What AI can and cannot do

In plain English

Today's AI is excellent at spotting patterns in large piles of examples: classify images, transcribe speech, summarize familiar document types, recommend the next likely word.

It is weak at guaranteed truth, long-horizon planning without tools, and situations far from training data. A confident tone does not mean a correct answer.

The right question is not can AI do everything? but is AI reliable enough for this task with these checks?

Strengths and sharp edges

Strengths: speed at scale, tireless drafting, embedding-based search, personalization, perception tasks where labeled data exists.

Limits: factual errors and hallucinations in generative models; sensitivity to prompt wording; struggle with precise arithmetic or rare logic without tools; potential bias; security issues (prompt injection, data leakage).

Mitigations: retrieval with citations, calculators and code interpreters, human review for high-stakes decisions, monitoring in production, clear escalation paths.

• Good fit: assistive drafting, search, classification with audit samples.
• Risky alone: medical diagnosis, legal advice, unsupervised autonomous actions.
• Always verify: citations, numbers, policies, safety-critical steps.

Going deeper

Evaluation matters. Offline benchmarks do not capture your users, your domain jargon, or adversarial misuse. Build task-specific tests before trusting automation.

Capabilities improve with tools: RAG for fresh documents, function calling for APIs, agents for multi-step workflows. Limits shift but do not disappear—each layer adds new failure points to design for.

Society-facing limits include energy use, labor impacts, and concentration of capability. Technical limits and human choices intertwine.

Common misconceptions

If the answer sounds polished, it is correct.

Generative models optimize for plausible language, not verified truth. Fluency is not fact-checking.

AI replaces the need for domain experts.

Experts are still needed to frame problems, validate outputs, and handle exceptions models never saw.

Zero mistakes is achievable with a bigger model.

Error rates can fall but rarely vanish. High-stakes workflows need thresholds, monitoring, and human override.

Key facts

• Modern AI excels at pattern matching and scalable perception on familiar data.
• Generative models can produce confident incorrect statements (hallucinations).
• Performance drops on out-of-distribution inputs unlike training examples.
• Tooling (retrieval, calculators, APIs) extends capability but adds system complexity.
• Human oversight remains essential for high-impact decisions.