What AI Can't Do
AI is remarkable — but hype often outpaces reality. Understanding limitations helps you use AI effectively.
1. Hallucinations
- Cite non-existent research papers
- Invent historical events
- Create plausible-sounding but wrong facts
Why: AI predicts likely sequences of words, not truth. "Likely" doesn't mean "true."
2. Reasoning Limitations
- Multi-step logical problems
- Novel situations outside training data
- Common-sense reasoning humans find trivial
Example: "I have 2 apples. I eat one and then go to the store. How many apples do I have?" AI often fails such questions.
3. No Real Understanding
- "Hot" relates to temperature sensation
- "Sad" relates to emotional experience
- Words connect to real-world referents
This is the "Chinese Room" problem — manipulation without comprehension.
4. No Persistent Memory
- Previous sessions
- What you prefer
- Past mistakes to avoid
(Some products add memory features, but core models lack it.)
5. Training Data Cutoff
AI doesn't know recent events. Information after training is invisible.
6. No Goals or Desires
- Intentions
- Curiosity
- Self-preservation instinct
Why Limitations Matter
- Verify AI outputs — especially facts
- Don't anthropomorphize — it's a tool, not a person
- Use for appropriate tasks — drafting, brainstorming, summarizing
- Maintain human judgment — AI assists but doesn't replace thinking
Related Reading
Listen to the Full Course
Explore AI in depth: Understanding GPT & AI: A Complete Guide.