Fully generated documents
A receipt, an invoice or a pay stub created from scratch by ChatGPT or Gemini. The output looks clean, but the texture of the image gives the generator away across the whole surface.
4.6 / 5 on Trustpilot
Detect an AI generated document
Detect an AI generated document
in 10 seconds
K-Detect, Koncile's vision model, spots receipts, invoices and IDs generated or edited by ChatGPT, Gemini or Flux. Upload a photo or a scan, then create an account to run the analysis.
- 22k+ users
- SOC 2 and GDPR
- Try for free
- SOC 2 Type II, GDPR
- Deleted after analysis
- Never used to train an AI
Create a free account to run K-Detect, our detection model for images and documents generated or edited by AI.
Create a free accountFor document volumes or an API, talk to an expert.
The AI generated document detector used by insurers, banks and large companies
4.8 / 5
How to tell if a document was generated by AI?
By eye, it has become close to impossible: across 287,000 images judged by 12,500 people, a 2025 Microsoft study measured 62% correct answers, barely better than chance. Three methods exist, from the simplest to the most reliable.
- MetadataSince August 2026, the EU AI Act requires generators to mark their images in a machine readable format. It is the first check, but a screenshot or a re-save wipes it out.
- WatermarksSome generators hide an invisible watermark in the image, such as Google's SynthID or Meta's marking. Not all of them do, and each one can only be read with its maker's own tool.
- Training on documentsThe most reliable signal is still a model trained on real and fake documents. That is K-Detect's job: it reads the image at full resolution and recognizes each generator's signature.
Comparison
Why not just ask ChatGPT?
On 314 receipt photos including 114 fakes, never seen in training, K-Detect, our machine learning model, catches 92% of the fakes, against 56% for the best LLM.
LLMs read the text, not the pixels
They miss almost every realistic edit
K-Detect
The AI generators K-Detect detects
K-Detect is trained specifically to detect the latest image generators fraudsters are likely to use, from frontier labs like OpenAI, Google, Black Forest Labs, ByteDance, Alibaba, xAI, Meta, and more. Koncile updates its model every month, checking its accuracy against the latest frontier AI models available.
- OpenAIGPT Image 1, GPT Image 1 mini, GPT Image 1.5, GPT Image 2, GPT Image 2.5, GPT 5 Image
- GoogleGemini 2.5 Flash Image (Nano Banana), Gemini 3 Pro Image, Gemini 3.1 Flash Image, Gemini 3.1 Flash Lite Image
- Black Forest LabsFLUX.2 Pro, FLUX.2 Max, FLUX.2 Flex, FLUX.2 Klein, FLUX 3, FLUX Kontext
- ByteDanceSeedream 4.5, Seedream 5 Flash, Seedream 5 Lite, Seedream 5 Pro, SeedEdit 3
- AlibabaQwen Image 3, Qwen Image 3 Pro, Qwen Image Edit, Qwen Image Max
- xAIGrok Imagine
- MicrosoftMAI Image 2.6
- TencentHunyuanImage 3, HY Image 3.5
- MetaMuse Image
- KuaishouKling Image O1, Kling Image O3, Kling Image V3
- RecraftRecraft V4
- OthersLuma Photon, Ideogram V3, Krea 2, Bria FIBO Edit, HiDream O1, Step1X Edit, OmniGen 2, Bagel, Riverflow 2, Ming Image
K-Detect
What K-Detect catches in a document image
K-Detect is a vision network trained by Koncile. It does not read the text of the document: it looks at how the image was produced.
Photos edited by AI
The most common case: a real photo where a generator rewrites the total, the date or the merchant, and regenerates a large part of the image on the way.
IDs and vehicle registrations
The model also learned from 2,625 real customer documents and their edits: IDs, vehicle registrations, driving licences, pay stubs and bank statements photographed.
Use cases
Where AI generated documents show up
Wherever a photo of a document counts as proof, an image generator is enough to fake it.
Expenses
Restaurant, taxi or fuel receipts edited to inflate an amount, read before every reimbursement.
Insurance
Repair invoices, claim photos, proofs of purchase: every document in a claim file checked on arrival.
KYC and KYB
IDs and supporting documents photographed at onboarding, analyzed together with their content.
Yes, in most cases. ChatGPT, Gemini or Flux leave a texture in the pixels that comes from how they work. K-Detect, trained on fakes from 39 generators, recognizes it even when the document looks perfect to the eye. The C2PA signature some tools add helps when it survived, but it disappears with the first screenshot.
A general purpose detector is not: it learned on photos of faces or landscapes. On receipts, K-Detect rates 98.1% of fakes as medium or high risk and wrongly rates only 1% of real receipts as high risk. The score remains a signal to verify.
Because an LLM reads the text and does not look at the pixels. On our test set, general purpose LLMs catch 31 to 68% of fakes and almost no realistic edit.
Photos and scans of documents: receipts, invoices, IDs, vehicle registrations, driving licences, pay stubs, bank statements. For a native PDF, Koncile also reads the file itself: producing software, versions, fonts.
The first document is free once you create an account. For volume or API access, the Advanced plan is €199 per month, about €0.20 per page.
No. It is analyzed securely, deleted after the analysis and never used to train a model.
Other documents
Explore our document specific detectors
Each document gets its own method: a proprietary ML model for photos, known metadata and forensics for native PDFs, and contextual cross checks.
Bank statement
Inflated balances, deleted lines, a PDF rewritten with software other than the bank's.
See the detection
Pay stub
Net pay changed, contributions that do not match gross pay, a layout that does not match the payroll software.
See the detection
ID document
Photo replaced, date of birth or number changed, an MRZ that no longer matches the fields.
See the detection
Invoice and receipt
Edited amounts, wrong VAT, duplicates and made up suppliers.
See the detection
Try Koncile's fraud detection on your own documents
Upload your bank statements, pay stubs or IDs: Koncile extracts the data and flags fakes in seconds.
Talk to an expert