The AI Asia Pacific Institute (AIAPI) has hosted a series of conversations with leading artificial intelligence (AI) experts to study ChatGPT and its risks, looking to arrive at tangible recommendations for regulators and policymakers. These experts include Dr. Toby Walsh, Dr. Stuart Russell, Dr. Pedro Domingos, and Dr. Luciano Floridi, as well as our internal advisory board and research affiliates. This briefing note outlines some of the critical risks of generative AI and highlights potential concerns.
Generative Artificial Intelligence (AI) systems have significantly advanced in recent years, enabling machines to generate highly realistic content such as text, images, and audio. While these advancements offer numerous benefits, it is critical that we are aware of the associated risks.
The AIAPI refers to a letter that has been recently released by the Future of Life Institute requesting for a Pause on AI Experiments. The call requests all AI labs to immediately pause for at least six months the training of AI systems more potent than GPT-4. The letter has received over 28 thousand signatures so far, including Elon Musk, Dr. Andrew Yang, and Dr. Stuart Russell, among others. We recognise that while safeguards are required, this technology’s current application in important sectors can lead to exponential benefits. Therefore, while a pause might be an unrealistic response to the current challenges, we believe safeguards must be put in place urgently. These safeguards are to be led by the public sector.
While this briefing paper provides an overview of the risks relating to language models, a forthcoming policy briefing paper from the AIAPI will expand on these issues, offering potential recommendations to regulators and policymakers.
An Overview of Large Language Models and Other Generative Models
Generative AI refers to a category of AI algorithms that generate new outputs based on the data they have been trained on. ChatGPT is trained with a blend of reinforcement learning algorithms and human input on over 150 billion parameters.
As summarised by Luciano Floridi:
today, artificial intelligence (AI) manages the properties of electromagnetism to process texts with extraordinary success and often with outcomes that are indistinguishable from those that human beings could produce. These AI systems are the so-called large language models (LLMs), and they are rightly causing a sensation.
Outlined below are some of the key risks and areas of concern related to the technology.
Impersonation and Disinformation
Impersonation encompasses a range of actions or techniques used to imitate or pretend to be someone else. It can occur in various forms, such as in-person interactions, social media accounts, or digital media manipulation like deepfakes or voice-mimicking. Impersonation can involve both intentional and unintentional acts and although misinformation and deepfakes are related, is not solely dependent on deepfake technology.
As technology advances, it’s increasingly difficult to distinguish between actual humans and technology systems. One of the most significant examples is the dating company Ashley Madison, which disclosed the personal information of more than 30 million customers as a consequence of a cybersecurity breach. It was subsequently found that customers were in fact unknowingly interacting with AI “bots”. The company was sued for deploying AI “bots” and falsifying user profiles to induce users to make purchases.
Another notable incident in 2022 included a deepfake video of Ukrainian President Volodymyr Zelenskyy surrendering and U.S. prisons using call-monitoring technology on their inmates.
Privacy and Security
Generative AI can inadvertently compromise individual privacy and security. For instance, AI algorithms trained on personal data may generate highly realistic and identifiable information, leading to potential breaches. This was the case involving a South Korean company that trained their language model on existing dating app data.
One of the most significant challenges is the need for more transparency in what data the models are trained. There is an argument that while large language models can aggregate more information, by virtue, they can also help build privacy models.
Bias and Discrimination
Generative AI models learn from the data they are trained on, which can introduce biases present in the training data. These biases can manifest in the generated content, perpetuating existing social biases and discrimination. For example, if a generative AI model is trained on biased texts, it may generate biased or offensive language. New data shows that large language models are also more capable of reflecting biases from their training data. The Artificial Intelligence Index Report 2022 states that a 280 billion parameter model developed in 2021 shows a 29% increase in elicited toxicity over a 117 million parameter model considered the state of the art as of 2018.
This issue requires careful consideration to ensure fairness, inclusivity, and the prevention of further social harm.
Intellectual Property Infringement
Generative AI has the potential to generate content that infringes upon intellectual property rights. The technology can replicate copyrighted works, including text, images, and music, leading to copyright violations and economic loss for content creators. We are already witnessing the start of lawsuits relating to potential infringements and plagiarism examples.
Protecting intellectual property rights in the era of generative AI poses significant challenges as identifying and addressing instances of infringement becomes more difficult.
Conclusion
It is vital to recognize and address the associated risks with generative AI. The majority of experts believe we are at the inflection point where control of such AI systems is still possible. Whether through upcoming regulation or not, control should be at the forefront of this technology so that it goes in the right direction. As these systems become more prevalent, it is crucial to develop robust safeguards to mitigate the risks of impersonation, privacy and data breaches, bias, intellectual property infringement, and other potential unintended consequences. A proactive approach that balances innovation with responsible deployment of generative AI will be critical in harnessing its benefits while safeguarding against potential harm. It is equally important to recognise this to be an ongoing work, where more questions will emerge as technology evolves, interacts and learns.