If you consider rolling out an AI-based facial recognition solution, we recommend committing to doing so ethically.
First, at the very beginning of your project, you are likely to face a choice: choose a readily available solution, go the custom route, or opt for library-based development. To make the right decision, weigh the options against the objectives you intend to achieve — the more specific the task, the higher the need for custom software. If, on the other hand, you develop a facial recognition solution aimed at the general public, it may be a faster option to implement an off-the-shelf solution or API, for instance, Microsoft Face API or Amazon Rekognition, or build on an existing facial recognition library, for example, DeepFace, FaceNet, InsightFace, or others.
Another critical aspect is asking people for informed consent for collecting and storing biometric data, as well as for other purposes, like using one's photos to train the algorithms further. While some facial recognition systems can de-identify the information, biometric data can hardly be wholly anonymized, so timely informing people is essential to maintain trust and transparency.
One more aspect to keep in mind is ensuring your solution is explainable. A user should understand why a system has come to a particular decision and revert it in case of false positives or false negatives. For example, rolling out a facial recognition solution, National Australia Bank intentionally chose to refer all user verification requests that the system could not verify to a human operator rather than rejecting them, which allowed reducing the error rate.