AI makes things up because producing plausible language is not the same job as checking facts. The practical response is to make important claims traceable to evidence, not to ask for a more confident answer.
Google’s explanation of language models describes predicting likely next words from the prompt, the text already generated, and patterns learned during training. That mechanism can produce useful explanations. It can also complete the familiar shape of a citation or instruction when the underlying fact is missing.
NIST calls this confabulation: confidently stated but erroneous or false content. “Hallucination” is the common label. Neither label means the tool witnessed something and misremembered it.
Supplying a document helps, but does not finish the job
A chatbot can use material you provide, and some products can retrieve sources. That gives it evidence to work with, not a guarantee that every sentence stays inside the evidence. Google’s Workspace source guidance warns that capacity limits can leave some supplied content unprocessed and that citations can be omitted, misapplied, or invented. That warning concerns Workspace features specifically, not a measured error rate for every chatbot.
A source-grounded summary therefore needs two checks: did the system receive the relevant passage, and did the answer preserve its meaning? Upload success answers neither question on its own.
Check a small example before trusting a long summary
Hypothetical source text: “Returns are accepted within 30 days for unopened items. Opened items are eligible only if faulty. Delivery charges are not refunded.” This is an invented policy for illustration, not any retailer’s terms.
Now consider the hypothetical summary: “You can return any item within 30 days for a full refund.” It is neat, short, and wrong. It drops the unopened-item condition, loses the faulty-item exception, and turns a limited refund into an unrestricted one.
A useful request would be: “Using only the quoted policy, list the return window, condition requirement, exceptions, and excluded charges. Give the exact supporting sentence for each. If the text does not specify something, write ‘not stated.’” This is a way to make checking easier, not a special phrase that prevents errors.
Compare each field with the original yourself. The deadline should remain 30 days. The condition should remain unopened unless faulty. Delivery charges should remain excluded. If the answer adds a refund-processing deadline, reject that detail: the supplied text says nothing about one.
For a real policy, also open the current authoritative version and check that it applies to the purchase type. An accurate summary of the wrong policy is still the wrong advice.
Do not ask the answer to certify itself
“Are you sure?” may produce more text, but that text is not independent evidence. Google explicitly warns that Gemini can misrepresent its own operation, including how it cites sources. Ask for a passage you can inspect, then inspect it.
When an error appears, correct the specific claim and recheck the revised answer against the source. A revision is another output, not a guarantee that the remaining details survived unchanged. If the source is incomplete, keep the gap visible instead of filling it with a plausible answer.
For medical, legal, or financial decisions, use the responsible professional or authoritative service rather than treating a polished summary as advice. For ordinary document work, check that each decision-changing claim has a matching passage, its conditions remain attached, and unsupported additions are removed.
AI can help turn a dense page into a checklist. The checklist is useful only if it still says what the page says.