What are the real capabilities and limitations of generative AI?
To use a tool well, you need to know how it works and where it breaks down. Without technical jargon: what to expect from AI, where it excels, where it makes things up, and how to work with it safely.
How it works, honestly
Generative models learn from enormous amounts of text to predict how a conversation continues. From this apparently simple capability, surprising abilities emerge: writing, summarising, translating, coding, reasoning about problems. But the nature remains the same: the model produces the most plausible continuation, not guaranteed truth. Plausible and true often coincide — but not always. All professional use of AI stems from this distinction.
Where it truly excels
Transform: summarize a fifty-page document, rewrite for a different audience, translate, extract structure from chaos. Explore: generate ten alternatives where you would have seen two, play devil's advocate, simulate an objection. Accelerate: first drafts, starter code, consistency checks. In all these tasks the cost of error is low because there is a human reviewing — and that is where AI multiplies productivity without risk.
Where it systematically goes wrong
Hallucinations: facts, numbers and citations invented with perfect confidence — never trust a verifiable piece of data without checking it. Biases: the model reflects the imbalances in the texts it learned from, and may reproduce them in its suggestions. Currency: knowledge stops at the training cutoff date, unless connected search tools are available. Precision mathematics and impact calculations: for those you need declared methods and coefficients, not plausible text — which is why on Simbial the scores are calculated by the server using public formulas, never by AI.
The craft of asking well
The quality of answers depends greatly on the quality of the questions. The rules that always apply: give context (who you are, what is needed, who it is for); ask for the format you want; proceed in steps rather than with one giant request; ask the model to show its reasoning and assumptions; and when the answer matters, also ask «what could be wrong with this answer?». Treat it like a brilliant but inexperienced collaborator: clear instructions, final verification.
The healthy division of labour
To AI: volume, speed, alternatives, first drafts, consistency checks. To the human: objectives, criteria, choices, verification of facts that matter, sign-off. This division is not a concession to nostalgics: it is what the data on the future of work shows — the professionals who grow are those who use AI to amplify their own judgement, not those who use it to avoid thinking.
Next stop: inside a real project
Now that you know what to expect from the tool, the next lesson brings it into the field: how AI is used in each of the 4 phases of an impact project — from ideation to measurement — with concrete examples from the Simbial method.
Frequently asked questions
What are AI hallucinations?
Invented content presented with confidence: plausible but false facts, numbers, and citations. These are a structural limitation of generative models: every verifiable piece of data must be verified before use.
Why doesn't AI calculate the Impact Score on Simbial?
Because scores require public and reproducible methods and coefficients, not plausible text: they are calculated by the server using declared formulas. AI helps structure data, but never invents numbers.
How do you write a good AI request?
Clear context, desired format, small steps, a request to show assumptions and reasoning — and for important responses, the final question: what could be wrong here?
← The full course · Next pill: AI in the 4 phases of an impact project →