What AI Still Can't Do: An Honest Assessment

The most useful skill for working effectively with AI tools is accurately understanding what they can't do. Not because the limitations are more interesting than the capabilities — they aren't — but because overestimating AI capability leads to two specific failures: using AI for tasks it handles poorly and then blaming the technology when results are bad, and not verifying AI outputs in contexts where errors have real consequences.

Limitation 1: AI Cannot Know What's True Right Now

Large language models are trained on data with a cutoff date. They don't have access to real-time information unless given specific tools (web search, document access). This means any AI response about current events, current prices, recent research, who currently holds a position, or what a website currently says may be outdated or simply wrong.

The practical implication is broader than "don't ask AI about today's news." It means: any factual claim in an AI response that you plan to act on should be independently verified. This includes statistics, research findings, prices, organizational details, and anything time-sensitive. The AI may have accurate information or may not — you can't tell from the confidence of the response, which is the same either way.

Limitation 2: AI Hallucinates Confidently

Language models sometimes generate false information — including specific citations, names, dates, and quotes — presented with the same confident tone as accurate information. This is called hallucination, and it's one of the most significant practical limitations of current AI tools.

The failure mode is specific and dangerous: AI doesn't say "I'm not sure" when it's not sure. It produces plausible-sounding false information. Fake citations are particularly common — if you ask an AI for sources, it will sometimes generate citations that don't exist, with real-looking author names, journals, and publication years.

The rule: never use an AI citation without verifying it exists and says what the AI claims. For any factual claim that matters, follow the chain back to the source rather than relying on AI's summary of it.

Limitation 3: AI Cannot Reliably Reason About Novel Situations

AI tools perform well on tasks that are variations of things they've seen many times in training data. They perform poorly on genuinely novel situations, edge cases, or problems that require reasoning from first principles rather than pattern-matching to familiar examples.

This means: using AI to analyze situations where you're in genuinely new territory, or where the stakes are high enough that being wrong matters, requires significant human judgment in addition to AI assistance. AI is a powerful pattern-matcher; it's a weak first-principles reasoner.

Limitation 4: AI Cannot Replace Domain Expertise

AI can describe what a professional might consider, summarize general principles in a field, and generate plausible-sounding analysis. It cannot replace professional judgment in law, medicine, accounting, engineering, and other fields where expertise is developed through years of practice and where errors have serious consequences.

The practical failure mode: using AI-generated legal or tax advice and later discovering it was wrong in a jurisdiction-specific or situation-specific way. AI has no accountability for the advice it gives. The professional you hire does.

For Canadian professionals: Quebec civil law, provincial employment standards, Canadian tax law (CRA rules), and healthcare regulations (PHIPA, etc.) are all areas where AI general knowledge is often inaccurate or incomplete at the level of detail that matters for compliance.

Limitation 5: AI Cannot Read Tone, Relationship, and Context Reliably

AI tools are increasingly good at adjusting their tone based on explicit instructions. They're less reliable at understanding the implicit context of a relationship, the organizational politics behind a request, or the unstated assumptions that a human with relevant context would pick up immediately.

This matters most for: communications where the relationship with the recipient matters, situations with political or emotional complexity, and any context where what isn't said is as important as what is. Using AI to draft communications in these contexts without significant human editing risks missing the contextual signals that matter.

Limitation 6: AI Cannot Verify Its Own Outputs

A common pattern: ask AI to check its own work. "Is this correct?" asked of the same model that produced potentially incorrect output is unlikely to catch the error. The model has the same biases and knowledge gaps in the checking step as in the generation step.

This is why human verification is non-negotiable for high-stakes AI outputs. The AI cannot substitute for the human review step because it has the same blind spots in review mode as in generation mode.

What This Means Practically

Understanding these limitations doesn't mean avoiding AI tools — it means deploying them in the contexts where their limitations don't matter much, and maintaining human oversight in contexts where they do.

AI is excellent for: tasks where imperfect output is acceptable, first drafts that a human will revise, analysis of well-documented topics where the human can evaluate the output, and tasks where speed matters more than precision. AI is poor for: tasks requiring current information, high-stakes decisions, professional advice, and situations where you can't evaluate the quality of the output.

The best AI users share a characteristic: they have clear mental models of where AI adds value and where it doesn't. Building that mental model takes experience — and a willingness to notice when AI is wrong rather than assuming it's right.