Generative AI has real, well-documented limitations — technical, operational, and ethical. Building with it responsibly means designing around these, not ignoring them.
Technical Limitations
| Limitation | What It Means in Practice |
|---|---|
| Hallucination | Fluent, confident-sounding output that is factually wrong. See LLM Hallucinations. |
| Knowledge cutoff | A model only knows about events up to its training data's cutoff date, unless given external context (see RAG) |
| Context window limits | Only a bounded amount of text can be considered at once. See Context Window. |
| Inconsistent reasoning | Can make logical or arithmetic errors, especially on multi-step problems, without any visible sign of low confidence |
| Non-determinism | The same prompt can produce different outputs across runs, complicating testing and reproducibility |
Operational / Business Risks
- Cost at scale — per-token API pricing means costs grow directly with usage volume and output length; an unmonitored feature can produce a surprising bill (see LLM API Cost)
- Latency — generation is sequential (one token at a time), making it slower than a typical database query or classifier for long outputs
- Vendor dependency — building tightly around one provider's API/model behavior creates switching costs and exposure to that provider's pricing or availability changes
Security & Trust Risks
- Prompt injection — malicious instructions embedded in input data can attempt to override intended behavior (see Prompt Injection)
- Data leakage — sensitive information included in a prompt may be logged, cached, or (depending on provider terms) used for further training unless explicitly configured otherwise
- Over-trust — users treating fluent output as inherently authoritative, especially dangerous in medical, legal, or financial contexts
Ethical / Societal Challenges
- Bias — models can reproduce biases present in their training data
- Misuse potential — the same generation capability that drafts helpful content can be misused for spam, deceptive content, or impersonation
- Attribution and copyright — training data provenance and output ownership remain unsettled legal questions in many jurisdictions
How This Shapes Real Engineering Decisions
These aren't abstract concerns — they directly drive architecture choices: RAG to reduce hallucination and knowledge-cutoff issues, structured output + validation for reliability, human-in-the-loop review for high-stakes actions, and evaluation pipelines before shipping (see LLM Evaluation).
Common Mistakes
- Treating hallucination as a rare edge case rather than an expected, ever-present behavior to design around
- Shipping an LLM feature with no cost monitoring, then being surprised by the bill at scale
- Giving a model real-world capabilities (sending emails, modifying records) without validation or human approval — see Human-in-the-Loop
Interview Relevance
"What are the main risks of deploying an LLM feature to production?" is a very common systems-thinking question — a strong answer touches technical (hallucination), operational (cost/latency), and security (prompt injection) dimensions, not just one.
Practice Question
Your team wants to let an internal chatbot answer questions using live company data with no review step. List three concrete risks and one mitigation for each.
