AI Capabilities and Limitations
AI Capabilities and Limitations
Know what AI does well — and where it will fail you.
Core mechanism: Claude doesn't look up facts. It predicts the next most statistically likely word — like a supercharged autocomplete. This is why it writes fluently and fabricates fluently.
What Claude Does Well
| Capability | What it means in practice |
|---|---|
| Versatility | Draft emails, translate, write code, explain concepts — same model, no switching |
| Pattern recognition at scale | Read 200 reviews or 500 support tickets → find themes in minutes, not days |
| Few-shot learning | Give 2–3 examples → Claude matches your exact format and tone |
| Tool use (agentic) | Execute code, search the web, connect to Slack/Gmail/Drive via MCP |
| Extended Thinking | Step-by-step reasoning for math, logic, complex debugging (+10–30% accuracy) |
Best for: Tasks in the "capability zone" — mainstream, stable, well-documented topics (e.g., "write a Python script", "summarize this report").
Where Claude Fails
| Limitation | What happens | Fix |
|---|---|---|
| Knowledge cutoff | Doesn't know anything after its training date — claims stale facts with confidence | Enable web search tool |
| Hallucination | Fabricates citations, case laws, statistics, URLs — with full confidence | Cross-check all specific numbers/names/citations |
| Context window cliff | When the chat fills up, oldest content is silently dropped — no warning | Start fresh with a summary; use /compact in Code/Cowork |
| Lost in the middle | Accuracy drops >30% for info buried in the middle of long docs | Put critical instructions at the TOP and BOTTOM of your prompt |
| Non-determinism | Same prompt → different answer each time | Set temperature = 0 for consistency-critical tasks |
| Inherited bias | Defaults to Western, English-speaking, tech-industry norms | Flag local context explicitly: "in the context of Vietnam..." |
The Hallucination Rule
Confidence does not equal accuracy. Claude says "I'm certain" the same way for both correct and fabricated facts.
Hallucinations concentrate in specifics: names, dates, numbers, DOIs, URLs, legal citations.
3 fixes:
- Add
"If unsure, flag it as [NEEDS VERIFY] — do not make things up"to your prompt - Use RAG (upload the source document) so Claude answers from your file, not memory
- Run a separate verifier pass: "For each claim, tag it HIGH / MEDIUM / LOW confidence"
AI vs. Human — Who Wins Where
| AI wins | Human wins |
|---|---|
| Speed, scale, pattern detection | Critical judgment, ethical decisions |
| Broad knowledge recall | Deep domain expertise |
| Parallel processing | Real-time awareness |
| Draft and summarize | Final call and accountability |
The right model: AI does the "broad and fast." You do the "deep and judgment." Combine both for output neither could produce alone.
Source: Claude Course 2026 — NotebookLM