Where AI will quietly lie to you

Where AI Falls Short (And Why It Matters)

This is arguably the most important lesson in the entire course. Understanding AI's limitations isn't about being pessimistic — it's about being strategic. When you know where the edges are, you can position yourself safely inside them and get consistently excellent results.

Knowledge Cutoff Dates

Every AI model has a knowledge cutoff — a date beyond which it has no training data. If the model was trained on data up to early 2025, it genuinely does not know about events, publications, or developments that occurred after that point. It's not being evasive when it says "I don't have information about that" — it literally doesn't.

This matters more than you might initially think. It means the AI might not know about:

• Recent changes to regulations or legislation
• Newly published research in your field
• Current events, personnel changes, or market developments
• Updates to software tools, APIs, or frameworks

Some AI platforms now include web search capabilities that partially address this limitation, but it's important to understand that the model's core knowledge has a hard boundary. When you need current information, always verify.

Hallucination: The Confident Fabricator

This is the limitation that catches most people off guard. AI can — and regularly does — generate information that sounds entirely plausible but is completely fabricated. The technical term for this is "hallucination," and it's not a bug that will be patched in the next update. It's a fundamental consequence of how prediction-based AI works.

Remember our mental model: the AI is predicting what text should come next based on patterns. If you ask it for a citation to support an argument, it knows what citations look like — author names, journal titles, volume numbers, page ranges. So it generates something that follows the pattern of a real citation. But it may have invented the paper entirely. The title sounds real, the journal exists, the date is plausible — but the actual paper does not exist.

This happens because the AI isn't retrieving information from a database; it's generating text that is statistically likely given the context. When it doesn't have a real answer, it produces a plausible-sounding one instead, with no indication that anything is amiss.

The practical implication is clear: never treat AI output as a verified source of facts. Always check claims, citations, statistics, and specific details against reliable sources. This is especially critical in professional contexts where accuracy matters — legal, medical, financial, or regulatory work.

Context Window Limits

The AI's ability to remember your conversation has a ceiling, known as the context window. Think of it as the AI's working memory — there's a maximum amount of text it can hold in mind at once, including everything you've said, everything it's said, and any documents you've shared.

When a conversation exceeds this limit, the AI starts to lose track of earlier content. It might forget instructions you gave at the beginning, overlook details from a document you shared, or contradict something it said earlier. Context windows have been growing rapidly (some models now support extremely long conversations), but the limit still exists, and performance often degrades well before the hard ceiling is reached.

For practical purposes, this means that very long conversations or very large documents may push against these boundaries. If you notice the AI forgetting things or becoming less coherent, starting a fresh conversation often helps.

Non-Determinism: Different Answers Every Time

Here's something that surprises many people: if you ask an AI the same question twice, you'll often get different answers. Not wildly different, usually, but noticeably so — different phrasing, different examples, sometimes different emphasis or even different conclusions.

This happens because the generation process involves a degree of randomness by design. When the model predicts the next word, there are usually several plausible options, and it samples from among them. This randomness is actually desirable — it's what makes AI output feel natural and varied rather than robotic and repetitive. But it means you can't treat AI output as a stable, reproducible reference.

If you need consistency, be aware that running the same prompt multiple times may yield varying results. This is particularly important when using AI for processes that need to be auditable or repeatable.

Reasoning Limitations

Despite how intelligent AI can seem, its reasoning capabilities have genuine gaps. It performs best on tasks that are well-represented in its training data — common logical patterns, standard analytical frameworks, and well-known problem types. But it can stumble on:

• Multi-step logic problems that require carefully tracking several conditions simultaneously
• Mathematical reasoning beyond straightforward arithmetic, particularly word problems that require setting up equations
• Spatial and physical reasoning — understanding how objects interact in three-dimensional space
• Novel problems that require genuine insight rather than pattern-matching against known solutions

The tricky part is that AI will often attempt these tasks confidently, producing answers that look well-reasoned but contain subtle errors. It might get nine steps of a ten-step problem right and then make a mistake on the tenth that invalidates everything. Always verify reasoning-heavy outputs carefully.

No Access to Your Private Data

Unless you explicitly provide information to the AI during your conversation, it has no access to your private data. It cannot read your emails, access your company's internal documents, check your calendar, or look at your files. It only knows what's in its training data and what you share in the conversation.

This is both a limitation and a reassurance. It means the AI can't proactively draw on your organisational knowledge, but it also means your private information isn't being accessed without your knowledge. When you need the AI to work with your specific data, you'll need to share the relevant information within the conversation (and be mindful about what you share, which we'll cover in a later module on responsible use).