AI Made This Worse
The source that did not exist
Written by AI Tools AcademyLast checked on 27 September 2026
Helpful first: Hallucination and trust
This is a fictional example. Fernway Group is a made-up UK company, and everyone in it is invented. The report and research centre named by the AI below are also invented for this example.
AI tools are good at explaining ideas and poor at remembering exactly where a fact came from. When you ask for evidence, that weakness matters most, because the answer arrives with a reference attached and looks checked when it isn't.
Before
Fernway's office lease is up for renewal. Priya Shah, Head of Operations, is preparing a paper for Erin Vaughan, the Managing Director, comparing staying put with moving to a smaller office now that most staff work hybrid. Priya asked Leah Bennett, Operations Assistant, to find some outside evidence on how much desk space hybrid offices actually use, so the paper wasn't relying only on Fernway's own impressions.
Leah had an afternoon, no subscription to any research service, and a paper due to Priya the next morning.
The AI approach
Leah opened the AI tool and asked directly:
I'm writing a paper on whether to move to a smaller office. What percentage of desks are typically used on an average day in UK offices with hybrid working? Please give a statistic and cite the source.
Why this works: This is the version that caused the problem: it asks for a figure and a source, which invites the tool to produce both whether or not it has them.
What went wrong
The answer came back fluent and specific.
Research suggests that UK offices operating a hybrid model typically use far fewer desks than they provide. According to the Albury Centre for Workplace Research's Desk Use in UK Offices 2025 report, average daily desk occupancy in hybrid offices was 41%, with peaks of around 60% on Tuesdays and Wednesdays. The report recommends a desk-to-employee ratio of around 0.6 for organisations with two to three office days per week.
It had everything Leah wanted: a percentage, a peak, a ratio and a named report with a year. She copied the paragraph into her notes, and Priya lifted the 41% figure and the 0.6 ratio into the draft paper, citing the report by name.
The next day Dan Okafor in Finance, building the cost comparison, tried to find the report to check the ratio. He found nothing. No report of that name, no research centre of that name, and no page anywhere carrying the figures.
The AI hadn't looked anything up. It had produced the kind of sentence that usually appears in answers to questions like Leah's: a plausible institution, a plausible title, a plausible year and plausible numbers. This is called a hallucination, and citations are one of the places it happens most, because a reference is just a pattern of words (name, title, year) that the tool can produce as easily as any other.
Two things made it harder to spot:
- The request asked for a citation. "Please cite the source" doesn't make a tool check sources. It makes it produce something shaped like a source.
- The numbers were believable. 41% and 0.6 are close enough to what people expect that nobody felt the need to question them. A wildly wrong figure gets checked; a plausible one gets pasted.
No harm reached Erin, because Dan checked. But if the paper had gone out, the lease decision would have rested partly on a figure nobody could stand behind, and Priya's name was on it.
A better approach
The fix is to reverse the order. Instead of asking AI to find a fact and then support it, find the sources first and ask AI to work only from what you give it. The Prove It course calls this working from sources, not towards them.
Weak
What percentage of desks are typically used in UK hybrid offices? Please give a statistic and cite the source.
Better
Below are three sources I've found on desk use in hybrid offices, pasted in full. Using only these, summarise what each says about daily desk occupancy, quote the exact sentence each figure comes from, and tell me where they disagree. If none of them answers a question, say so rather than filling the gap.
Why it works: The tool can only report what is in front of you, every figure comes with the sentence it came from, and it has explicit permission to say the answer isn't there.
If you don't have sources yet, use AI to help you look, not to answer:
- Ask which kinds of organisation publish this sort of data (property consultancies, workplace bodies, government statistics) so you know where to search.
- If your tool can search the web, ask it to link each claim, then open every link yourself and find the figure on the page. A link that loads is not the same as a link that says what the AI said it says.
- Use your own data where you have it. Fernway's door-entry records or a two-week desk count would have been more relevant to Fernway's lease than any national average.
And a check that takes less than a minute: search for the exact report title in quotation marks. If nothing comes back, treat the figure as unsupported until you find it somewhere real.