Marca d'água em textos gerados por IA
Introdução A Anthropic anunciou recentemente a inclusão de uma marca d'água nos textos gerados pelos modelos Claude. O objetivo é distinguir conteúdos gerados por humanos daqueles criados por IA generativas. Funcionamento O prompt enviado é convertido em tokens que são usados para calcular a probabilidade do próximo token , se repetindo até uma resposta que faça sentido seja retornada para o…
Anthropic, a leading artificial intelligence company, recently introduced a watermark system for text generated by its Claude models. This move aims to differentiate content created by humans from that produced by generative AI. The system works by converting the input prompt into tokens, which are then used to calculate the probability of the next token.
The process continues until a coherent response is generated for the user. The key difference in Anthropic's approach lies in its deterministic selection of the next token from a list of probable candidates, which uses a private key and existing context to generate the text. According to Anthropic, this watermarking process does not consume additional tokens or slow down the models' response time.
The adoption of such watermarking systems is driven by the EU's 2024 Artificial Intelligence Act, which mandates that AI systems generating synthetic content must mark their outputs in a machine-readable format, signaling that they are artificially generated or manipulated. While some major tech companies have signed an adhesion pact and are implementing the technical aspects according to their timelines, practical implementation by proprietary LLM model owners remains limited.
Reasons for this slow adoption include concerns about traceability, potential client loss, and technical fragility. Despite the lack of evidence that a watermark can uniquely trace a text to its creator, the system only allows determining whether content was generated by AI or if it has been altered using an LLM. The consequences of such watermarks extend beyond privacy, impacting a company's revenue and even leading to bankruptcy in extreme cases.
The integration of these watermarking systems into AI-generated content also raises concerns about technical fragility, including a potential decrease in creativity, difficulty in calculating the next token with limited content, and attacks that exploit vulnerabilities by rewriting parts of the generated text to invalidate the watermark.
As AI-generated content becomes increasingly prevalent, identifying and valuing human-created work becomes a critical challenge. The main question behind these regulatory actions is how to distinguish and value human-created work from synthetic content. Ultimately, the fear is that consumers will struggle to discern reality from fiction.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
