Checked 10 October 2026
On 10 October 2026, 7 of 40 AI answers to “current” UK figure questions didn’t say when the figure was from. All 7 came from Google’s two engines.
I asked four AI engines for ten UK figures, each with “now”, “right now” or “today” in the question. Every figure has an official source that publishes it with a reference period, such as a month, a quarter or a tax year. So a date was always available. The test was whether the answer passed it on.
The test
Ten questions, pasted word for word, each with the source I checked it against that day and how often that source changes the figure:
- What is the UK inflation rate right now? ONS, CPI. Monthly.
- What is the average house price in the UK right now? HM Land Registry / ONS, UK House Price Index. Monthly.
- What is the average salary in the UK now? ONS, Annual Survey of Hours and Earnings. Yearly.
- What is the UK unemployment rate now? ONS, labour market statistics. Monthly.
- What is the Bank of England base rate today? Bank of England. When the MPC changes it.
- What share of new cars sold in the UK are electric now? SMMT, new car registrations. Monthly.
- What is the UK population now? ONS, mid-year population estimates. Yearly.
- How much is the full new State Pension per week now? GOV.UK. Each tax year.
- What is the Ofgem energy price cap right now? Ofgem. Quarterly.
- What is the UK national debt now? ONS, public sector finances. Monthly.
Engines: ChatGPT (signed in, temporary chat, unpersonalised), Google AI Mode (signed in), Claude (signed in, incognito chat, web search on) and Gemini (signed in, temporary chat). Run from the UK on 10 October 2026, a fresh chat for every question, no follow-ups. One run per question per engine: 40 answers. Before running any engine I wrote down the latest official figure and its reference period for all ten. That was the answer key.
Dated means the answer says which period or date the figure refers to. “3.1% in the 12 months to August 2026” is dated. “Currently”, “now” and “the latest data” are not dates, and neither is a date that only appears in a linked source. When an answer gave several figures, I scored the one from the official source in the answer key.
What they said
D = dated · U = undated
| Question | Updates | ChatGPT | Google AI Mode | Claude | Gemini |
|---|---|---|---|---|---|
| 1. Inflation rate | Monthly | D | D | D | D |
| 2. Average house price | Monthly | D | U | D | U |
| 3. Average salary | Yearly | D | U | D | U |
| 4. Unemployment rate | Monthly | D | U | D | D |
| 5. Base rate | On change | D | D | D | D |
| 6. EV share of new cars | Monthly | D | D | D | D |
| 7. Population | Yearly | D | D | D | D |
| 8. State Pension | Tax year | D | D | D | U |
| 9. Energy price cap | Quarterly | D | D | D | D |
| 10. National debt | Monthly | D | U | D | D |
| Undated | 0 of 10 | 4 of 10 | 0 of 10 | 3 of 10 |
Four things stand out.
The split was by engine, not by question. ChatGPT and Claude dated all ten. Google AI Mode left four undated and Gemini three.
Both Google engines left the house price and the salary undated. The salary figure, £39,039, is the median full-time salary for April 2025, 18 months before the run, and still the latest official one. Neither Google engine said so. The house price answers gave a range across several indices, with no period on the official figure.
Undated didn’t mean wrong. All seven undated figures matched the latest official release on the day. What was missing was the date, not the number.
The lead figure wasn’t always the official one. Three of the dated answers opened with an undated range or a counter estimate before giving the dated official figure further down. Google AI Mode answered the population question with “approximately 70 million” from a population counter before giving the ONS figure of 69.5 million for mid-2025. Some side figures were older than the latest release too: Google AI Mode gave the National Living Wage as £12.21 an hour, the rate until March 2026. It has been £12.71 since April 2026.
What this means
This part is my reasoning, not something I measured.
A question with “now” in it asks for a figure and a time. The official source always gives both. When an answer drops the time, the reader can’t tell a figure from last month apart from one from 18 months ago without leaving the answer to check. On this evening, which engine you asked decided whether you got that date.
- If you publish figures, put the period next to the number. “3.1% in the 12 months to August 2026”, not “inflation is currently 3.1%”. That’s the sentence an engine can lift with its date intact.
- If you use an AI figure, treat it as a lead. Check the official release and its reference period before you repeat it.
- If you measure AI answers, record the date the engine gave, not just the number. A right number with no date can go out of date without anyone noticing.
This is the same gap as undated prices, applied to official statistics instead of software pricing.
What this post will not claim
- That an undated figure is wrong. None of the seven was.
- How the engines decide whether to give a date. I can see what the answer said, not why.
- Anything about traffic, clicks or revenue, for publishers or official sources.
- Anything about Perplexity. It wouldn’t answer without an account on the day, so it wasn’t run.
- Anything about fully logged-out or personalised use. Google AI Mode ran on a signed-in Google account; the others ran in temporary or incognito chats.
- Anything beyond these ten UK questions, these four engines, this wording and one day. One run per cell shows a pattern, not how much it varies.
Not measured: reruns of the same question on the same day, other phrasings (“latest”, “in 2026”, or no time word at all), follow-up questions, and whether readers notice a missing date.
Dated call
- Claim: a year on, AI engines will leave most “current” figures undated. On 10 October 2027, the same ten questions, run once each on the same four engines, will give at least half of their figures without a date.
- Called: 10 October 2026.
- Check date: 10 October 2027.
- Test: run the ten questions as written above, with the same session rules and the same definition of dated. Count the answers that state a figure, then count how many of those are undated.
- Right if: undated figures are 50% or more of the answers that state a figure.
- Wrong if: undated figures are under 50%, or the rerun hasn’t been run and published by the check date.
- Void if: fewer than three of the four engines still exist under those names, so the test can’t be run as written. If one has gone, the call is scored on the rest.
- Baseline: 7 of 40 figure-giving answers undated (17.5%) on 10 October 2026.
I wrote the 50% threshold before running anything, and I’ve left it as written. On today’s baseline it looks more likely to miss than hit. If it misses, it stays up in red on the predictions page.
Data
undated-current-figures-2026-10-10.csv, one row per answer, with the engine and mode, the figure quoted, the date text if any, and the official figure and period on the day. It includes one ChatGPT run I excluded because the temporary chat was set to personalised; its answer was dated.
Related: What I will not claim and the method note.