What if organizational knowledge becomes obsolete, can LLM unlearn it?

Organizational knowledge metamorphoses or it may simply become obsolete as product and service offerings evolve. New products may get introduced and some products/ services may be decommissioned. What will happen to the knowledge associated with dismissed products/ services? How can Large Language models (LLMs) accommodate for these changes? Can LLMs unlearn the expired knowledge? What will be the process of removing unwanted knowledge? To extend the question to the mainstream audience, what if someone asked LLMs to forget his/ her personal information from public GPT technologies, for example, ChatGPT, Claude, Gemini…

Rights to be forgotten or the right to erasure.

The European policy GDPR (General Data Protection Regulation) grants the right to Europeans to request the removal of their personal information from the internet. It is known as the right to be forgotten or the right to erasure.

The right to be forgotten means that people can tell online services, like social media, search engines, or websites, to delete their personal information if they don't need it anymore if they change their mind, or if they use it in the wrong way. But this right is not always true and must be fair with other important rights, such as freedom of expression and information.

The legal case between the New York Times and OpenAI.

In a complementary situation, The New York Times claims that OpenAI copied millions of its news stories, opinions, reviews, and other content without permission or payment, and used them to train its artificial intelligence products, such as ChatGPT. The New York Times argues that this harms its journalism and its ability to provide independent and reliable information to the public. OpenAI denies the allegations and says that it supports journalism and has partnered with other media organizations. OpenAI also says that its AI tools are not substitutes for the original sources, but rather enhance the creativity and diversity of human expression. The NYT asked OpenAI to remove its content from ChatGPT and OpenAI models are requested to unlearn NYT-specific content.

The unlearning challenges.

Should OpenAI retrain its ChatGPT from scratch? If this is the case, it will cost hundreds of millions of dollars to retrain the model. So can LLMs unlearn undesired knowledge in a novel and efficient way? And comply with institutional requests and/ or human rights to be forgotten?

What is unlearning and what are the techniques?

Unlearning is a technique that aims to make LLMs forget undesirable or harmful behaviors, such as generating offensive or inaccurate responses, by modifying their parameters or training data. Unlearning can help align LLMs with human preferences and values and also protect the privacy and intellectual property rights of the data owners.

There are different methods of unlearning, such as fine-tuning, pruning, distillation, and data removal. Each method has its advantages and disadvantages in terms of effectiveness, efficiency, and scalability. Some recent research papers have explored how to perform unlearning on LLMs and evaluated their results on various tasks and datasets.

Who’s Harry Potter?

Microsoft researchers have recently published a study addressing the challenge of unlearning specific subsets of data from large language models (LLMs) without retraining them from scratch, due to ethical, legal, and technological concerns. The researchers applied their technique to unlearn the Harry Potter books from the Llama2-7b model, a generative language model open-sourced by Meta.

The study proposed a novel technique that aims to identify idiosyncratic terms and replace them with target labels. For example, replace ‘Harry’ with ‘Jack’ and ‘Hermione’ with ‘her’. They overwrite the base model with a refined model while targeting the unique and specific n-grams and expressions of the Harry Potter saga with generic labels. However, the challenge is more complicated with non-fictional texts where idiosyncratic expressions are less dense and unique. Journals, magazines, blogs, and articles are often constructed with higher levels of abstraction, ideas, and sentences. It’s uncertain to which extent this replacement technique can address the unlearning LLMs.

Conclusion and insights

LLMs can unlearn, but it is not a trivial or easy process. It’s a common belief that ‘set it and forget it’ is not the wise approach to deploy a Generative AI application, for example, RAG (Retrieval Augmented Generation) Question-Answering System.

Unlearning/ relearning in the organizational context requires robust governance, a clear conversational model, knowledge architecture, and active involvement of knowledge curators and brokers. It’s ultimately the role of knowledge managers to create adaptive spaces and communities where knowledge can be continuously co-constructed, consolidated, and ready to be ingested into the Large Language Models. ?

For more information and insights on AI Integration Strategy in the context of Learning and Knowledge Management, I may recommend watching these two 3R? videos (video1, video2) or engaging with us at [email protected]