Your right to be forgotten in the age of AI

22 September 2023

share this:
person sat on a sofa with a laptop

By Alice Trend, David Zhang and Thierry Rakotoarivelo

Earlier this year, ChatGPT was briefly banned in Italy due to a suspected privacy breach. To help overturn the ban, the chatbot’s parent company, OpenAI, committed to providing a way for citizens to object to the use of their personal data to train artificial intelligence (AI) models.

The right to be forgotten (RTBF) law plays an important role in the online privacy rights of some countries. It gives individuals the right to ask technology companies to delete their personal data. It was established via a landmark case in the European Union (EU) involving search engines in 2014.

But once a citizen objects to the use of their personal data in AI training, what happens next? It turns out, it’s not that simple.

CSIRO cybersecurity researcher Thierry Rakotoarivelo is co-author of a recent paper on machine unlearning. He explains that applying RTBF to large language models (LLMs) like ChatGPT is much harder than search engines.

“If a citizen requests that their personal data be removed from a search engine, relevant web pages can be delisted and removed from search results,” Thierry said.

“For LLMs, it’s more complex, as they don’t have the ability to store specific personal data or documents, and they can’t retrieve or forget specific pieces of information on command.”

So, how do LLMs work?

LLMs generate responses based on patterns they learned from a large dataset during their training process.

“They don’t search the internet or index websites to find answers. Instead, they predict the next word in a response based on the context, patterns and relationships of words provided by the query,” Thierry said.

Another CSIRO cybersecurity researcher David Zhang is the first author of Right to be Forgotten in the Era of Large Language Models: Implications, Challenges, and Solutions. He has a great analogy for how humans use training data they have learned for speech generation as well.

“Just as Australians can predict that after ‘Aussie, Aussie, Aussie’ comes ‘oi, oi, oi’ based on training data from international sports matches, so too do LLMs use their training data to predict what to say next,” David said.

“Their goal is to generate human-like text that is relevant the question and makes sense. In this way, an LLM is more like a text generator than a search engine. Its responses are not retrieved from a searchable database, but rather generated based on its learned knowledge.”

Is this why LLMs hallucinate?

When a LLM outputs incorrect answers to prompts it is said to be “hallucinating“. However, David says hallucination is how LLMs do everything.

“Hallucination is not a bug of Large Language Models, but rather a feature based on their design,” David said.

“They also don’t have access to real-time data or updates post their training cut-off, which can lead to generating outdated or incorrect information.”

How can we make LLMs forget?

Machine unlearning is the current front-runner application to enable LLMs to forget training data, but it’s complex. So complex, in fact, that Google have issued a challenge to researchers worldwide to progress this solution.

One approach to machine unlearning removes exact data points from the model through accelerated retraining of specific parts of the model. This avoids having to retrain the entire model, which is costly and takes time. But first you need to find which parts of the model need to be retrained, and this segmented approach could generate issues with fairness by removing potentially important data points.

Other approaches include approximate methods with ways to verify, erase, and prevent data degradation and adversarial attacks on algorithms. David and his colleagues suggest several band-aid approaches, including model editing to make quick fixes to the model while a better fix is developed or a new model with modified dataset is being trained.

In their paper the researchers use clever prompting to get a model to forget a famous scandal, by reminding it the information is subject to a right to be forgotten request.

The case to remember and learn from mistakes

The data privacy concerns that continue to create issues for LLMs might have been avoided if responsible AI development concepts were embedded throughout the lifecycle of the tool.

Most well-known LLMs on the market are “black boxes”. In other words, their inner workings and how they arrive at outputs or decisions are inaccessible to users. Explainable AI describes models where decision making processes can be traced and understood by humans (the opposite of “black box” AI).

When used well, explainable AI and responsible AI techniques can provide insight into the root cause of any issues in models – because each step is explainable – which helps find and remove issues. By using these and other AI ethics principles in new technology development, we can help assess, investigate and alleviate these problems.

Read the research in full

Learn to Unlearn: A Survey on Machine Unlearning, Youyang Qu, Xin Yuan, Ming Ding, Wei Ni, Thierry Rakotoarivelo, David Smith.


            AIhub is supported by:

Related posts :

Madagascar’s ancient baobab forests are being restored by communities – with a little help from AI

The collaboration between communities and scientists aims to restore baobab forests in Madagascar to their natural state.
24 May 2024, by

DataLike: Interview with Wuraola Oyewusi

Ndane and Isabella talk to Wuraola Oyewusi about challenging and rewarding aspects of research and how her background in pharmacy has helped her data and AI career

European Union AI Act receives final approval

On 21 May, the Council of the EU formally signed off the artificial intelligence Act.
22 May 2024, by

#ICLR2024 invited talk: Priya Donti on why your work matters for climate more than you think

How is AI research related to climate, and how can the AI community better align their work with climate change-related goals?
21 May 2024, by

Congratulations to the #ICRA2024 best paper winners

The winners and finalists in the different categories have been announced.
20 May 2024, by

Trotting robots offer insights into animal gait transitions

A four-legged robot trained with machine learning has learned to avoid falls by spontaneously switching between walking, trotting, and pronking
17 May 2024, by

AIhub is supported by:

©2024 - Association for the Understanding of Artificial Intelligence


©2021 - ROBOTS Association