AI capability has advanced quickly, but it is uneven. A clear view of strengths and weaknesses helps you pick sensible projects.
Where AI Performs Well
- Perception: recognising objects in images, transcribing speech, reading documents.
- Language: drafting, summarising, translating, classifying and extracting information from text.
- Pattern-finding at scale: fraud detection, recommendations, demand forecasting.
- Code: generating, explaining and reviewing routine code.
- Narrow, well-defined tasks with plenty of representative data.
Where It Struggles
- Reliability on specifics: exact figures, citations and rare facts can be invented.
- Novel situations: performance drops when inputs differ from training data.
- Long chains of reasoning without feedback or verification.
- Common sense and physical understanding in unfamiliar settings.
- Accountability: a model cannot take responsibility for a decision.
Signs of Hype
Be sceptical of claims that a system is "unbiased", "fully autonomous" or "understands" in a human sense; of accuracy figures without a description of how they were measured; and of demos run only on the vendor's own examples.
A Practical Rule
Use AI where mistakes are cheap to catch and correct, keep people accountable for important decisions, and measure performance on your own data before relying on it.