Turkey's virtual public
What would Turkey say?
We asked 20,000 virtual people who learned from real surveys. Make your guess first, then see the answer and how reliable it is.
Example question
20,000 virtual peoplelearned from 27,688 real interviews243 tests, all public, failures includedWhat the virtual panel cannot do: vote forecasts · province detail · ad reactions
Game · Guess
How well do you know Turkey?
Six questions. Guess first, then see the virtual panel's answer. The panel's answer is an estimate too, but one tested against independent surveys.
Surveys it learned its views from: WVS · EVS · ESS · ISSP · CSES · Pew · LiTS — 27,688 interviews conducted in Turkey. Eurobarometer and Gallup surveys were used only for testing. Population and vote: TurkStat, Election Board.
The street
160 people, out of 20,000
Every face is a record from the 20,000-strong virtual panel, not a real person. Age, province, education and facial expression (life satisfaction) come from that synthetic record.
Everyone in this crowd is a synthetic record from the virtual panel, not a real person · hover over them
Your twin
The person in virtual Turkey most like you
Choose your age, gender, education and province, and we'll show you the person in the virtual public most like you, and what they say. Your choices are not sent anywhere and not saved.
How it works
Build, test, ask
Build
Population, province, age, education, work and vote from TurkStat and the Election Board; values and views from 27,688 real survey respondents.
02Test
Results on real survey questions are checked against independent surveys the panel has never seen (scenarios and provincial values are outside this testing). The rule is written before the result; failed tests are published too.
03Ask
See the AI's experimental, unvalidated estimates for questions nobody has asked, by age, province and party.
The country
Meet Ahali
The same faces, as a globe. Drag to spin it, click a face.
160 virtual citizens · drag · click
What it is for
What it knows best
Where it is strong
- Values and religionIt knows who is more religious and who is more conservative, very close to real surveys. Where groups clearly differ, it gets the order right 97 times out of 100.
- The make-up of party basesWe checked which education and age groups each party's voters come from against independent election surveys. Education came out very close; age is still weak, and we are improving it.
- Things that do not changeWhich groups lead on values barely changed from 1990 to 2018; what changed are the shares themselves.
- Audience and channelIt knows, very close to real data, which groups use which channel (TV, social media, internet), who shops online and who is self-employed.
How to read it
- Elections and votesThe virtual panel does not forecast votes: party figures come from actual results, and the views of party groups are estimates. The forecast on the Election page is a separate method: the average of polls and its track record in past elections.
- Provincial detailProvincial values come from regional and population structure; they are sturdier at regional level.
- Fast-changing topicsOn topics that change every year, such as migration, the EU or the press, read it alongside a recent survey.
Experimental estimates for questions nobody has asked
An ad campaign, a new law, a price rise: see the AI's unvalidated estimates for ready-made scenarios, by age, province and party.
Plate II · Atlas
Turkey province by province, district by district
Turkey, province by province, as the virtual panel answers.
Colours show the estimate for each province on the question you pick. These estimates come from region and the age and education make-up of the population; each province was not checked separately. Even small gaps can show up as different colours. Click a province to open its districts with the real 2024 local election results.
Turkey · 81 provinces
Differences between provinces are usually small (often about 8 points between the highest and lowest province) and come mostly from region and population make-up; provinces were not measured one by one. That is why maps for different questions look alike. Colours are set to each question's lowest and highest province.
Finding
Where age, sex, education and regional groups are closest and furthest apart
For each question, the largest gap between age, sex, education, region and urban-rural groups. On some topics every group says almost the same thing.
Segments
What do Turkey's segments say more, and say less?
Distinct parts of the population, with the views on which they differ most from Turkey overall. The line is Turkey, the dot the segment.
Where is Turkey divided?
Same question, very different answers
The sharpest divides the virtual panel found. Each card shows the largest gap between groups.
Over time
How has Turkey changed?
Not a model: change measured in real surveys repeated in Turkey. The shaded band is the survey's own margin of error.
Source: World Values Survey, Turkey: 1990, 1996, 2001, 2007, 2011, 2018.
Explore for yourself
Pick a question and a group
The orange line is the Turkey average.
AI · innovation
How well does AI know Turkey?
On topics asked in a real survey but never seen by Ahali, we compared the AI's group-by-group estimates with the real results (TAYA 2016).
On shares the AI was off by 10 points on average and by more than 25 on some topics (e.g. religion in choosing a spouse: estimate 49%, actual 76%). But it mostly got right which group is ahead. TAYA results are published, so the AI may have seen them before; that is why we also tested it on surveys that were never published. Backed by our data, the AI ranked the groups correctly on most topics; without data it knew nothing. Our data does the work: the AI never works alone, only together with real data.
Scenarios · experimental
What would Turkey say to questions nobody has asked?
Topics that real surveys haven't covered: an ad campaign, AI assistants, the economy and working life. Here the AI gives the answers: it plays 240 virtual citizens one by one, and the results are added up to match Turkey's population.
These results are not validated. On similar questions AI on its own is off by about 9-10 points; read them as a rough direction, not exact figures.
Real network · #Secim2023
Twitter in the election: share within your camp, argue across
67.8 million Twitter interactions during the 2023 election (March-June; shares and charts use April-June, leaving out an unusual March surge) and the politician-follow network of 37.7 million accounts. No tweet text; who retweeted, replied to and followed whom.
Twitter users do not represent Turkey (younger, urban, politically engaged). These are findings about the digital public sphere, not the population.
