The Definition Problem

When someone says "GPT watermark," they might mean any of three fundamentally different things. In the broader technology space, "watermark" refers to any signal embedded in content to indicate its origin or authenticity. In the context of AI text, this term has been applied loosely to cover both intentional signals and accidental artifacts.

The confusion matters because the three types have different properties: different detectability, different removability, and different implications for privacy and policy. Treating them as interchangeable leads to inaccurate advice and misplaced concern.

Let's go through each type in detail.

Type 1: Cryptographic Watermarks (Proposed, Not Deployed)

A true cryptographic watermark for AI text would work as follows: during the token generation process, the model uses a secret key to bias its token selection. Instead of sampling purely from the probability distribution over tokens, it systematically prefers tokens that belong to a particular set defined by the key. The resulting text reads normally — the bias is imperceptible to a human reader — but the pattern of token choices creates a statistical signal that can be verified by anyone who knows the key.

This approach was described in detail by researchers at the University of Maryland in a widely-cited 2023 paper. OpenAI researchers have referenced similar work internally. The key property of this approach is that it produces a watermark that is nearly impossible to remove without significantly degrading the text, because removing the watermark requires knowing which tokens were biased and substituting alternatives systematically.

The practical implication: you cannot currently detect a cryptographic watermark in ChatGPT text because no cryptographic watermark exists in current ChatGPT output. Any tool claiming to detect "OpenAI's cryptographic watermark" is making a false claim.

Type 2: Statistical Watermarks (Naturally Present)

Statistical watermarks are not deliberately embedded — they are natural properties of AI-generated text. Language models produce text with characteristic statistical signatures: low perplexity (predictable word choices), low burstiness (uniform sentence length), consistent argument structure, and specific vocabulary preferences.

These patterns emerge from how language models work: they optimize for coherent, grammatically correct, high-probability text. Human writing has more statistical entropy because thought and expression are naturally more variable. The resulting signature is "watermark-like" in that it identifies AI origin, but it is not deliberate.

Statistical watermarks are detectable without any key — just probabilistically, not definitively. This is why AI detector scores are expressed as probabilities ("83% likely AI") rather than certainties. They are also reducible by editing: adding variety, changing vocabulary, and restructuring paragraphs all reduce the statistical AI signature.

Type 3: Unicode Artifact Watermarks (Accidentally Present)

The third type is what most people encounter in practice when they talk about "invisible watermarks." AI models sometimes produce text with invisible Unicode characters embedded within it — zero-width spaces, zero-width joiners, soft hyphens, byte-order marks, and directional formatting characters.

These are not deliberate watermarks. They are artifacts of the generation process — characters that appear in training data and are reproduced by the model at similar positions in its output. They are consistently more common in AI-generated text than in human-typed text, which makes them useful as secondary detection signals.

Unlike statistical patterns, Unicode artifacts are binary: either the character is present or it is not. They can be removed completely with the right tools, without affecting the visible content of the text in any way. The ChatGPT Watermark Detector and Invisible Character Detector scan specifically for these characters.

Why Companies Want to Embed Watermarks

The push for genuine AI watermarking comes from several directions. Governments, academic institutions, and media organizations are all interested in reliable provenance tracking for AI-generated content. The uses range from preventing academic fraud to limiting deepfake misuse to enabling copyright attribution.

How to Detect What Is Currently Detectable

Given the current state of AI watermarking — no deployed cryptographic watermarks, natural statistical patterns, and accidental Unicode artifacts — detection tools focus on the latter two. Here is what you can reliably detect with available tools: