AI watermarking doesn't visibly alter text. Instead, it uses a secret key to slightly boost the probability of certain words appearing in a sequence. This creates a statistically significant pattern that can only be detected by a tool with access to the original model and the secret key.

The necessity of rewriting AI content to avoid watermarks isn't a bug; it's a feature. It compels marketers to adopt the best practice of using AI for ideation and organization while retaining a human-driven final edit, which should have been the standard workflow all along.

Claude's watermark is a subtle pattern of word choices, not hidden characters. To bypass it, provide your own draft and instruct the AI to only fix grammar and punctuation, explicitly telling it not to rewrite the content. Anthropic confirms this prevents a detectable watermark.

Anthropic's move to embed invisible watermarks directly into all AI-generated text to comply with EU regulations has ignited controversy. Developers and users fear this will constrain the model's creativity, degrade output quality for tasks like coding, and set a worrying precedent for content integrity.

A new form of analysis compares the semantic similarities (e.g., diction, phrasing) of outputs from different AI models. This technique is being used to create 'fingerprints' that can suggest if one model was illicitly 'distilled' or trained on the outputs of another, a key concern in the AI arms race.

Heuristics for spotting AI writing, like the overuse of em dashes, are becoming obsolete as models learn from human feedback. For instance, ChatGPT now uses em dashes *less* frequently than human writers at The Economist, flipping the old tell on its head and complicating detection efforts.

Initiatives like Google's Synth ID aim to standardize detection of AI-generated content. However, these systems are vulnerable. Simple user actions like screenshotting can strip metadata, and blending AI-generated assets with real footage can easily confuse detection algorithms, limiting their effectiveness.

AI labs face a trade-off with watermark detection tools. Making them widely available promotes public transparency, but it also allows bad actors to use the detector's feedback to reverse-engineer and train other AI models to become more effective at removing the watermarks, undermining the system's long-term security.

Critics argue that the EU's watermarking rules could unfairly credit AI for work that is merely edited by a model. This could discourage creators from using valuable AI tools, as their entire work might be labeled "AI-generated," diminishing their own contribution and creative ownership.