How the AI report works

This article explains how Maptionnaire's AI report is built: how it turns open-ended answers into tags and a narrative, why it's built that way, what model runs it, and where your data is processed. For the steps to generate and export a report, see Generate a Survey Results Report Using AI.

How the AI writes the narrative

Language models are unreliable when handed thousands of open answers and asked to summarize them directly. Asked to summarize, a model isn't counting — it has no dependable way to track how often something was actually said, and what it surfaces depends partly on where an answer sat in the pile.

So Maptionnaire never asks the model to summarize your raw data. Open answers go through three separate rounds instead:

Round 1 — Propose tags. For each open question, the AI reads the answers and proposes the tags that actually recur in them. It returns at most ten tags per question, and is instructed to prefer fewer.

Round 2 — Apply tags. The AI goes through the answers one at a time and assigns each answer the tags that fit it. An answer can get several tags, or none. Very short answers are skipped.

Round 3 — Write the narrative. Only now does the AI write, and it does not see your responses at this point. It sees the tag statistics ("mentioned in 87 of 421 answers") and up to five example answers per tag. A tag one person raised cannot be written up as a common concern, because the count says otherwise.

The narrative is a description of the tags. If the tags are right, the narrative is right — which is why checking the tags matters more than checking the prose.

Why it's built this way

This is how qualitative researchers have analyzed open text for decades: content analysis. Agree a set of tags, apply them to every response, then report how often each one occurs. Maptionnaire does the same thing, with the AI as the coder and you as the reviewer.

That split matches what language models are actually good at. A 2026 study comparing GPT-5 to human coders on survey text found LLMs align much more closely with human coders when applying a fixed set of tags (Adjusted Rand Index 0.61, close to the 0.68 measured between human coders) than when generating the tags in the first place (0.54) — Bergmann et al., "LLMs for Survey Text Analysis" (preprint). Round 2 (applying tags) is the reliable part. Round 1 (proposing tags) is the step worth your attention, which is why you can replace it with your own tags at any time.

Maptionnaire follows the guidance of AAPOR's Task Force on Responsible AI Integration in Survey Research (2026), which addresses this exact task — an AI sorting open answers into tags:

Remove bias by defining your own tags

If you're unsure whether the AI's own tags slant the results, don't rely on our word for it: click 'Edit' next to the tags group, type the tags you want, and re-tag. The AI then sorts answers into your tags, and its own reading of the data no longer shapes what the report can say. Everything downstream — the counts, the narrative, the quotes — follows from tags you chose.

This is also the more rigorous way to work if you already have a coding frame, a policy vocabulary, or tags from an earlier round of engagement that this survey needs to be comparable with.

Quotes are copied, not written

The AI does not type the quotes in your report. It picks an example answer by number, and Maptionnaire inserts that respondent's text verbatim. A quote in the report is always something a respondent actually wrote — it cannot be paraphrased, merged, or invented.

Cluster analysis is arithmetic, not AI

The map cluster sections work the same way: the analysis is mechanical, and only the wording is written by AI.

• Clusters are found with a standard density-based clustering algorithm (DBSCAN), which groups map pins that sit close together. You control it: from Clustering settings, set the cluster radius and density, preview the clusters on a map, and pick which ones the report should cover.
• For each cluster, Maptionnaire compares the responses inside it against the responses in the rest of the survey, and keeps the differences that are statistically significant. You can open the list of these differences per cluster and read the numbers.
• The AI's only job is to put that comparison into a sentence.

What the AI does not decide

Charts, figures, and percentages are computed from your response data the same way as everywhere else in the analysis tool. The AI groups elements into sections and writes prose; it never produces a number.

Nothing in the report is final: every section can be edited, reordered, or removed, and the underlying views are ordinary saved analysis views you can open and change.

The model, and where your data goes

As of August 2026, Maptionnaire uses Claude Sonnet 4.6, running on Amazon Bedrock — the same AWS infrastructure that hosts the rest of Maptionnaire. Every step described above uses it: reading the answers, tagging them, and writing the narrative.

Your responses are processed in AWS's US and Ireland data centers — the same regions where Maptionnaire already stores your data, not just "the same part of the world." See the Compliance page for the full processing and subprocessor details, and Customer Privacy for the Data Processing Agreement, which includes the European Commission's Standard Contractual Clauses. The model does not retain your responses after processing them, and does not use them to train models.

Limitations

Less widely spoken languages. AI models are trained on far more English than Finnish, Somali, or Greenlandic, and tagging tends to be correspondingly less sure-footed in those languages. If your survey ran in a less widely spoken language, give the tags a closer look than usual before trusting the narrative.

Small sample sizes. With very few responses to an open question, tag statistics like "mentioned in 3 of 8 answers" carry little weight, and a single respondent's wording can look like a pattern.