The inaugural International AI Safety Report provides a comprehensive insight into General-purpose AI’s current state, future potential, and associated risks. General-purpose AI refers to AI models or systems that can perform a wide variety of tasks, as opposed to Specialized AI.
Synthesizing insights from over 90 independent international AI experts, the report summarizes understanding around three core questions: What can general-purpose AI do? What are the risks associated with general-purpose AI? And what mitigation techniques exist against these risks?
A lot has changed since an interim report was published in May 2024. So, let’s review the report and assess the current situation.
The Current State of Play
As the report states, ‘What AI can do is a key contributor to many of the risks it poses, and according to many metrics, general-purpose AI capabilities have been progressing rapidly.’ General-purpose AI can handle all sorts of data, such as text generation, audio/video generation, question answering, and even robotic actions. It is based on Deep Learning Technology, which uses multilayered neural networks.
Examples of general-purpose AI include Language models such as GPT, which is in its fourth iteration. Currently, the process of developing and deploying a general-purpose AI model has six distinct stages:
1. Data Collection and pre-processing: Raw data is collected, cleaned, labelled, standardized, and transformed into a format the model can learn from.
2. Pre-training: The model learns from a vast array of data, gaining a broad understanding of language, concepts, or images. It is essential but compute-intensive.
3. Fine-tuning: Takes a pre-trained machine learning model and further trains it on a smaller, targeted data set to refine the AI system further through feedback and data.
4. System Integration: The AI model is combined with other components such as User Interfaces, filters, and mechanisms.
5. Deployment: When others use the model.
6. Post-deployment monitoring: Feedback is collated and performance tracked to resolve issues and improve performance.
Training a general-purpose AI model can cost millions of dollars, primarily due to the need for powerful GPUs and the vast size of the training datasets. The report predicts that by 2027, the cost to retrain some of the largest models could exceed a billion dollars – and it’s not simply about the algorithms; it’s about having the resources to train them.
The environmental impact of these models can’t be underestimated either. GPT-3, the precursor to GPT-4, consumed about 1300 MWh (megawatt-hours) of power, while GPT-4 requires approximately 1,750 MWh of energy. This usage is equivalent to the yearly energy consumption of about 160 average American homes. The report states that ‘AI energy demand is expected to grow substantially by 2026, with some estimates projecting a doubling or more, driven primarily by general-purpose AI systems such as language models.’
AI is increasingly demonstrating its capability to move beyond passive tasks and into more active roles. A significant advancement since the interim report were the early test results and subsequent release of OpenAI’s new AI model, o3, which has even outperformed some human experts and a breakthrough in an abstract reasoning test thought to be previously unattainable.
A Risky Business
AI is undoubtedly a high-stakes race, one that organizations and even governments intend to be at the forefront of. The report predicts that by the end of 2026, some AI models could be using 100x more compute than the most advanced models from 2023, potentially 10,000 times more by 2030. But whether we will have the resources, data chips and energy to keep up with such growth is another matter.
There are broadly three main types of risk identified in the report: systemic risks, malicious use, and malfunctions.
Systemic risks
This encompasses broader societal risks associated with AI growth and deployment, including the previously touched-upon environmental cost. Current estimates identify data centers and data transmissions as accounting for an estimated 1% of global energy-related GHG (greenhouse gas) emissions, with that figure predicted to at least double by 2026.
Economic disruption and the accelerating trend of AI-automated jobs are also important systemic risk factors to consider. AI systems can reflect and even amplify existing biases in data, which can play out in areas like hiring, lending, and even criminal justice.
Malicious use
There is undoubtedly huge potential for the malicious use of general-purpose AI. The predicted rise in growth necessitates a significant financial outlay, and the report already identifies how three companies, Amazon (AWS), Microsoft (Azure), and Google, already control over ⅔ of global cloud computing services. Services that are essential AI training and deployment infrastructure. What does that mean for global access to AI and its benefits? Will everyone benefit from AI?
Another area for malicious use that stands out is the creation of fake content. AI-manipulated or completely AI-fabricated imagery depicting an individual explicitly or compromised in some way can be used for blackmail or extortion. Cybercriminals can do this or more directly undertake cyberattacks on companies or infrastructure.
Malfunctions
Bias and misalignment are two important issues to consider when considering malfunctions. AI systems can reflect and even amplify existing biases in data. This can play out in areas like hiring, lending, and even criminal justice. It’s crucial to recognize that bias can creep in at many stages – from the data used to train the models to the system’s design itself.
Misalignment refers to how AI could deviate into pursuing goals that don’t align with human values, even if it is following its programmed objectives. Two ways this could occur are through Goal misspecification and Goal misgeneralisation. Goal misspecification refers to a mismatch between the objective given to an AI and the developer’s intention. Whereas Goal misgeneralisation is where an AI system follows a command in training but carries out the instruction differently outside of training in another environment.
Safety First
The report highlights the potential for accelerated development coupled with the need for careful oversight to ensure alignment with human values. Now, we are starting to see AI being used to fine-tune other AI; combined with a very tense and fractured geopolitical situation and growing wealth inequality, the need for strong governance is more important than ever. Although the report concedes that ‘Policymakers face a challenge in determining which capabilities warrant stricter regulations while supporting beneficial research…’
Reassuringly, the report does state that ‘Existing AI systems are not capable of undermining human control.’ And that ‘Experts agree that their current capabilities are insufficient to create any meaningful risk of active loss of control.’
The report concludes with an important message that ‘AI does not happen to us; choices made by people determine its future.’ With the impact predicted to be far-reaching and ‘profound,’ the call-to-action goes out for ‘an urgent need to work towards international agreement and to put resources into understanding and addressing the risks of this technology. Constructive scientific and public discussion will be essential for societies and policymakers to make the right choices.’
Adam Parlett is a cybersecurity marketing professional who has been working as a project manager at Bora for over two years. A Sociology graduate from the University of York, Adam enjoys the challenge of finding new and interesting ways to engage audiences with complex Cybersecurity ideas and products.
The opinions expressed in this post belongs to the individual contributors and do not necessarily reflect the views of Information Security Buzz.
