Updated OECD AI Principles to keep up with novel and increased risks from general purpose and generative AI

On May 3, 2024, the OECD released updated AI Principles that build upon the 2019 version with some notable differences that respond to risks emerging from latest technological developments such as general purpose and generative AI systems.
This article contains a summary of the changes as well as a line-by-line comparison of the old and new AI Principles.
The most notable changes are:
- Principle 1.1 now explicitly calls out environmental sustainability, acknowledging increased concerns around the environmental footprint of large language models over the past years.
- Principle 1.2 received a new heading which now includes respect for the rule of law, human rights and democratic values, and privacy, while fairness has been there all along. However, all of these elements were previously included in the Principle itself. The Principle also specifically mentions that mis- and disinformation amplified by AI must be addressed while respecting freedom of expression. Some more details are added around the required safeguards with an emphasis on human agency and oversights as well as risks resulting from misuse.
- Principle 1.3 saw only slight changes. One of them is the specification of information requirements enabling an understanding of the AI system, which now explicitly include capabilities and limitations.
- Elements of Principle 1.4 have been moved to Principle 1.5 “Accountability” while override, repair, decommission, and information integrity requirements were added.
- Principle 1.5 receives a large addition covering ongoing systematic risk management and responsible business conduct, calling out specifically harmful bias, human rights including safety, security, and privacy, as well as labour and intellectual property rights. Net new among those is bias, which did not find any mention in the previous version of the Principles
Summary
New additions to the OECD AI Principles that had not been included previously are environmental sustainability, mis- and disinformation, as well as bias.
Increased emphasis is placed most notably on risk management practices highlighting novel risks and mechanisms to address risks such as override, repair, decommissioning, and cooperation between AI actors. Privacy, too, takes a more prominent position than before with now 3 vs. 1 mention in the Principles.
How Private AI can Facilitate Adherence to Privacy Principle and Bias Mitigation
The best way to preserve the privacy of individuals is not to contain personal identifiers in the training or fine-tuning data of LLMs. Where personal data redaction or replacement with synthetic data is an option from a data utility standpoint, technology can now be leveraged to greatly facilitate the task of removing personal information from vast amounts of data, even unstructured data.
Private AI has developed solutions that do exactly that, and with unparalleled accuracy. Our AI-driven technology is trained to detect over 50 entities of personal information, using the context of the text to determine what constitutes a personal identifier. Available as on-prem deployment or via an API key, the tool works across many different file formats and over 50 languages. You can try it here on your own data.
Removing indirect personal identifiers such as origin, race, gender, and physical attribute can also help with the mitigation of bias. During inference, this can be achieved by using common LLMs with the added PrivateGPT interface of Private AI. PrivateGPT filters out any selected personal identifiers, including those to which biases most commonly attach, before they are sent to ChatGPT or other LLMs. Before the response is sent to the user, the original identifiers are automatically included back into the output for a seamless user experience. Try PrivateGPT here for free.
Table 1
The table below displays a colour-coded line-by-line comparison of the old and new principles with green indicating new additions, yellow a re-organization, and red deletions.