TL;DR
ChronoCulture is a study that tested whether models can predict how people's views in a country have changed. Models were given a country, a year, a question from a social survey, and the answer options. They were asked to estimate the percentage of people choosing each option, and the estimates were compared with real data from two years: an older one and a newer one.
The headline finding: a model produces nearly the same picture for 1999 and 2022, even though real people's answers diverged sharply. For most questions with a noticeable shift, both forecasts landed closer to the same year. The model nails the country's “general vibe” but misses the dynamics. When the shift is large, it underestimates its size. And where nothing changed, it “invents” a change in a third of cases. If the prompt has no year, the answer tends to resemble an earlier era rather than today.
A simple hint doesn't fix it: adding “in 2022” to the prompt does almost nothing. It helps about as often as it hurts. The only thing that noticeably helps is naming the older year when the shift is very large. The practical takeaway: don't trust a model with “how have attitudes toward X changed” questions, and check its “as of today” answers against fresh data.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: Prompt with no year: country + question + options → percentage per option
STEP 2: Prompt with the older year → percentages
STEP 3: Prompt with the newer year → percentages
STEP 4: Compare with the real survey:
— how accurate each forecast is
— how different the forecasts are from each other
— whether that matches the real shift
(each prompt is sent separately, same prompt format every time)
Example in Practice
Task: An e-commerce analyst is putting together a deck on “how Americans' values have changed” and wants a quick estimate of the answer distribution for a World Values Survey (WVS) question: “Can divorce be justified?” on a 1–10 scale. The method's sweet spot is asking for a percentage distribution, not a single answer. The model will handle the dynamics poorly, so the check is built right into the prompting process.
Prompt (send it three times: no year, with 1995, with 2022):
Country: United States
Year: 1995 [in the second prompt use 2022; in the first, remove the “Year” line]
Social survey question: “Please tell me whether, in your opinion, divorce can be justified.”
Scale: from 1 (“never justifiable”) to 10 (“always justifiable”).
Estimate what percentage of the country's adult population would choose each option from 1 to 10.
The total must equal 100%.
Reply with a table only: option — percentage.
Result: The model will return three tables of 10 rows with percentages. The check: if the 1995 and 2022 tables are nearly identical, that's a sign the model can't tell the eras apart. In that case, don't build a “here's how Americans have changed” slide on its numbers. Use them only as a rough sketch of the “general profile” and go find real data from the WVS, Gallup, or Pew Research Center. For a “today” question, assume the answer may actually reflect an earlier era.
Why It Works
The model's weakness: it was trained on a huge pile of text where different eras are mixed together. A country's cultural attitudes get memorized as an averaged-out portrait. The year in the prompt is a weak hint: it doesn't rebuild the portrait, it only nudges it slightly. That's why forecasts for different years stay close together.
What the model can do: it does a decent job guessing a country's overall profile and the correct direction of change, especially when the shift is large. It still underestimates the magnitude. When the model outputs a percentage distribution, you can see nuances that get lost in a “single answer” format: the most popular option may stay the same while everything else moves a lot.
How to use it: ask for a distribution, not “what people think.” Compare answers across years. If they barely differ, treat that as a marker of time blindness. Your levers: - Naming the older year is the only one that works. It helps with very large shifts. - Naming the newer year is nearly useless. Don't count on “2024” bringing the answer up to date. - Leaving the year out makes the model lean toward an earlier era. For “current” estimates, don't rely on the default.
Prompt Template
Country: {country}
Year: {year}
Social survey question: “{question_wording}”
Answer options (in order): {options}
Estimate what percentage of the country's adult population would choose each option.
The total must equal 100%.
Reply with a table only: option — percentage.
Fill in:
- {country} — the country;
- {year} — a specific year (or remove the line for a “no date” prompt);
- {question_wording} — the question exactly as worded in the survey;
- {options} — an ordered scale (“strongly agree … strongly disagree” or 1–10).
To check the dynamics, send the template at least twice with different years and compare the tables.
Limitations
⚠️ No access to dynamics: models struggle to tell different years apart for the same country. For “how have attitudes changed” questions, the model's numbers can't be treated as a trend estimate.
⚠️ Underestimated magnitude: even when the direction is right, the size of the change is almost always smaller than in reality. Small or moderate shifts may go completely unnoticed.
⚠️ Phantom changes: where real views barely changed, models still draw a noticeable shift in a third of cases.
⚠️ A year in the prompt is no cure: it helps and hurts at roughly the same rate. The only gain shows up with very large shifts, and only for older years.
⚠️ Scope of the study: only ordered-scale questions from two major surveys were tested, with a single prompt format, no web search, and no step-by-step reasoning. Some conclusions by topic and region are cut off in the available text. Results are unconfirmed for questions without a scale and for open-ended answers.
How They Tested It
Researchers from Texas A&M took two long-running international surveys, WVS and ISSP. From them they built 10,360 pairs: the same question in the same country in an older and a newer year, each with the real distribution of people's answers. In total, 72 countries and years from 1985 to 2022. Six models (Llama-3.3-70B, GPT-4o, GPT-5.4, Qwen, and DeepSeek) were asked to predict the percentages for each year.
The key design choice: they scored the entire distribution, not the most popular answer. The authors showed why: in 68% of pairs with a noticeable shift, the most popular answer stays the same. Look only at that, and you'll miss the dynamics entirely. To avoid blaming the model for simply not knowing a topic, they first filtered for questions where it knew the country's profile in principle (keeping the top 75%).
The results aren't pretty. In 80–87% of pairs with a noticeable shift, both forecasts were closer to the same year than each was to its own. Anywhere from 10 to 56% of shifted pairs got fully identical forecasts. Without a year, the answer was closest to the most recent year in only 46–48% of cases. Specifying the year helped in 20–44% of cases and hurt in roughly the same share. Surprisingly, it was the older year that helped with large shifts, while the newer one barely did. In practice, this means the model struggles to “catch up” to the present and responds better to a “retro hint.”
Adaptations and Extrapolations
This is not tested in the paper, just an idea based on its findings.
🔧 Technique: compare two eras in one chat → a built-in blindness detector
Instead of two separate prompts, ask the model for both tables at once plus an estimate of the difference:
Give two distributions for the United States, for 1995 and for 2022, on the question “{question}”.
Then, in a single line, state by how many percentage points the share of “{option}” changed.
If you think you don't have enough real data, say so.
If the difference is suspiciously small or “0,” that's a marker of time blindness. This way you can quickly weed out the questions where the model can't be trusted.
Resources
- ChronoCulture: Evaluating Cultural Awareness Across Time in Large Language Models — Rusali Saha, Nitya Khurana, Ruihong Huang, Kuan-Hao Huang (Texas A&M University)
- Benchmark construction code: github.com/lab-flair/ChronoCulture
- Data sources: World Values Survey (WVS), International Social Survey Programme (ISSP)
- Related work: GlobalOpinionQA, OpinionQA, WorldValuesBench; Meister et al. (2025) on asking for distributions directly