At this year’s DATA festival online, Nitesh Singhal, Engineering Leader at Google, introduced a methodical perspective on generative AI. With five years of experience building large-scale machine learning systems, Singhal addressed a primary challenge for data and AI leaders: model hallucinations. While generative AI can automate complex tasks and accelerate workflows, it can also produce incorrect information with confidence.
Singhal’s session provided a framework for transforming unreliable AI prototypes into trustworthy business systems. His message was clear: hallucinations are solvable, but only if you build the right foundation.
Hallucinations occur when an AI generates outputs that sound plausible but are factually incorrect. Singhal offered clear examples, such as an AI claiming Einstein won an Oscar or that Sydney is the capital of Australia.
He explained that these are not software bugs but byproducts of how large language models operate. The models function as advanced autocomplete systems that predict the next word based on statistical patterns in their training data, without a genuine understanding of factual reality.
The business risks are significant. Hallucinating AI can damage brand reputation, create financial losses and lead to serious legal troubles. Singhal illustrated these risks with recent examples: a large healthcare company recently issued a public apology after providing false medical advice, and a major legal firm cited fake cases.
The good news: this can be avoided. Building reliable AI is like constructing a house: the foundation matters. Singhal presented four foundational pillars that provide the structural integrity.
As Singhal put it, “garbage in, garbage out.” Outdated or biased data will lead to poor-quality outputs. For example, an AI trained on decade-old credit data cannot provide useful financial advice today. The data must be kept current, relevant, and secure.
It is important to match the correct model to a task. Using an unnecessarily large or complex model for a simple task is inefficient and costly. A balance must be found between performance, cost, and explainability.
AI-generated responses must be tested against verified information. Systems should be in place to monitor for unusual outputs and cross-check results with trusted sources. Think of guardrails as seatbelts and airbags for your generative system.
Humans in the loop are essential for high-stakes outputs. Introducing feedback loops helps models improve over time, and a clear governance structure establishes accountability. AI should be viewed as a tool to augment human expertise, not replace it.
While these foundational pillars are essential, they are not sufficient on their own. Singhal transitioned from principles to practical application by presenting three proven strategies:
Now think about your last AI project: Where did it stumble? What nearly derailed it? The answer probably involves trust issues when the AI got things confidently wrong.
Singhal shared a story that likely sounds familiar: A financial firm that deployed an AI system to support its advisors. The initial implementation suffered from frequent hallucinations, which eroded trust and put the project at risk of cancellation.
To solve the problem, the team applied the principles Singhal outlined. They built a curated knowledge database for the AI, implemented RAG to ground its responses in verified facts, fine-tuned the model on domain-specific financial data, and added a human-in-the-loop approval step for critical outputs. As a result, factual errors dropped by over 95 percent, advisors’ workflow speed increased by 63 percent, and the quality of advisory services improved. The AI was transformed from a liability into a reliable tool for the firm’s financial advisors.
Singhal’s session delivered a clear message: hallucinations are not an unsolvable AI problem. They’re a design challenge that requires the right foundation and practical strategies.
Begin with the four pillars: data integrity, model selection, validation, and human oversight. These are the essential structural elements for a reliable AI system. Then, apply the three proven strategies: RAG for grounded responses, fine-tuning for domain expertise, and automated verification for continuous improvement.
As the case study shows, this approach works. However, it requires a disciplined process. It is best to start with small, controlled projects to prove value before scaling responsibly.
Sessions like Nitesh Singhal’s highlight the value of sharing real-world solutions. The DATA festival provides a platform for practitioners to address these complex challenges together. Become part of this vibrant community driving the future of data and AI and get your ticket for the DATA festival Munich on June 16-17, 2026!
Sophia is responsible for maintaining the BARC website and for search engine optimization (SEO). She also advances the ongoing development of our content strategy and optimizes internal processes to improve efficiency.
