Behavioural Biases and Data-Level Biases.
Types of Algorithmic Biases Faced by AI
Data-Level Biases: When the data mirrors pre-existing human prejudices, historical inequalities, or culturally skewed worldviews, it is called Historical Bias or Societal Bias. For example, if historical corporate recruiting data shows a massive imbalance favouring male hires, an AI trained on it may naturally downgrade resumes containing words associated with women.
Then there’s Representation Bias or Sampling Bias, which occurs when specific populations, perspectives, or edge cases are omitted or heavily underrepresented in the dataset. A classic example is facial recognition software trained primarily on lighter skin tones, causing it to perform poorly on individuals with darker skin.
And finally, Measurement Bias. This happens when the tools, metrics, or methods used to log data are fundamentally flawed or favour specific outcomes. For instance, if a predictive policing algorithm measures “crime” based solely on arrest numbers rather than actual criminal acts, it over-indexes heavily monitored and minority neighbourhoods.
System and Behavioural Biases: When an AI assistant has a strong behavioural bias toward validating the user’s beliefs or stating what it thinks the user wants to hear, rather than sticking to objective truth, it is called Sycophancy Bias.
Then there is Aggregation Bias, which occurs when an algorithm applies a blanket generalisation about an entire group to a specific individual, overriding their unique metrics.
Finally, we have Automation Bias, which creates a human-to-AI feedback loop. Humans tend to overtrust automated systems and stop scrutinising their outputs, assuming computer logic is inherently neutral. If biased outputs are treated as facts, they feed back into the internet and pollute future training cycles.
Conclusion
AI models inherit the biases, gaps, and skewed perspectives embedded within their training data. We’ll learn about the techniques engineers use, such as RLHF (Reinforcement Learning from Human Feedback), and how human feedback, including prompts, is used in attempts to strip systemic bias from AI.
Day 2 / 30: Tomorrow in Part 3, we transition from the problems to the solutions. We’re opening up the hood on RLHF to see how ChatGPT & Claude try to fix toxic internet data—and why it sometimes backfires.
What's your take? Have you noticed Al models giving overly agreeable or shortcut responses instead of deep reasoning? Let's discuss below!