OpenAI engineers are developing a tool to apply “watermarks” to content generated by artificial intelligence systems, according to Scott Aaronson, a computer science professor invited by the company, who spoke at a lecture at the University of Texas at Austin.
According to the researcher, Hendrik Kirschner has already created a working prototype, which they hope to integrate into future products.
“We want to make it harder for AI-generated results to be mistaken for human work,” said Aaronson.
He added that it would help prevent academic plagiarism and the mass spread of propaganda.
Systems like the chatbot ChatGPT understand input and output text as strings of “tokens”, which can be words, parts of words, or punctuation marks.
AI algorithms continually generate a mathematical function called a probability distribution to determine the next emitted lexeme based on previously provided information.
In models such as ChatGPT, after the distribution is created, the OpenAI server performs token sampling according to it. There is some randomness in this process, so the same text prompt may lead to a different answer.
According to Aaronson, the tool being developed by the company acts as a “wrapper” around existing text generators. It uses a cryptographic function operating at the server level for a “pseudo-random” selection of the next “token”.
When such a system is engaged, ordinary users will see the AI-generated text without modification. However, those who have a “key” to the cryptographic function will be able to detect the “watermark”.
Independent scientists and industry experts expressed doubts about the tool’s correctness. They say that since the system is server-based, it will not be able to work with all AI text generators. Opponents of the method will also be able to circumvent it rather easily, the researchers say.
Allen Institute for AI researcher Jack Hessel pointed to difficulties in unobtrusively removing fingerprints from algorithm outputs, since each “token” is a discrete choice.
Too obvious an identifier would lead to the selection of odd words that could degrade fluency of speech.
Co-founder of AI21 Labs Yoav Shoham urged a more sophisticated approach to determining the sources of text, which would include differential rather than static “watermarks”. They would allow marking different parts of the text differently.
In his lecture, Aaronson acknowledged that OpenAI’s proposed scheme would operate only in a world where all companies working in this field agree to become responsible players.
However, according to the scientist, if the lab demonstrates the tool’s viability and that it does not affect the quality of generated text, it could become an industry standard.
In December, OpenAI released the chatbot ChatGPT, which can answer questions, admit mistakes, debate and reject inappropriate requests.
In September the company unveiled the Whisper speech recognition system, enabling transcription in multiple languages.
In January OpenAI released a new version of GPT-3, which produces fewer offensive expressions, misinformation and errors overall.
Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!