Asked Twice, a Reading Should Not Move

The 29th ACM Conference on Computer-Supported Cooperative Work and Social Computing opens today in Salt Lake City and runs to October 14. One of the papers accepted to it studies something this site has a direct stake in: how people in China actually use generative AI for fortune-telling. Long Ling, Xiyu Zheng, Gengchen Cao and Ray LC interviewed 22 regular users aged 18 to 52, and ran a three-week digital ethnography in March 2025 across 1,842 posts and interactions on RedNote, WeChat and the community forum of the Cece app. The preprint has been public since late March, so the text is not new; the occasion is. Their participants found the tools accessible, convenient and efficient, and said the readings felt flat and without ritual. The finding worth sitting with is about what people did when they did not like the answer they got.

“

If I feel there's a slight deviation, I'm not satisfied, I might recalculate

The authors describe users asking again and again, rewording the question slightly each time, until a reading feels satisfactory, and in their discussion they tie it to confirmation bias. No tradition can stop anyone doing that, and nobody needed a study to suspect it happens. What the study is useful for is sharper: it makes you ask what exactly was being re-asked. Because the systems we compute here divide cleanly in two, and which half you are in decides whether a second ask is bias or a bug report.

Two things that look identical from the reader's chair

The first half is counting. Take today. October 10, 2026, is the first day of the ninth lunar month. The year pillar is Yang Fire Horse, the month pillar Yang Earth Dog — the month that opened at Cold Dew on October 8, which is why English-language columns are calling this the Earth Dog month — and the day pillar Yin Fire Snake. The day officer is Danger. Its lunar mansion is Willow (Earth Roebuck). Every one of those is a position in a cycle, derived from a rule one or two lines long. Ask again tomorrow what today's day pillar was and the answer is Yin Fire Snake, because there is no second way to count to Yin Fire Snake.

So when the study reports community members trading prompts to correct the models' miscalculated lunar dates and BaZi charts — it does, in its section on community practices — those users were not re-rolling until they liked the answer. They were debugging, and they were right to. A model that returns a different day pillar on the second ask has not offered a second opinion. It has made an arithmetic error, and the error is findable, because any printed almanac prints the same Yin Fire Snake.

The second half is reading. What a Yan Nian (longevity) star between two people is telling them, whether a Danger day suits the particular thing you are planning, what it means for you that this month's pillar is an earth month — these are interpretation. Ask twice and you get different words, and nothing adjudicates between them. This is where the paper's confirmation-bias reading bites hardest, and it is also where no better model helps at all, because there is nothing here to be more accurate about. A reader who re-asks until the words please them has not been misled by a machine; they have been handed a layer that was never going to push back.

A count can still have a margin, and today's has one

Today is the right day to say the harder half of this, because today's own lunar date is the borderline case. The ninth month begins at the new moon, and the month start is defined by which Beijing calendar day that new moon falls in. This site's code puts it at 23:50 Beijing time on October 10 — ten minutes before midnight. Ten minutes decide whether the first day of the ninth month is today or tomorrow, and every lunar date for the rest of this month moves with it. Two sources computing the same three rules with slightly different lunar theory can honestly land a day apart here, and neither is doing anything wrong.

That is not a license to keep asking until you get the month you wanted. It is the one case where the margin is the answer. Our almanac prints it: a new moon falling within a quarter of an hour of Beijing midnight is reported on the page rather than smoothed over, because a reader who knows the boundary is ten minutes away can go and compare two almanacs and see which side each one took. A reader handed a flat date has nothing to compare. The honest version of a determinate claim is the claim plus its margin, and a tool that quietly re-rolls near the boundary is doing to itself exactly what the study's participants were doing by hand.

The test, and it applies to us too

That pair is the whole argument in one example. The inputs are a birth date and a recorded sex; the boundary is a solar term; the output is eight directions. It is checkable end to end, and anybody who disagrees with it has to disagree about a count. What it cannot tell you is whether you should put your desk against the west wall this year, and we should not pretend otherwise.

Our honest read

We should say what this study is about us, since it plainly is. We have no mystique either, and that is a decision rather than an omission: no ritual, no ceremony, no score at the end. The participants who missed those things were missing something real, and we do not supply it and do not intend to. What we offer instead is the one property the re-prompting behavior destroys — an answer that does not move, with its margin stated where it legitimately could. That is a much narrower product than a fortune. It is also the only one we can stand behind.

And the finding is not really an indictment of AI. A person who dislikes an answer has always been able to ask a second practitioner, or keep the almanac that said what they wanted and put the other one down. What a model changes is only the price of the second ask, which is now nothing, and the speed, which is now instant. The defense has not changed either: a reading with its derivation written down, so that somebody who asks twice can see which answer moved and go and establish which one was wrong. That is also why we will not grow our list of portents past the six rules we can state and check, and why the tools warn instead of guessing when a facing sits near a 24-mountain boundary. A confidently wrong chart is the thing a second ask cannot fix.