Model monitoring must continue after deployment because input distributions, feature-target relationships, and model performance can change in production.
2
Explainability and monitoring help organizations adopt complex machine learning models in regulated settings where legal and compliance teams need to understand their behavior.
3
Responsible AI work starts with questions about whether machine learning is appropriate, whether data represents affected groups, and whether performance remains fair across users.
Summary
Krishnaram Kenthapadi explains why model monitoring belongs inside a broader responsible AI and governance process. Before deployment, teams validate models and inspect their explanations. After deployment, they need to watch for data drift, changes in feature-target relationships, performance degradation, feedback loops, and differences across user groups. He argues that monitoring matters in every sector because model failures can affect business results, while explainability and fairness become especially important in areas such as financial services, healthcare, and hiring. The conversation also covers the social work behind responsible AI. Teams need input from customers and people affected by their systems, along with diversity of perspectives inside the organization. Kenthapadi recommends starting with feature and data explanations, then adding monitoring that can alert teams when problems emerge. He is candid that business deadlines can conflict with ethical work, although better fairness and monitoring can sometimes improve business performance too.
Monitoring continues after a model reaches production
Krishnaram Kenthapadi says customers first focus on validation and explainable AI before deployment. Once models are live, they want to know whether the models behave as intended, whether the input distribution has drifted, and whether the relationship between target variables and input features has changed. He compares this shift with the way security moved from an afterthought to a standard concern after major breaches. Training and testing a model before release is therefore only one part of the work. Teams need to keep checking models in production because real-world conditions can change.
Complex models make explanations necessary for people inside and outside the team
Kenthapadi describes the move from simple rule-based systems to deep learning models and linear models with large numbers of parameters. Even model developers may not understand how the trained model behaves. He separates understanding the training method from understanding the trained model itself. Stochastic gradient descent is easy to state, while the resulting deep learning model can be difficult for humans to interpret. Explainable AI helps with this tension because complex models can perform well while remaining hard to understand. In future systems with people working alongside models, explanations may become a prerequisite for human decisions.
Outside scrutiny is useful, and companies should investigate before it arrives
Kenthapadi says researchers and auditors probing commercial models may create an uncomfortable situation for the company involved, but those checks provide needed balances. He advises organizations not to wait for regulation, legal action, or academic researchers to expose problems. He gives examples involving image search, autocomplete, facial analysis APIs, and the Apple credit card, where outputs or decisions appeared to differ across demographic groups. His recommendation is to test models thoroughly before release and to repeat that work after deployment. Problems can be found even when a company believes its model is not publicly visible.
Responsible AI needs participation from the whole organization and affected users
Kenthapadi says leadership may already accept the value of explainability and governance, while engineers and data scientists have not yet adopted those practices as part of their own work. He wants responsible AI to become a shared expectation rather than a top-down instruction. The people affected by a model should also help shape its design. He describes work at LinkedIn on fairness in candidate ranking, where the team listened to enterprise customers and LinkedIn members. He calls this a socio-technical problem. Teams need customer input, diverse perspectives, and varied lived experiences to find blind spots.
The right questions span the full machine learning lifecycle
Kenthapadi recommends asking whether a problem should use machine learning at all, whether the data represents intended uses and user groups, and whether the dataset contains measurable or correctable bias. During training and validation, teams should compare predictive behavior across groups rather than looking only at overall performance. A/B tests should also check whether improvements reach all relevant user groups. After deployment, teams should watch for degrading performance and feedback loops that can make a system biased or produce other harmful behavior. He points to tutorials, case studies, and responsible AI materials as ways for practitioners to learn which questions to ask.
Ethically sourced data requires consent, representation, and fair treatment
The discussion of datasets goes beyond whether data can technically be collected. Kenthapadi says teams need to ask whether people consented to their images, text, or biographies being used to train models. Facial analysis and speech systems also need data that represents different demographic groups and user experiences. Building such datasets can require developing trust with communities, and compensation alone may not persuade people to share biometric data. The conversation also raises the treatment of crowdsourced workers who label large benchmark datasets. Kenthapadi says responsible collection, consent, representation, and bias remain a work in progress.
Small organizations can begin with feature, data, and production monitoring
For a new organization, Kenthapadi suggests starting by understanding which features matter most to a model's predictions. This can reveal undesirable features, help data scientists improve the model, and support complaint or regulatory investigations. Data explanations can show which examples contribute negatively, helping teams find mislabeled data. He also recommends adding monitoring early so teams receive alerts when something goes wrong instead of relying on scheduled retraining or occasional checks. Earlier detection can reduce lost revenue or profit, which may repay the time and money spent on monitoring.
Business pressure changes the order of investment, but the goals can overlap
Kenthapadi acknowledges that deadlines and business goals can conflict with work on monitoring, explainability, and fairness. Leadership has to decide that responsible behavior matters alongside near-term metrics. When teams need a direct business case, monitoring is often the easiest starting point because model degradation can affect revenue or profit. Explainability can help with debugging and model improvement, while fairness may depend on regulation, social responsibility, or external scrutiny. He also says the goals can reinforce each other. Improving performance for groups that receive less benefit may improve overall model performance and business results.
"I think this is still very much a work in progress, how to collect datasets with the consent of users, how to make sure that these datasets represent all potential stakeholders, and how to ensure that these are free from different types of biases."Krishnaram Kenthapadi41:57
Who should watch
You are deploying models and need a practical checklist for validation, monitoring, and post-deployment checks.
Your organization is moving from rule-based systems to machine learning and needs explainability for legal, compliance, or business stakeholders.
You work on datasets or models that affect people and need to include consent, representation, fairness, and user feedback in the development process.