Interactions between users with an identified camp, April-June 2023. Twitter users do not represent Turkey.
From camp → to camp, %
Camp hashtags (April-June)
Campaign slogans dominate the last two months. On 18-26 March (omnibus bill period) occupational groups' demand hashtags surged: {mart}. That surge also coincides with a change in data collection; the analysis was repeated for April-June and camp shares did not change.
Daily volume: the balance flipped between election days
The opposition was more active on 14 May, the government side on 28 May. Which side was more active on Twitter did not predict the result.
Follower growth is not votes (1 March = 100)
Kılıçdaroğlu grew fastest before the election; Erdoğan grew among the slowest and won. Oğan jumped after the first round and lost followers after endorsing Erdoğan.
Source: #Secim2023 open dataset (Twitter interaction and follow networks, 2022-2023). Analysis: Ahali.
Ahali Apartments
Door to door, Turkey
Pick a province; panel records from that province move into the building. The building answers every question through its windows. Click a flat: the camera moves in and shows this virtual person's details and estimates.
Voices
The people behind the numbers
Pick a question and meet the virtual citizens who speak to it. What they answer is decided by statistics; AI only puts that answer into their own words.
Why should you trust it?
We tested the virtual panel against a real survey
Similar projects have AI invent their virtual people and never measure how accurate they are. We compared ours with the answers of 1,179 real people the panel had never seen.
Got the direction right
97 / 100
When there was a clear difference between two groups (say, young and old), the panel correctly said which group was ahead in 97 out of 100 differences.
How close?
2.7 points
For large groups such as age, education and region, our prediction was 2.7 points away from the real survey on average. For provinces and small groups it is not this close.
AI on its own?
9.2 points
We asked AI the same questions on its own: it was 9.2 points off. It assumed people were more religious than they are and always gave the 'expected' answer. That's why we never use AI on its own.
Where it was measured clearly
What do we know?
We compared our predictions with real surveys
🔒 can't be memorised For questions with this mark, the survey we compared against was never published anywhere, so the AI could not have known the answer in advance.
Sealed forecasts
Written and sealed before the results
Report card
What do we know, and how well?
The longer the bar, the better we got the differences between groups (women and men, young and old). The bar never fills completely: even two halves of the same survey don't match perfectly. The best achievable value is 0.93.
By topic
By group
Average miss by method (points; lower is better)
How many points each method missed the real share by, on average. For scenarios: the miss we expect on questions never asked before.
Legal information
Imprint, personal data and terms of use
In short: we don't ask for your name, email or phone number, and we use no cookies, ads or tracking tools. Details below.
Imprint
- Site
- Ahali
- Owner and responsible person
- Emirhan Taftaf ve Doğa Akdeniz
- Address
- İstanbul / Maslak
- emirtaftaf@gmail.com
- Hosting
- Cloudflare Pages (Cloudflare, Inc.)
- Nature
- Non-commercial education and research project
Published as required by Article 3 of Turkish Law No. 5651.
Personal data: privacy notice
Data controller: Emirhan Taftaf ve Doğa Akdeniz (contact: emirtaftaf@gmail.com).
What data: When you open the site, your browser technically sends your IP address, browser details and the address of the page you open to the hosting service (Cloudflare) and the font service (Google Fonts). Like on any website, this is needed for the page to reach you. We keep no separate record of it and make no attempt to identify you.
We don't ask for your name, email or phone number. Anything you choose on the site (e.g. age and province in "Your twin") is processed only in your own browser; it does not reach us and is not stored.
What for: To deliver the site to you and protect it against attacks.
Legal basis and method: Collected automatically when you connect; the legal basis is legitimate interest (Turkish Personal Data Protection Law No. 6698, Art. 5(2)(f)).
Who receives it: Only the two services above. Their servers may be outside Turkey, in which case your data is transferred abroad (Art. 9).
Your rights (Art. 11): To learn whether your data is processed, request information, learn the purpose, know who receives it, ask for correction or deletion, object to an adverse result produced solely by automated analysis, and claim compensation. Write to emirtaftaf@gmail.com; we reply free of charge within 30 days at the latest (Art. 13).
Cookies and browser storage
We use no cookies. There are no ads, tracking or visitor analytics tools.
Purely for your convenience, a few small items are written to your browser's own storage; none of them is sent to us: your light or dark theme, the last question you viewed on the Questions page, your Question of the day guess and your best score in the guessing game. To delete them, clear this site's data in your browser.
Terms of use
- Ahali is a non-commercial education and research project; no services are sold.
- Results on the site are a virtual panel's estimates, not a real survey, and describe groups, not individuals.
- The virtual people's words were written by AI; they are not real people's words.
- The election page forecast is the average of published polls; it does not come from the virtual panel.
- Nothing here is investment, political, legal or health advice. We give no guarantee of accuracy; do not base important decisions on this site alone.
- Data from the institutions we cite (TurkStat, the Supreme Election Council, the Central Bank, polling firms and others) belongs to them; they are not affiliated with Ahali and are not responsible for the results here.
- Ahali must not be used for individual targeting, misleading propaganda or voter manipulation.
- Please cite Ahali when quoting.
- To ask for content to be corrected or removed, write to emirtaftaf@gmail.com.
Last updated: 29 October 2026
Sources and licences
Every number has a source
Ahali does not publish any source's raw data; the site shows only the virtual panel's aggregated results and indicators the sources themselves publish. Each source is cited under its own terms of use.