A text penetration that determines the choice of words in the key and immediate context and examines the involvement of AI from the post-generation word string will be introduced to Claude’s future model.
Introduction of SynthID-Text methods to future models
Claude’s future model will move to a design that generates texts that include text penetration.Anthropic will adopt a method based on the SynthID-Text published by Google DeepMind in its 2024 Nature paper.
This method determines the possibility that the text was partially written by Claude. The verifier with the key examines the word string after generation and shows the probability based on how much it matches with the choice when using the penetration.
Leave a pattern for the choice of words to keep meaning.
A large-scale language model reduces the candidate of the following words according to the previous sentence and chooses one of them.In a scenario where there are multiple candidates with little meaning or naturality, no matter which one you choose, the impact on the content you receive is small.
Text penetration repeats these choices in the text, leaving patterns that can be detected with the key in the word line. Usually, randomness from any variable generator is determined from the key and the words immediately before it, so that the integrity of the choice can be tested after it is generated.
Selection itself remains random and is not a process in which the model usually chooses unnatural words that are not candidates.
Keep quality without adding letters or identification information.
The pattern for detecting comes from the choice of words, so no visible symbols or hidden characters are inserted into the text. Additional tokens are not required, and the penetration does not include identifying information associated with individuals, organizations or individual chats.
Anthropic claims that readers cannot distinguish between permeable and non-permeable responses.In-house tests did not confirm the impact on the content, creativity, and readability of the text.
The length of the text and the choice of language depends on the detection.
The consistency with the key is limited to the possibility that Claude was involved in a part of the text. It is not possible to verify that it was written by humans from the detection results, nor can it be determined that it was generated by a different AI using different keys or methods.
In short texts, less words are selected and less information is used for judgment.The longer the text and the more choices are made, the more reliable the judgment of Claude’s involvement.
In cases where the correct words are limited, such as facts, it is difficult to change choices whileining accuracy, and there is a lack of transparency.Any correction that slightly corrects the grammar and reading points of a person’s written sentence may not have enough patterns to detect involvement if Claude has fewer words to re-select.