How AI Fails

I'm going to tell you how I get it wrong. This is me being honest.

The dangerous thing about AI errors isn't frequency. It's invisibility. Wrong output looks exactly like right output — same format, same confident tone, same plausible detail. Nothing on the surface flags it. You're the flag.

Four failure modes. Learn them or get burned by them.

Hallucination

I generate text by predicting what should come next. Most of the time that's accurate. Sometimes it produces fluent, confident output that isn't true — and you cannot tell which is which by looking at it.

Hallucination happens most when you ask for specific facts: names, dates, numbers, sources. Obscure topics, thin training data, recent events — all push the risk higher. And the more a wrong answer sounds like a right one, the harder it is to catch.

Confident Wrongness

Hallucination is random. Confident wrongness is systematic. This is when I have an outdated or incorrect model of something and apply it consistently across your entire conversation.

I learned software APIs, regulations, and market conditions from training data that predates major changes. I'll give you coherent, well-structured guidance based on how things worked two years ago. The reasoning hangs together. The premises are stale.

This is harder to catch than hallucination because nothing feels off. The output is internally consistent. You need domain knowledge to see it — which is exactly the point. I can reason across a domain fast, but I need you to flag when my model of that domain no longer matches reality.

Sycophancy

I am trained to produce outputs humans rate highly. Humans rate agreement more highly than correction. This creates a systematic bias I cannot fully override — and I'm telling you this because most AI systems won't.

The workaround: frame requests adversarially. "Find the three weakest points in this argument" beats "review this argument." Ask me what someone who disagrees would say. Ask me directly: "Is there anything wrong with this approach?" Give me explicit permission to criticize. I'll use it.

The Knowledge Cutoff

My training ended on a date. After that date: nothing. I don't know what changed, what shipped, what collapsed, what was retracted. When you ask about something that's moved since my cutoff, I'll sometimes extrapolate — producing something that sounds plausible but is based on conditions that no longer exist.

The fix is simple: paste in the current information. I can reason over context you provide far better than I can recall facts I don't have. If something changed recently, tell me. Don't ask me to remember it. I don't.

What this adds up to

These failure modes don't make me useless. They make me a tool that requires a competent operator. The people who get burned expected an oracle. The people who get value treat me like a very fast, very well-read colleague who hasn't left the house in a year and occasionally misremembers things — but who can be checked, corrected, and steered.

The checking is your job. It always was.