Building trustworthy AI systems

If you were asked the following questions in an interview, how would you handle them? I am providing my answers today so you can build on them.

Why is there a concern about trust in AI and Gen AI?

The machine learning model creation approximates the relationship between inputs and outputs. Models are probabilistic by nature. Large language models (LLMs) hallucinate. Sometimes AI and Generative AI can provide incorrect answers. The output can be biased, unethical or even dangerous. AI has the potential to break or make our lives on the planet, depending on what we use it for.

How can we make AI and Gen AI trustworthy?

We must do the following to make AI trustworthy. First, customers must trust that the business will handle their Personal Identifiable Information (PIIs), financial data, health information, employee details, authentication data, legal records, educational records, and government ID details responsibly. Second, users must trust that the data scientist has trained the machine learning model using a genuine, certified training dataset. Third, the AI solution designer must introduce explainability as much as possible so that users can trace the reasons behind the system’s recommendations. Fourth, the LLM provider must create and share model cards for each model, including details on how to use and misuse the AI model. Technical performance metrics and their values must be published.

Fifth, users must continue to follow and refer to the publicly available model leaderboards to understand their strengths and weaknesses. Certain models are well-suited for mathematical computations, while others may be better for coding, reasoning, or answering domain-specific questions. Sixth, the LLM creators must include built-in guardrails when preparing the training dataset. Seventh, developers must include guardrails when building applications using models. Prompt engineering must be conducted responsibly. The applications’ output must be moderated to remove self-harm-promoting, racial, explicit, and abusive content. I guess you can think of a few more along similar lines.

Domains where AI and Gen AI are trusted? Domains where AI is not trusted?

AI is viewed sceptically, especially in healthcare, and for good reason. The cost of errors is very high when they negatively impact someone’s health. AI errors in the retail domain, leading to lower-than-expected sales volume for an item, are something I would worry about a little less for the moment. What are your thoughts?

Who takes responsibility when AI or Gen AI goes wrong and has a severe negative impact?

It is difficult to answer this now. Regulations and accountability distributions are not concrete, at least to my knowledge. When an LLM goes wrong, shall we blame the LLM provider who gave the architecture and predictive power to the model, the developer who built the application leveraging the LLMs capabilities, the regulator who didn’t anticipate the error, or the developer’s employer who didn’t have enough tests done of the application’s characteristics, or the user who unquestioningly trusted the model’s recommendations without doing much homework?

Will AGI increase or reduce our trust in AI? How far are we from AGI?

While we are far away from Artificial General Intelligence (AGI), the definitions and expectations are vague. It is easy to call it an AI model having “Human-Level Intelligence” rather than defining or designing the system. In my opinion, AGI will be more complex than generative AI and will therefore require significant effort to gain public trust.

Are large language models smart?

No.

Trustworthy AI and Gen AI – theory vs. practice

Theory – We are getting there. Practice – no comments.

Can responsible Gen AI catch up with rapid Gen AI innovation?

It is interesting to note that the players on the “Responsible AI” and “AI Innovation” teams are the same. We are trying to catch up with ourselves. My short answer to this is that we are not able to keep pace with innovation.

Do you think we are building explainable AI and Gen AI systems?