There is a particular kind of Australian argument that goes on for years without anyone thinking to check the equipment.

For the past two decades we have been conducting one of these about religion. The Census says the proportion of Australians with no religion has climbed from under 1 per cent in 1911 to 38.9 per cent in 2021.

Both sides have made a great deal of this number.

Secularists cite it as proof that the churches are finished and should be taxed accordingly. Christians cite the two-thirds who tick something as proof they are not.

Everyone agrees the number means something. Nobody, in 115 years, thought to test whether the number was measuring what they assumed.

So, we did.

We took 5,012 Australians, nationally representative, and gave each of them the Census religion question. But not the same version. We recreated four designs the Australian Bureau of Statistics has genuinely used since Federation and assigned people to them at random. Same population, same beliefs, same afternoon. One difference: the question.

• The 1911 version was a blank box. What is the person’s religion?
  • 20 per cent wrote nothing, or none, or something to that effect.
• The 1971 version added six helpful words: if no religion, write none.
  • 47 per cent.
• The 1991 version introduced a checkbox list with No Religion at the bottom.
  • 27 per cent.
• The current version, in use since 2016, moved No Religion to the top of that list.
  • 32 per cent.

20-47 per cent… A 27-point spread, on the same people, produced by nothing but the arrangement of words on a page. If a bathroom scale did this, you would throw it out. We have been using it to allocate school funding.

Run those measured effects back through the historical record and the picture rearranges itself. Strip out the design distortion and the 2021 figure lands closer to 26 per cent than 38.9 per cent. Roughly a third of the number that anchors this entire debate is not Australians losing their faith. It is the instrument.

Now, before anyone gets excited: secularisation is real. It is one of the largest social shifts of the past century and nothing here disputes the direction of travel. What is in question is the speed. Australia’s secularisation is genuine. The speedometer is broken. We have been reading the pace off a dial that has been rebuilt four times, and nobody checked what the rebuilding did to the reading.

The most unsettling part is not the 27 points. It is that nobody felt a thing. We asked respondents afterwards whether the wording had influenced them. Fewer than 8 per cent thought it had. 89 per cent were confident in their answer. 86 per cent said it accurately reflected their identity. They were not confused and they were not lying. A great many Australians sit in genuinely ambiguous territory, somewhere between a christening they do not remember and a conviction they have never quite formed. They are not withholding a settled answer that the right question would unlock. The question is quietly deciding which side of the ambiguity gets written down, and it is doing so invisibly.

This ought to embarrass both lobbies, because both have spent two years insisting the question is biased against them and neither bothered to run an experiment. The Rationalist Society wants a yes/no filter. The Bishops want the status quo. What our data shows is that the most inflationary design in Australian history was not the current one but the 1971 wording, that innocuous instruction to write none, which added 27 points over an open question. The bias has not always run in the direction people assume. It has simply run.

The ABS, to its credit, half knows this. Its own documentation notes that moving No Religion to the top in 2016 ‘may result in higher responses’ in that category. Its 2016 methodology paper traces the same pattern back to 1933. The Bureau has understood for decades that the instrument moves the result. What was missing was a number. The number is 13 points.

They proposed a change for this year’s Census. Both camps lobbied. In February 2025, the Bureau reversed itself and kept the old design, citing comparability. This is the part that defeats me. The argument for keeping the question is that changing it would break the time series. The time series broke in 1971, and again in 1991, and again in 2016. We have simply been treating a discontinuous line as though it were continuous and building policy on the difference.

There is a fix, and it is almost insultingly cheap. Randomise the order of the options. Every household sees the same categories in a different sequence, so nothing benefits from being first. It is standard practice in survey research and has been since before most of us were born. 79 per cent of the Census is now completed online, which makes it a configuration setting rather than a redesign. We already do exactly this with ballot papers, having accepted long ago that top billing is worth votes.

The 2026 Census will cost $726 million. For that we hire 30,000 people and build a national digital platform. Randomising a list costs approximately nothing.

If we are going to spend three quarters of a billion dollars counting the country, we might at least count it properly.

Dr Dorian von Freyhold is a survey methodologist at Compass Polling. The study surveyed 5,012 Australians in June 2026 and replicates an earlier trial of 2,005 respondents.