AI transparency

In short

Synthestat results are produced by AI models simulating a synthetic population. No real person answered. Figures are estimates with stated uncertainty, not responses from real people and not survey measurements, and they may be biased where the source data or models are. Use them to explore and prioritise, and check consequential decisions against real data.

Briar Hamsfest Consulting s.r.o. provides the Synthestat AI system under its own name and uses AI models from third parties inside it.

Where the portal uses AI

• Survey runs: large language models answer questionnaires as synthetic respondents drawn from a synthetic population. For closed questions a run can use the “Jev” model (TypeSafe), which picks typed answers. Tables, percentages and uncertainty are then computed in code from those answers.
• Focus groups: the participants are AI-simulated personas, not real people, including when you moderate a session live. Transcripts and their analysis are AI-generated.
• Simulations and the Replication Lab: synthetic people and twins act or answer through language models; the Replication Lab's optional judge, leakage probe and literature search also call a model.
• Profiles and scenarios: models summarise news sources for media-diet profiles and draft reusable scenario templates.
• Media contrasts: speech in media items is transcribed by a speech-recognition model running on our own servers.

Multiverse comparisons and Serious Gaming exercises are computed in code; they call no model. When no AI model is available, the sandbox engine fills a run with rule-based answers computed in code; such output is labelled as sandbox output and marked as synthetic, not AI-generated.

Models are reached through OpenRouter (optionally only through its EU endpoint) or directly from their providers; the privacy notice lists them. Your account's model settings show which models your organisation can use and what a request costs, and each run records the exact model it used.

Academy videos

The video guides are animations of live copies of portal pages with artificial example data, drawn in your browser. The pictures are not AI-generated. The voiceover is: it is generated by Deepgram Aura-2 through OpenRouter, with a stock synthetic voice (“aura-2-thalia-en”) that does not imitate a real person. The narration is written with AI assistance and reviewed by people; Briar Hamsfest Consulting s.r.o., the operator, holds editorial responsibility for it (contact info@synthestat.com). Each guide starts with a notice that the voice is AI-generated, shows the EU “AI GENERATED” label from its first frame, and names the voice under the player.

The EU "AI GENERATED" label

These are the official icons the European Commission published with the Code of Practice on marking and labelling AI-generated content (Annex 1). “AI GENERATED” marks content generated by AI, such as the Academy voiceover. The round “AI” icon marks AI use together with our own text, as in the footer, and links to this page. The icons are free to use; using them does not mean that we have signed the Code.

Machine-readable marking

Article 50(2) of the EU AI Act requires providers to mark AI-generated content in a machine-readable way. Synthestat has not yet been placed on the EU market. Article 50 applies to it from the day it is; because that is after 2 August 2026, the transitional period until 2 December 2026 for systems placed on the market earlier does not apply. The marking below is in place from launch.

We mark these downloads:

• Report PDFs of survey runs and studies, focus groups (including the transcript) and the Replication Lab: in the document metadata and an XMP packet.
• The Replication Lab comparison (JSON) and questionnaires the agents generated (JSON export): in an ai_marking object.
• Academy voiceover audio files: in ID3 tags.

The marking states that the content is AI-generated, the generator (Synthestat, Briar Hamsfest Consulting s.r.o.), the models the run recorded, the IPTC digital source type trainedAlgorithmicMedia, the generation time and a link to this page. Output of the sandbox engine, which runs when no AI model is used, is marked as synthetic but not AI-generated: ai_generated: false, synthetic: true, generator “Synthestat sandbox (no AI model)”, digital source type algorithmicMedia.

Single words, coded answers and numeric tables are not marked one by one. Questionnaire files for survey platforms (for example QSF or XLSForm) carry no marking, so that those platforms can import them; use the JSON export when you need the marked version.

Signed marking and how to check a file

We sign the marking of report PDFs and JSON exports with our Ed25519 key. The signature covers the marking, the SHA-256 checksum of the file's content and the time of signing, so it shows that Briar Hamsfest Consulting s.r.o. issued the marking and whether the file changed afterwards. The time comes from our server's clock; no independent time-stamping authority is involved. Academy voiceover files are signed only when they are tagged with the operator key.

Our public keys are listed at /.well-known/ai-marking-keys.json. A key that has been replaced stays listed as retired, so files it signed can still be checked:

Loading keys…

To check a file, send it to /api/ai-marking/verify, for example curl -F file=@report.pdf https://<portal>/api/ai-marking/verify. The answer says whether the signature is valid and shows the marking. Or ask info@synthestat.com.

Human oversight

You decide what to run: the questionnaire or discussion plan, the population, the scenario and the model. Synthetic answers, transcripts and narratives are generated automatically and are not reviewed by a person before you see them. Where a workflow has a human approval step, such as the risk review and publication gate for high-risk studies, the portal shows it. Review outputs before you rely on them or pass them on, and state the extent of human review when you publish (as the ICC/ESOMAR Code asks of research that uses AI). Editorial responsibility for the Academy narration lies with Briar Hamsfest Consulting s.r.o..

Limits of synthetic data

• Synthetic personas imitate patterns in source statistics; they are not people and have no opinions of their own.
• Models can be biased, can compress the spread of opinion, and can be weak for small groups or topics their sources do not cover.
• Results change with the model, the population release and the wording; each run records what it used.
• Agreement with published figures is not proof that a result is right for a new question.

Using and publishing outputs

Say that outputs are synthetic and AI-generated when you publish or pass them on, name Synthestat as the source, and do not present them as observed survey data. If you publish AI-generated text on matters of public interest, label it or take editorial responsibility for it. The terms of use set out the full rules. Questions about AI in the portal: info@synthestat.com.