Let us please cease spreading conspiracy theories regarding legitimate usage of Artificial Intelligence. There are some real concerns about AI but the current policing and suspicion by the general public borders on tactics used during the Inquisition.
The recent Thélyson Orélien novel controversy is just the latest example. So what if the author and publisher used AI-assisted tools for editing, text or image generation or more. AI is but a tool — and a legal tool at that. The author is regardless personally responsible and liable for published content — AI-assisted, or not. AI is here to stay. It is everywhere and it is unavoidable. And do not misinterpret AI-detection as being a watch dog for all AI-use:
Anthropic is releasing a Detection API in private preview to identify text generated by its Claude models.
How Anthropic’s Detection Works
- Invisible Watermarking: Anthropic applies an invisible, statistical watermark to text generated by Claude. It subtly influences token and word selection without altering the overall meaning or writing style for human readers.
- Who Can Access It: The Detection API is not a public, open-to-all consumer web tool. It is currently available in private preview to eligible organizations with compliance obligations under EU law—such as regulators, law enforcement agencies, educators, fact-checkers, and enterprise clients.
- Content Checker Tool: Anthropic also launched a multimedia verification tool (
[claude.com/check-content](https://claude.com/check-content)) to check if images, audio, or video files were created or edited using Claude via C2PA standards, though text formats are handled separately through the Detection API. - Limitations: The watermark can break or become undetectable if the text snippet is very short or if the majority of the content has been heavily rewritten by a human or another program.
Note on AI Usage: The general public currently does not have access to this detection tool. Furthermore, the legal status surrounding the use of AI tools remains evolving. Using AI for legitimate, non-criminal purposes is neither illegal nor unethical—when in doubt, simply include a brief disclaimer acknowledging that AI was used in creating the content.
And both Google (Gemini) and OpenAI (ChatGPT) are also actively using and expanding similar text watermarking and detection technologies, primarily driven by regulatory requirements like Article 50 of the EU AI Act. Hello Growth
Like Anthropic, both companies rely on statistical token watermarking—meaning they do not insert hidden characters or secret tags, but rather subtly steer the model’s word choices during generation to leave an invisible mathematical pattern. Medium
How Google (Gemini) Approaches Detection
- Technology: Google uses its SynthID-Text system (developed by Google DeepMind). It steers word selection behind the scenes without changing the quality or readability of the output.Medium+ 1
- Current Status: Google integrated SynthID directly into Gemini for text, image, video, and audio generation. Google DeepMind
- Access & Detection: For multimedia (images, audio, video), users can verify if content was made by Google AI through Gemini, Search, or Chrome. For text, access to detection API tools remains focused on compliance, enterprise, and research applications rather than open-to-all consumer text checkers. Google DeepMind+ 1
How OpenAI (ChatGPT) Approaches Detection
- Technology: OpenAI uses a system called textGrain. Similar to SynthID and Anthropic’s methods, it subtly influences the model’s probability scoring when picking words. Bitdefender+ 1
- Current Status: OpenAI started rolling out
textGrainacross ChatGPT and Codex text in response to regulatory mandates, particularly for users in the EU. Free AI Humanizer - Access & Limitations: OpenAI’s text detection tool is not widely open to the general public or schools to grade student work; access is restricted to approved research, enterprise, and regulatory organizations. Developers using the OpenAI API have control over enabling or disabling text provenance for their projects. Free AI Humanizer+ 2
Key Takeaways Across All Three (Anthropic, Google, OpenAI)
- Shared Detection Limits: Statistical text watermarking works well on long, unedited texts. However, across all models, if the text is very short (e.g., under 150–200 words) or heavily edited/paraphrased by a human or another model, the watermark signal decays or disappears. Hello Growth
- No Public “Catch-All” Web Tool: None of the major AI labs offer a free, open web-based text detector for the general public, as false positives and false negatives remain too common for high-stakes consumer enforcement. Hello Growth
- No Personal Data Stored: The watermarks detect whether an AI model produced the pattern, not who prompted it or generated it.
Is this akin to police entrapment? No, it is not legally or conceptually akin to police entrapment. The term “entrapment” has a very specific legal definition, and AI watermarking operates on fundamentally different legal,technical, and ethical principles. What Entrapment Actually Means: Entrapment occurs when a law enforcement officer (or state agent) induces, tricks, or coerces a person into committing a crime that they otherwise would not have committed, for the purpose of prosecuting them.For an action to be entrapment, three things must happen: Government action: The scheme must be initiated by law enforcement or government authorities. Inducement: Law enforcement must actively persuade or pressure someone to break the law. And lack of predisposition: The person had no prior intention of committing the crime until the officers coerced them.Why AI Detection/Watermarking is Not Entrapment1. Private Services, Not Police Sting OperationsAI companies (Anthropic, Google, OpenAI) are commercial entities providing a service. They are not secret police officers luring you into breaking a rule. Using AI is Generally Legal: Generating text, essays, code, or emails with Claude or ChatGPT is completely legal for standard, non-criminal tasks. You are using a paid or free consumer product as intended. Because generating content is not inherently illegal, offering the service cannot be considered “entrapment.” Watermarking is a Provenance Label, Not a Trap: Watermarking acts like a digital signature or manufacturer’s stamp.- When a camera manufacturer bakes EXIF metadata into a photograph identifying which camera took it, it isn’t “trapping” the photographer.Similarly, when an AI model subtly guides word selection so that the output carries a statistical signature, it is simply tracking the origin (provenance) of the media.
