Implications of AI & Generative AI Use

Negative Aspects of Large Language Models (LLMs) and Generative AI

Energy Usage:
- Large language models and the data centers that power them consume a considerable amount of energy.
- This energy consumption occurs during both the operational phases and the training phases of the models.
- Building the necessary data center infrastructure also requires significant energy resources.
- Predominantly, fossil fuels have been the primary energy source for data centers.
Increasing Demand:
- The complexity of LLMs and improvements in versions (as seen with OpenAI's ChatGPT updates) lead to increased electricity demand due to larger datasets and enhanced model refinement.
- The energy needed correlates with the aggregate outputs from these data centers.
Research Confirmation:
- Individual energy consumption when using AI technology is relative to other electronic devices.
- Studies suggest that, under certain comparisons, generative AI might not significantly raise energy demands beyond those of other common electronic items.

Cooling Requirements:
- Data centers with GPUs (Graphics Processing Units) generate a significant amount of heat requiring cooling measures.
- Cold water is typically used for cooling, resulting in a substantial consumption of fresh and clean water.
- There is ongoing debate on whether 'clean water' is necessary for cooling or if alternative sources could be used.
Fresh Water Access Concerns:
- As demand for clean water increases to cool data centers, access to fresh water becomes a significant concern.

Indirect Environmental Impact:
- The production of hardware such as GPUs requires critical minerals, leading to indirect environmental concerns from mining activities and the use of toxic chemicals.
- This emphasizes the concept of 'scope three emissions', which include all emissions not directly related to energy usage in data centers, but relevant to the entire lifecycle of data center infrastructure and hardware.
Mitigating Awareness:
- Environmentally, there’s little evidence to suggest that generative AI will meaningfully combat climate change in the short-term; however, there are discussions on operational adjustments to potentially mitigate negative effects.

Operational Energy Efficiency:
- Proposals include reducing GPU energy consumption without affecting model performance.
- Analogy provided: Dimming light bulbs can maintain effectiveness while reducing electricity use.
Energy Use Comparison:
- Approximately 50% of electricity spent training models focuses on achieving the last 2-3% of accuracy while outputs are often sufficient with lower energy consumption.

Technological Improvements:
- Average computer chip efficiency is improving rapidly, meaning less hardware is needed for the same processing power.
- Improvements of 50-60% annually in computational capabilities could considerably lower overall energy requirements, as just one chip now can accomplish tasks that required two previously.

Digital Infrastructure:
- Existing infrastructures for computing, especially in older systems like MS DOS mainframes, are less energy-efficient compared to newer developments.
- Discussion of the need for developing energy-efficient architectures at initial stages, before becoming entrenched in outdated systems.

Adaptive Infrastructure:
- Suggestions for potentially relocating data centers or improving their operations to lessen energy draw from non-renewable sources.
Using Generative AI for Environmentalism:
- Exploring how generative AI can predict energy trends, optimize renewable sources, or identify issues within green infrastructure to reduce carbon emissions.

Energy Use Comparison:
- Each AI prompt (such as those for ChatGPT) reportedly consumes approx 3 watt-hours of electricity.
- Comparatively, an average UK citizen generates around 4,500,000 watt-hours of electricity annually, indicating a negligible contribution from a single AI query.
Carbon Emission Factors:
- Estimated emissions of 2-3 grams of CO2 per query; this becomes more relevant when scaling up to multiple queries, but still very small compared to traditional daily energy usage.

Misinformation & Disinformation Risks:
- Rapid spread of misinformation aided by generative AI capabilities could lead to societal issues, including erosion of trust in credible sources.
- Heightened need for transparency and accountability within AI outputs.
Bias in AI Models:
- Concerns about inherent biases in training data perpetuating stereotypes (e.g., gender roles in images generated).
- Key demographic representations in datasets (given the 22% female professional representation in AI) tend to neglect experiences and perspectives from marginalized groups.

Digital Identity and Exploitation:
- Issues surrounding digitally created individuals (e.g., avatars) leading to possible financial exploitation of minority identities by their creators, typically from dominant cultures.
Intellectual Property Rights:
- Concerns about whether AI systems use individuals' intellectual properties without compensation raises ethical questions about data ownership and consent.

Cybersecurity Risks:
- Generation of realistic phishing scams using AI could increase susceptibility among individuals.
- The potential for deep fakes escalates the threat to data integrity and veracity.

Emerging Regulations:
- Different regions, notably the EU and Canada, are beginning to establish AI regulations emphasizing safety, ethical usage, and human monitoring.
- Challenges exist in achieving uniformity across jurisdictions and understanding the balance of innovation and regulation.

Urgency in Action:
- Continual evaluation regarding the role of AI within society is crucial as technology evolves faster than regulatory frameworks can often keep pace.
Long-term Consequences:
- Identifying appropriate mechanisms and best practices to engage with AI responsibly to ensure societal benefits while mitigating potential harms.