How can you responsibly implement AI software in critical applications?

it's crucial for a product to have a proper use case and fit. It's not wise to incorporate trendy and viral engineering innovation into a product just for the sake of it, especially if it doesn't benefit your business or solve customer problems. Recently, I've noticed that many consumer applications have integrated GPT use-cases into their core product. While ChatGPT is undoubtedly cool and useful, in most cases, the GPT feature is either irrelevant or poorly executed in a way that's not customer-friendly. Most of these products assume that customers are willing to write prompts to find what they want. While writing a good prompt is the right way to use ChatGPT, forcing similar behavior across other consumer apps is not thoughtful!

The requirements and constraints should include measurements for success: When is the software good enough? What type of failures do you accept? Without these questions use cases stay mainly tests to see what's possible. We know what's possible, but it remains unsolved what is fit for purpose.

In addition to the use cases, due focus needs to be given to the target audience. Definitions of what is acceptable varies by demographic and the AI should know how to ensure it respects the beliefs and boundaries of the target demographic. Trying to please all demographics will unnecessarily cripple the AI, it is better to focus on your primary demographic and build some basic guardrails for the rest.

All of these steps should be a joint effort of developers, business owners and users. This requires good communication and mutual understanding. Include standards beyond engineering: management system to ensure quality and enforce a strong governance for decision making.

This is something all AI teams should tell up front to the business team requesting a solution to a problem. Most of the time the business team or the clients would be coming into a meeting with the AI team riding on the hype created by the latest AI model from the top research companies. I think its one of the responsibilities of AI team to show the stakeholders what AI can and cannot achieve with current technology and most importantly with the current data we have. And sticking to the best practices of software development will make it easy to transition an experimental module in AI to be easily deployed and integrated into the rest of the application.

This is a key consideration, reducing technological bias and misinformation. However, generally, the documentation and guidance for Best Practice techniques are fundamentally still being defined locally and regionally. The practice of following this step in the process is vital, even moreso as AI evolves.

Ethics and social mores are unique and ever changing. Yes, we should build some basic guard rails in AIs to prevent exploitation beyond the intended purpose but lets not try to make AIs human. Lets understand the limitations of AIs and build human involvement to ensure that ethical requirements are met.

Ethical principles and values should be included in the engineering process. If a solution cannot be responsible used it is no need to start building it. As a part of the development cycle businesses are flexible to incorporate risk measures in time for building a responsible and sustainable solution.

One thing that always helps in any type of software development is to looping in the stakeholders and end users, this will guide our development in the right direction and help prioritize the features. In the AI use case, this feedback would help create the right metrics for evaluation and monitoring the model's after the deployment.

In order to communicate, talk in terms of the business. What can the solution do and more importantly: what not? What are the parameters to tune, in other words what are the choices to make and what is the impact on the results and performance.

If you cannot understand what you're doing: just don't. A fool with a tool is still a fool. Engineers should help their stakeholders and users to build the knowledge and understanding of both the results and their role in the complete AI lifecycle.

Explainability is the single most important feature every AI must have. It is also one of the tougher features to build. Thus, we need to put in the necessary effort to ensure AI is explainable. Without explainability it is impossible to evaluate if a particular bad response from the AI can be truly fixed. Without explainability, the best you can do it put filters on the output.

AIs today are in many ways like children. They need constant supervision and correction, not just in research labs but also in live environments. Ensure you are allocating resources to monitor the output and finetune the AI as necessary.