GretchenAI: Check an image, then make up your own mind

What GretchenAI offers

The provider describes a web dashboard that accepts files or links. It aims to detect signs of manipulation and research the origin and context of media online. Results are meant to include clues and sources. We have not tested how well this works in a particular case.

Source: Gretchen AI: Produkt

A real photograph can tell a false story

Imagine a photograph of a flooded street. Its caption claims it was taken today in your town. It could actually be years old and from another country. This is an invented example to explain the distinction, not a GretchenAI test result.

Looking for artificially generated parts of the image would not, by itself, resolve the false location. You would need the original post, a date or other images of the same event. Even if the photograph, place and date are genuine, a further claim about the cause could still be wrong.

Our guiding question is therefore: which part of the post was actually checked? A finding about the image does not automatically establish whether the accompanying claim is true.

Where detection can go wrong

GretchenAI acknowledges false alarms on real images and missed fakes. Screenshots and heavy compression make analysis harder. The provider describes a focus on faces and images centred on people; landscapes and objects are harder to assess. Dates found during context searches can also mislead.

Source: Gretchen AI: FAQ

What a high detection rate does not tell you

Research shows why everyday material needs to be distinguished from test conditions. Deepfake-Eval-2024 examines media that actually circulated online. The open detectors studied performed substantially worse than on earlier benchmark datasets. The commercial models evaluated also fell short of the human specialists. These findings do not establish an error rate for today’s GretchenAI dashboard.

Our conclusion: an advertised detection score does not tell you how reliable the assessment of your particular image will be. Understanding that would require information about the method, test data and errors. We therefore do not turn a marketing figure into a promise of success.

Source: Chandra et al.: Deepfake-Eval-2024

Access, cost and what happens to your images

As checked on 11 October 2026, the pricing page lists a Starter plan at €9.99 a month with 100 scans. It also advertises a free 14-day trial with a limited allowance and no credit card required. We have neither created an account nor checked the checkout process. Review the conditions of the specific offer before signing up.

The privacy notice we read mainly describes website use. We do not take it as an assurance about how long dashboard uploads are stored or how they are used. We have not fully resolved those questions. For an initial experiment, we would therefore choose a non-sensitive file without private or confidential information.

Source: Gretchen AI: Preise · Gretchen AI: Privacyerklärung

Why its use in Berlin sparked a debate

On 26 August 2026, Berlin’s interior administration announced the “Gretchenfrage” service ahead of the 20 September election. People would be able to submit suspicious posts via Instagram. The administration described its aim as traceable guidance with sources and explicit uncertainty. This records the announcement at the time; we have not checked operation after the election.

In a Berliner Zeitung opinion piece on 27 August, Marcel Luthe criticised the state-backed service as paternalistic. His criticism concerns the state’s role in interpreting political content. It does not, on its own, demonstrate a technical fault in the tool.

For Hexpex, this raises a question that matters beyond elections: can you see which sources support an assessment, what remains unknown and how to challenge a mistake? An official sender cannot replace that scrutiny, nor can a familiar company name.

Source: Berliner Innenverwaltung: Start der Gretchenfrage · Marcel Luthe: Kommentar in der Berliner Zeitung

What to do with a result

Open the cited evidence. Does it address the same image and the same claim? Check the original publication date and whether a later page is merely repeating an older report.

If you cannot follow the assessment, the case remains unresolved. You do not have to turn an uncertain answer into “real” or “fake”. If you share it, a precise description helps: “I have not yet been able to check where this came from.”

GretchenAI would be useful to us if it provided a verifiable lead for our own research. This source review does not establish whether it does that reliably for your images. Our checking guide explains how to investigate origins and context yourself.

Open the Hexpex guide to checking images and videos ↗