ChatGPT Memorizes Copyrighted Poems: Ethical Concerns Arise
Cornell University researchers have conducted a study revealing that ChatGPT, a popular language model, has the ability to memorize entire poems, even those protected by copyright laws. The study raises ethical questions regarding how artificial intelligence models such as ChatGPT are trained, potentially using data scraped from the internet.
The researchers found that ChatGPT is particularly adept at memorizing poems, especially well-known ones that are readily available online. This discovery has sparked concerns about privacy, as large language models retaining vast amounts of text can pose a privacy risk, especially when the source of the data used for training is unknown.
Lyra D’Souza, the study’s first author, explained that proprietary models like ChatGPT could be trained on private data, but the public remains unaware of the specifics. This lack of transparency raises further concerns about the usage of personal information and highlights the need for responsible and transparent AI training practices.
The study focused on poems as they are concise enough to fit within the context size of a language model. Although many of the poems examined were technically under copyright, they are widely available from reputable sources like the Poetry Foundation, adding a layer of complexity to the issue.
D’Souza tested the memory capabilities of ChatGPT and three other language models using prompts requesting specific poems. The research revealed that the most reliable indicator of memorization was whether the poem was included in the 1983 edition of the Norton Anthology of Poetry.
As the model evolved over time, ChatGPT’s responses changed. Initially, it would fabricate or recycle poems if it was unfamiliar with a specific one. Later on, the model would question the existence of the poem, effectively shifting the responsibility to the user. Furthermore, while there were initially no copyright limitations on ChatGPT’s ability to produce copyrighted poems, the model started occasionally refusing to reproduce them by July.
While this study only focused on American poets, the next step is to explore how chatbots respond to requests in different languages and investigate whether factors such as poem length, meter, and rhyming patterns influence their likelihood of being memorized.
D’Souza emphasized the need to utilize ChatGPT responsibly and transparently, considering its potential as a powerful tool in our lives going forward.
In conclusion, the study conducted by Cornell University sheds light on the ethical concerns surrounding AI language models, specifically ChatGPT’s ability to memorize copyrighted poems. Given the widespread use of these models, it is crucial to address concerns regarding privacy, data usage, and responsible AI training practices to ensure their responsible and transparent integration into society.
