General-purpose AI has achieved remarkable capabilities but struggles in real-world private settings. This talk examines three barriers to deployment.
1. Human-to-machine communication. Encoding human intent into AI remains hard. Modularity and agentic architectures offer promising solutions.
2. Machine-to-human communication. Users must understand AI decisions. This drives research in explainable AI and training data attribution.
3. Privacy and security. Deployment in sensitive domains is blocked by PII leakage, adversarial attacks, and membership inference risks. Regulatory compliance adds further constraints. I present our recent work on these challenges.
Future directions. AI research is shifting from model scaling to adaptation, personalisation, and agent-based interfaces. I conclude with a new direction - separating knowledge from intelligence in AI systems. Software engineering decoupled code from data 50 years ago. A similar decoupling in AI could unlock transparent, editable, and trustworthy systems.
