EP 77: They Transferred $25M to a Deepfake — Why Your AI Isn't as Governed as You Think | Dr. Latha Karthigaa

Inside AsembleAI EP 77 features Dr. Latha Karthigaa, Director & Head of AI Governance at the Global AI Certification Council, discussing why corporate AI governance is often inadequate despite companies believing they're compliant. The episode opens with a critical distinction: having a policy is not the same as having governance—Karthigaa uses a highway guardrail analogy to reframe governance as something that enables faster innovation rather than merely restricting it. The conversation centers on concrete failure cases. The headline example is a $25 million wire fraud at engineering firm Arup, where an employee transferred funds after a video call with his CFO and colleagues—all deepfakes generated from publicly available videos. This illustrates why employee training and human verification matter more than policies alone. Karthigaa cites alarming statistics: 2,000+ logged AI incidents, 400+ in 2025, with over 52% deepfake-driven, including a SpaceX/Elon Musk crypto scam that aired during a live rocket launch. The episode also examines systemic failures: IBM Watson's oncology tool was shut down due to biased training data, and the Dutch childcare-benefits algorithm falsely accused thousands of families, ultimately bringing down a government. A key theme is accountability: Karthigaa emphasizes that ChatGPT itself bears no responsibility—the deploying organization does. Every AI use case needs a named human owner. The episode addresses ISO/IEC 42001, the only certifiable AI standard, discussing when startups should pursue it and what it signals to customers. Karthigaa notes that AI governance roles are growing 150% year-over-year, reflecting rising demand. The core message: "Policy is like potential power. Execution is the real power." The episode positions governance not as compliance theater but as a competitive advantage and risk mitigation strategy, with emphasis on human judgment, empathy, and accountability structures rather than purely technical or policy-based solutions.

Verdict: A substantive, well-focused episode with a credible expert, concrete examples, and non-obvious insights about AI governance that go beyond compliance theater—valuable for anyone deploying AI systems.

Technology rubric

• Technical accuracy 4/5 20%

  Cites real incidents (Arup $25M deepfake, IBM Watson, Dutch childcare algorithm) and concrete statistics (52% deepfake incidents, 150% YoY growth in governance roles); no obvious errors detected in show notes, though depth of technical deta

• Insight beyond the headlines 4/5 20%

  Reframes governance as enabler rather than constraint (guardrail analogy); distinguishes policy from execution; emphasizes human accountability and training over compliance theater—these are non-obvious insights beyond typical AI risk cover

• Clarity for non-experts 4/5 15%

  Uses accessible analogies (highway guardrails); explains ISO/IEC 42001 in context; avoids heavy jargon while discussing technical failures; makes governance concepts concrete through real-world examples.

• Substance 4/5 15%

  Covers multiple dimensions: policy vs. execution, deepfake incidents, systemic failures, accountability structures, standards, and labor trends; substantive guest with relevant expertise; limited padding evident in show notes.

• Delivers on its title 5/5 15%

  Title promises explanation of why AI governance is inadequate and what responsible governance looks like; show notes confirm both are directly addressed through case studies, statistics, and Karthigaa's framework.

• Value for time 4/5 15%

  33 minutes for a focused governance discussion with named expert, multiple case studies, and actionable takeaways suggests efficient pacing; no evidence of filler or tangents in show notes.

How we score Technology

Worth it if you work in AI deployment, risk management, or leadership roles where you need to understand why governance failures happen and how to build accountability structures that actually work.

Based on the episode description. AI-generated; it can be wrong.