- The Trap Feeling: A student or employee uses a public AI tool to draft a paper, assuming it’s private. Later, an institution runs a detection tool and flags them for academic dishonesty or policy violations.The Reality: The “trap” isn’t created by the AI company; it’s a conflict between your institution’s rules (which prohibit unauthorized AI use) and the technology’s transparency features (which prove the content was generated by AI).
Article 50 of the European Union Artificial Intelligence Act (EU AI Act) introduces strict transparency obligationsdesigned to fight deception, deepfakes, and illegal misrepresentation by making AI-generated content clearly identifiable.
ai-act-service-desk.ec.europa.eu
Rather than banning AI outright, Article 50 enforces a clear rule: people have a right to know when they are interacting with an AI system or viewing synthetic media.
EU AI Act
The Four Core Rules of Article 50SSL.com
Article 50 sets out four main obligations split between providers (companies building AI models, like OpenAI or Anthropic) and deployers (businesses and organizations using AI):
EU AI Act
ARTICLE 50 OBLIGATIONS
│
┌───────────────────┬───────┴───────────┬───────────────────┐
▼ ▼ ▼ ▼
1. Chatbots 2. Technical 3. Biometrics & 4. Deepfakes &
& Agents Provenance Emotion AI Public Text
(Disclosure) (Watermarking) (Prior Notice) (Visible Labels)
1. Disclosing Direct Human-AI Interaction (Art. 50(1))
- Who it applies to: AI systems designed to interact directly with people (e.g., customer service chatbots, virtual assistants, or AI agents).Shaping Europe’s digital future – European Union
- The Rule: Providers must design systems so users are clearly informed upfront that they are speaking to or chatting with an AI.Shaping Europe’s digital future – European Union
- Exceptions: Notice is not required if it is glaringly obvious to a “reasonably well-informed person” that an AI is involved, or for specific authorized law enforcement operations.EU AI Act
2. Machine-Readable Watermarking (Art. 50(2))Shaping Europe’s digital future – European Union
- Who it applies to: Providers of generative AI systems that output synthetic text, images, video, or audio.Shaping Europe’s digital future – European Union+ 1
- The Rule: Generative AI outputs must be marked in a machine-readable format that makes them detectable as AI-generated.Shaping Europe’s digital future – European Union
- How it works: Companies must embed reliable, tamper-resistant metadata, digital signatures (such as C2PA standards), or invisible statistical watermarks (such as Google’s SynthID or Anthropic’s text watermarking).SSL.com
- Exceptions: Basic assistive editing tools (like spell-check or grammar correctors that do not substantially alter original meaning) are exempt.EU AI Act
3. Emotion Recognition & Biometric Categorization (Art. 50(3))EU AI Act
- Who it applies to: Organizations deploying AI that detects human emotion or categorizes people biometrically (e.g., mood detection or demographic sorting).EU AI Act
- The Rule: Deployers must notify individuals before or at the moment they are exposed to the AI system.Shaping Europe’s digital future – European Union+ 1
- Context: While using emotion AI in workplaces and schools is heavily restricted or banned under other sections of the AI Act (Article 5), where it is allowed (such as in certain customer service or safety research settings), transparent notice is mandatory.EU AI Act
4. Labeling Deepfakes & Public-Interest Text (Art. 50(4))SSL.com
- Who it applies to: Deployers publishing AI-generated content for public consumption.SSL.com
- The Rule:
- Deepfakes: Audio, visual, or video media that creates a realistic illusion of real people, places, or events must carry a clear, visible label disclosing that it was artificially generated or manipulated.ai-act-service-desk.ec.europa.eu
- Public-Interest Text: AI-generated articles or news text published to inform the public on matters of public interest (e.g., politics, public health, law) must also be explicitly labeled.SSL.com
- Artistic Exception: For satirical, artistic, or fictional works, the disclosure must be done in a way that informs the audience without ruining the artistic experience or display.EU AI Act
Why Article 50 Matters
| Aspect | Impact |
|---|---|
| Global Reach | Even if an AI company is based in the US (like OpenAI, Google, or Anthropic), if their system’s output is used or accessible in the EU, they must comply. |
| Combating Misinformation | It forces platforms to adopt standardized provenance tracking (e.g., C2PA metadata, imperceptible watermarks) so fact-checkers and social platforms can verify content origin. |
| Enforcement | Non-compliance with transparency rules can trigger significant fines under the EU AI Act enforcement framework. |
In short, Article 50 is the legal driver behind why AI companies are currently rolling out invisible text watermarks, disclosure badges, and verification tools (like Anthropic’s Detection API or Google’s SynthID) across the tech industry. EU AI Act
— Gemini and Adam Donaldson Powell
Leave a Reply