In media coverage of AI, the term “algorithm” is frequently mentioned, often serving as a catch-all phrase to encapsulate the complex mechanisms that drive AI systems. This widespread usage underscores the pivotal role algorithms play in the functionality of AI technologies, from curating personal news feeds to making autonomous driving decisions. However, despite its prevalence in discussions about AI, many people lack a clear understanding of what an algorithm actually is. This gap in understanding can lead to misconceptions about how AI works and its implications.

For example, recent news stories on algorithmic bias in facial recognition technology highlight the need for greater transparency and understanding of these systems. Reports from reputable sources such as The New York Times, The Guardian, and BBC have shed light on how algorithms, despite being perceived as objective, can perpetuate biases present in their training data. These stories illustrate the critical importance of demystifying algorithms, not only for the sake of informed public discourse but also to foster awareness about the ethical considerations and challenges in AI development.

To keep it as simple as possible, consider an algorithm as a set of instructions or steps that a computer follows to solve a problem or complete a task. Remember those complex algebraic expressions that we learned in high school? Something like the following:

F = (P – (P x R)) + (P – (P x R) x T

An algorithm can make sense of the different parts of the equation to determine the value and then, based on that value, the computer follows a set of instructions. The algebraic expression above determines the final price (F) of a product by consider the original price (P), discount rate (R) and the current tax rate (T).

For another example, we might show the computer a bunch of pictures of different fruits like apples, oranges, grapes, and bananas. We then set up the equation that helps AI to identify the different fruits based on specific characteristics like shape, color, size, texture, taste, firmness, and so on. Once the computer determines the value based on the combination of characteristics, it will then make a decision to identify the fruit in a different picture.

The previous examples are very simple explanations of the brain behind how AI systems work, making decisions based on learned data.

In the case of Google Gemini, the algorithm’s effectiveness was compromised due to the limitations of the data used in its development. This situation can be understood using our earlier simplified example involving fruits. Remember, the algorithm is designed to recognize various fruits from images. If the data used to train this algorithm consists exclusively of images that meet very specific guidelines – such as only pictures of fruits taken under certain lighting conditions, from certain angles, or only including fruits of a particular size – the algorithm’s ability to accurately recognize fruits under different conditions becomes limited.

This limitation is comparable to what happened with Google Gemini, where the data used to develop the algorithm did not adequately represent the diverse conditions under which the algorithm was expected to operate. As a result, the algorithm could perform well only under the specific conditions represented in the training data but failed to generalize to other situations. Thus, an algorithm can become skewed or biased based on the data it is trained on.

The flaw in the algorithm due to skewed data underscores the importance of regulatory agencies requiring that algorithms undergo several assessments, including thorough documentation of the results. These assessments aim to ensure that algorithms are not only effective but also fair, unbiased, and capable of performing accurately across a wide range of conditions and scenarios. By documenting the development process, the data used, and the results of various assessments, developers can identify potential biases or limitations in their algorithms and make necessary adjustments. This process helps in creating more robust, reliable, and equitable AI systems that can be trusted to perform as intended in diverse real-world applications.