Study Shows Removing Sensitive Data from AI Models Like ChatGPT Remains a Challenge

Researchers find it's difficult to fully remove sensitive data from large language models like ChatGPT and Bard.
Deleting info from LLMs is hard because models are trained on massive datasets and data is embedded in undefinable weights.
Guardrails like reinforcement learning can limit unwanted outputs but don't delete data.
New study shows factual info can still be extracted from LLMs even after editing methods try to delete it.
Defending against attacks to extract sensitive info is an ongoing challenge as new attack methods emerge.

cointelegraph.com

Relevant topic timeline:

8/30/2023

A.I.’s un-learning problem: Researchers say it’s virtually impossible to make an A.I. model ‘forget’ the things it learns from private user data

A.I. models pose a challenge to data privacy as it is difficult to remove user data from these models without resetting or deleting the entire model, presenting a collision course with inadequate privacy regulations, according to experts.

9/13/2023

Tech Companies Update Privacy Policies to Allow Broader Use of User Data for AI

Companies such as Rev, Instacart, and others are updating their privacy policies to allow the collection of user data for training AI models like speech-to-text and generative AI tools.

9/18/2023

Microsoft Exposes 38TB of Sensitive Employee Data on Misconfigured Azure Server

Microsoft AI researchers accidentally exposed tens of terabytes of sensitive data, including private keys and passwords, while publishing open-source training data on GitHub, but no customer data was exposed as a result.

9/23/2023

Concerns Grow Over Unchecked A.I. Voice Cloning and Data Use

As AI technology progresses, creators are concerned about the potential misuse and exploitation of their work, leading to a loss of trust and a polluted digital public space filled with untrustworthy content.

10/19/2023

AI Training Data Lacks Transparency, Raising Concerns Over Privacy and Bias

Generative AI systems, trained on copyrighted material scraped from the internet, are facing lawsuits from artists and writers concerned about copyright infringement and privacy violations. The lack of transparency regarding data sources also raises concerns about data bias in AI models. Protecting data from AI is challenging, with limited tools available, and removing copyrighted or sensitive information from AI models would require costly retraining. Companies currently have little incentive to address these issues due to the absence of AI policies or legal rulings.

10/19/2023

AI Transparency Falters as New Models Stay Secretive

A recent study by Stanford University reveals that major AI language models, such as GPT-4, are shrouded in secrecy, posing concerns about accountability, reliability, and safety within the field of AI. The study found that no model achieved more than 54% on their transparency scale, urging for greater openness and reproducibility in AI research.

10/25/2023

Study Finds Ethical and Legal Issues in Many AI Data Sets

The Data Provenance Initiative has found that approximately 70% of fine-tuning data sets used by AI developers have improper licensing or are mislabeled, leading to a lack of clarity on copyright restrictions and usage requirements. This has raised concerns about the fair use of text taken from the internet, particularly for training large AI systems. The initiative aims to increase transparency and provide visibility into the ecosystem of data used in generative AI models.

10/26/2023

Meta's AI Data Deletion Form Called Ineffective by Artists

Meta's data deletion request form, which some artists and journalists interpreted as an opt-out program, has proved to be ineffective and frustrating for those attempting to delete their personal data from Meta's generative AI models.

Topics

Posted 10/2/2023, 5:33:02 PM

Study Shows Removing Sensitive Data from AI Models Like ChatGPT Remains a Challenge

Researchers find it's difficult to fully remove sensitive data from large language models like ChatGPT and Bard.
Deleting info from LLMs is hard because models are trained on massive datasets and data is embedded in undefinable weights.
Guardrails like reinforcement learning can limit unwanted outputs but don't delete data.
New study shows factual info can still be extracted from LLMs even after editing methods try to delete it.
Defending against attacks to extract sensitive info is an ongoing challenge as new attack methods emerge.

cointelegraph.com

Relevant topic timeline:

8/30/2023

A.I.’s un-learning problem: Researchers say it’s virtually impossible to make an A.I. model ‘forget’ the things it learns from private user data

9/13/2023

Tech Companies Update Privacy Policies to Allow Broader Use of User Data for AI

Companies such as Rev, Instacart, and others are updating their privacy policies to allow the collection of user data for training AI models like speech-to-text and generative AI tools.

9/18/2023

Microsoft Exposes 38TB of Sensitive Employee Data on Misconfigured Azure Server

9/23/2023

Concerns Grow Over Unchecked A.I. Voice Cloning and Data Use

10/19/2023

AI Training Data Lacks Transparency, Raising Concerns Over Privacy and Bias

10/19/2023

AI Transparency Falters as New Models Stay Secretive

10/25/2023

Study Finds Ethical and Legal Issues in Many AI Data Sets

10/26/2023

Meta's AI Data Deletion Form Called Ineffective by Artists