After Algorithm

Digital content inspection methodology

Take back control of your content

Every day we read, copy and share text and images. Beyond what the eye can see, digital content can also carry invisible characters, provenance information, privacy metadata or markers added by some AI systems. After Algorithm acts as a technical revealer: it inspects, explains what it observes and lets you decide what to keep or change.

Signaux et niveaux de confiance

Four evidence levels, without shortcuts

AfterAlgorithm separates directly observable facts from elements that require interpretation.

Factual and deterministic

Bytes, metadata, Unicode characters or structures that are directly present.

Structural

A container or marker is detected without automatically claiming it is valid.

Cryptographic

A signature or provenance is called verified only when cryptographic validation has actually succeeded.

Careful interpretation

Heuristic or statistical signals remain indicators, never stand-alone proof.

Inspection is not cleaning

Inspection never modifies content silently. When a cleaned copy is available, AfterAlgorithm states what changed and keeps the original untouched.

Reveal structural traces

1. Text analysis

Digital text is not just a sequence of letters. It is made of Unicode code points, graphemes and sequences that can be invisible on screen. These elements do not prove AI authorship: they may come from an editor, copy and paste, a particular writing system, steganography or a watermarking system.

The text scanner works like a microscope. It locates observable elements, places them in context, separates anomalies from evidence, and only suggests changes that can be explained.

In practical terms, the tool looks for these invisible or unusual elements for you:

  • Invisible and zero-width characters such as ZWSP, ZWNJ, ZWJ, Word Joiner and Soft Hyphen.
  • Homoglyphs and mixed scripts, such as a Cyrillic “o” that looks like a Latin “o”.
  • Bidirectional controls that can change apparent text direction or order.
  • Unusual spaces, punctuation, Unicode normalization and character sequences.
  • Repeated structures that may act as a steganographic channel.
  • Statistical or provider watermarks when a compatible documented detector is available.

Text watermarks: what can be verified today

LLMs do not all leave the same signal. Google uses SynthID for some generated content. Anthropic says compatible Claude models released from August 2, 2026 embed a text watermark across supported distribution channels. Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry are third-party platforms that can provide access to Claude, not Anthropic products. After Algorithm only confirms a provider watermark when a compatible detector is actually integrated. No detected signal never proves human origin.

Model ecosystems the tool may be asked to inspect

This is a broad but non-exhaustive list. It does not mean every model below adds a watermark or that After Algorithm can currently attribute text to any of them.

OpenAI
ChatGPTGPT familyOpenAI APIGPT Image / DALL·E
Anthropic
ClaudeClaude API / PlatformClaude CodeClaude Cowork
Google
GeminiGemini APIGoogle AI StudioVertex AIImagen
Microsoft
CopilotMicrosoft FoundryAzure OpenAIPhi
Meta
Meta AILlamaMuse Image
Mistral AI
Vibe (formerly Le Chat)Mistral modelsMixtralMistral API
Amazon
BedrockNovaNova CanvasTitan family
xAI
Grok
DeepSeek
DeepSeek
Alibaba
Qwen
Cohere
Command
AI21 Labs
Jamba

Claude is also distributed through third-party cloud platforms including Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. Those platforms belong to Amazon, Google and Microsoft respectively; they provide access to Claude models but are not part of Anthropic.

AI / provenance

AI markers: what can actually be verified

Not all markers are visible or built in the same way. AfterAlgorithm separates deterministic observations, statistical mechanisms and provenance information so that a clue is never turned into a verdict.

Deterministic characters and structures

Invisible characters, bidirectional controls, unusual spaces and metadata can be observed directly. AfterAlgorithm can report them and, for some text, offer a cleaned copy while showing exactly what changes.

Statistical watermarks

A statistical watermark cannot be inferred from a single invisible character. Verification requires a detector suited to the actual scheme and may require specific parameters or keys. Without a publicly verifiable method, AfterAlgorithm labels it unverified.

Provenance and C2PA

Detecting a C2PA structure only establishes that provenance data is present. Cryptographic validation is a separate check. C2PA alone is not proof that content was generated by AI.

Claude case, as of 24 August 2026

Anthropic says that some Claude models launched on or after 2 August 2026 support text watermarks and, where the format supports it, signed C2PA provenance metadata. Anthropic also says technical detection details will be published later. AfterAlgorithm therefore does not label a Unicode anomaly as a verified “Claude watermark”.

Official Anthropic documentation ↗

Understand provenance, C2PA and metadata

2. Image analysis

An image can carry several layers of information outside its visible pixels: EXIF, XMP, IPTC, ICC profiles and sometimes Content Credentials based on the C2PA standard. C2PA is not a universal AI detector. It is a provenance mechanism that can record signed information about a file’s origin and transformations when a tool chooses to provide it.

Think of it as a digital provenance label. It can help explain the declared history of a media asset, but it can be absent, lost during conversion or removed. A detected C2PA structure is only treated as cryptographically verified when a compatible validator actually confirms its signature and trust chain.

Examples of AI image generators and tools

Support for C2PA, Content Credentials, SynthID or other markers varies by product, model, format and version. The names below therefore do not imply that every output carries C2PA.

OpenAI
ChatGPT ImagesGPT Image APIDALL·E
Adobe
FireflyContent Credentials
Midjourney
Midjourney Webimage generation
Google
Gemini imageImagenGoogle AI Studio
Microsoft
CopilotDesignerMicrosoft Foundry / image models
Stability AI
Stable ImageStable Diffusion
Amazon
Nova CanvasAmazon Bedrock
Meta
Meta AIMuse Image
Other ecosystems
Leonardo AIIdeogramRecraftRunwayLuma AIKling AIFLUXxAI / Grok

PDF · Office · ZIP

3. Batch processing

After Algorithm can inspect several JPEG or PNG files in one operation. Each file keeps its individual audit while the overview highlights images carrying C2PA structures or metadata. When batch cleanup is requested, files are processed sequentially and packaged into a ZIP archive with a traceability report.

After Algorithm removes targeted segments without re-encoding the image stream when the format and operation allow it. When the report says “image stream preserved”, the before/after technical comparison confirmed that the image stream was not recompressed. Otherwise, the tool makes no such claim.

After Algorithm can flag active structures and PDF risk indicators. This check is not antivirus software and does not guarantee that a malicious PDF becomes safe.

Filigrane Lab

Watermark resistance-test method

The Lab processes document batches locally. Automatic mode proposes a heuristic mask from distributed colour patterns and estimates an orientation when sufficient signal exists; users can correct the mask before adaptive directional interpolation. It can also create straight, diagonal or oscillating watermarks. Visual disappearance does not prove recovery of the original or absence of traces.

Multi-file · Auto / Assisted / Custom · estimated automatic removal · multiple exports · local processing

2.4.2 release: multipage PDFs are supported when every page is a full-page JPEG image. Vector or mixed PDFs are explicitly rejected rather than partially analysed.

Test a watermark

Analysis without application retention

4. Our commitment

Text and images are processed only to produce the requested analysis or copy. They are not written to an application database, are not used to train AI and are not sent to a third-party AI service in this version. Temporary uploads are deleted at the end of application processing. Text history remains optional and local to the browser.

Local-first, with explicit limits

After Algorithm does not send inspected content to the server. Local history is optional. No clean result is turned into proof of absence.

Keep in mind

After Algorithm observes technical signals. A signal is not attribution, and no signal is not proof of human origin. Secret, proprietary or key-dependent watermarks can remain unverifiable until a compatible detection mechanism is available.

Main technical references