LLM Failure Modes

The explainability failure

Jack Dorsey identifies a general societal risk distinct from any single coding failure: "One of our biggest risks in artificial intelligence going forward is we are building a lot of black boxes that can't necessarily explain why they made a decision or what criteria they used. And we're trusting them more and more, from lending decisions to content recommendation to driving to health."¹ For Dorsey, the real chasm between current systems and something closer to general intelligence is the ability to explain why, to give a natural-language account of the reasoning behind a decision. Until that gap closes, AI systems are being handed consequential, sometimes irreversible decisions they cannot justify after the fact. He is careful not to frame this as despair: humans also frequently cannot explain their own decisions, and current systems may be judged somewhat unfairly on a standard humans do not meet either. What makes the AI case more urgent is scale and irreversibility, the sheer number of decisions being handed off and how hard some of them are to undo.

Hallucination blocking real deployment

Kevin O'Leary offers the most direct practitioner confirmation that hallucination is still blocking deployment at real scale. He built an AI agent system, deployed across dozens of portfolio companies for social media management, CRM, and customer retention, and reports it plainly: "These agents hallucinate. After a while they start wasting your time and we are nowhere near yet."² His planned fix is not to wait for the models to improve on their own but to build a hybrid operation, a call center combining AI, agent-assisted, and human components, on the reasoning that a business can never fully get away from human beings. The failure mode requires a human in the loop not as an efficiency nicety but as a reliability necessity, and because this account comes from outside software development entirely, applied to marketing and customer operations, it suggests the failure pattern generalizes rather than being specific to coding agents.

The root failure: no model of consequences

Mark Cuban names what may be the generator behind the other failures rather than another item on the list: today's large language models have no model of consequences. "AI right now doesn't understand the consequences of its recommendations. It has no idea what happens next. A two-year-old kid with a high chair and a sippy cup knows if it pushes the sippy cup off the high chair, mom's coming running. Large language models don't understand."³ Reframed this way, an AI system that could actually simulate the consequence of an assumption, if I assume this and I'm wrong, the build breaks, would be far more likely to catch its own wrong assumptions, avoid overcomplicating a solution, or notice when it is about to mislead someone, all without needing to be told separately not to do each of those things. A system trained only on correlations across text has no such action-to-outcome loop built in.

This framing does double duty. It bounds how dangerous a system without a working model of the physical or causal world can be, since capability stays tied to specific, narrow applications rather than becoming general, and it has a security dimension too: a system with no underlying physics prior can be fooled by something as simple as a painted road on a wall, the same absence showing up in a different form.

Why it matters

These three failure modes sit at different levels and compound rather than duplicate each other: Dorsey's concerns the system's inability to justify a decision after the fact, O'Leary's concerns the system's tendency to simply fabricate information during ordinary operation, and Cuban's concerns the system's inability to anticipate outcomes at all before acting. Taken together they argue for a specific kind of supervision, not proofreading for syntax errors but actively catching wrong assumptions, unjustified decisions, and consequences nobody checked for in advance, which is a higher-level judgment task than debugging code used to be and one that benefits disproportionately from actual seniority and domain experience.