As large language models (LLM) integrate into more applications, clear safety standards ensure reliable results and measure performance and effectiveness. At SXSW 2025, Scale AI’s Summer Yue, HeyGen’s Lavanya Poreddy, and Fortune’s Sharon Goldman sat down for the Beyond the Hype: Building Reliable and Trustworthy AI panel to explore pioneering AI safety standards, like rigorous model testing and high-quality data, to ensure a more transparent future for AI adoption. They also discussed best practices for mitigating risks and addressing ethical considerations in AI development.
Learn how leading organizations establish AI safety processes to ensure trust and real-world adoption.
Understanding AI and the challenge of evaluation
AI models are often described as “black boxes”—powerful but opaque. Summer noted that even AI researchers rely on informal sources like Twitter and Reddit to understand model capabilities. This trust gap raises concerns about AI’s unpredictability and biases.
Lavanya likened AI to a newborn child—it learns from data but lacks independent reasoning. AI isn’t autonomous intelligence; it’s human-backed intelligence that must be monitored and refined.
One major issue is how to measure AI performance. Traditional AI models recognize patterns, while reasoning models generate outputs based on different thinking times, making evaluations harder. Even with extensive testing, AI models still hallucinate—generating false or misleading information. This is particularly concerning in safety-critical applications like healthcare or law enforcement.
AI bias and content moderation challenges
Bias remains a major challenge. Lavanya explained that if an AI model is trained only on golden retriever images, it will always assume a “dog” looks like a golden retriever. The same problem applies to hiring models or content moderation, where skewed training data can reinforce biases. The solution? More diverse and inclusive training datasets. However, even with improved training, AI still struggles with context and intent, making human oversight essential.
Trust in AI depends on who sets the rules. AI companies have vastly different policies, leading to inconsistencies in content moderation and safety standards. AI automatically filters harmful content like hate speech or nudity. However, gray-area content (e.g., political discussions) is flagged for human review. Lavanya emphasized that AI cannot replace human decision-making, especially in complex ethical situations.
Summer noted that AI-generated content varies significantly across companies. Some models are highly restrictive, while others are more lenient, creating inconsistent user experiences.
AI in real-world applications
The discussion covered AI’s impact across industries, including:
- Healthcare – AI can assist with scheduling and insurance claims but should never make life-or-death medical decisions.
- Self-driving cars – AI follows traffic laws but lacks human intuition. Unlike a human driver, an AI-powered vehicle like Waymo cannot understand gestures or social cues.
- College admissions bias – Lavanya used AI to help her son apply for college. Unlike human counselors (who may hold racial or gender biases), AI provided unbiased recommendations based purely on academic interests.
These examples show that AI can enhance efficiency, but human oversight remains crucial. Lavanya and Summer agreed that greater transparency is key to building trust. AI should be seen as a collaborator, not a replacement for human decision-making. Lavanya and Summer believe:
- AI should assist with tasks but cannot fully replace human judgment.
- Transparency, fairness, and human oversight must guide AI development.
- Regulations will push AI companies toward greater accountability.
This SXSW session underscored the importance of responsible AI development. While AI’s potential is undeniable, ensuring it is fair, ethical, and transparent remains a critical challenge.
