Learn about AI

Introduction

Welcome to the AI City!

We want Southampton to become an example of how to make the best use of artificial intelligence.
We want to create an AI vision that works for everyone in Southampton.

What does artificial intelligence mean for Southampton’s future? What should AI do for the city, and what should it absolutely not do?

We are planning community collaboration to establish a shared vision for AI implementation in Southampton.

A series of workshops over the coming weeks will bring together residents, businesses, politicians, charities, scientists, and other community groups

Participants will meet and work together in groups to understand AI opportunities and risks in Southampton.

What is AI?

AI just means computer programs that perform tasks typically associated with humans.

Current examples include image recognition and text or image generation.

Narrow AI: Algorithms designed for specific, single tasks

Broad AI: Algorithms capable of performing multiple different tasks

General AI: Hypothetical - but could potentially surpass human intelligence

AI can...

• Predict: Analyse past data patterns to forecast future trends (beyond weather to new applications)

• Classify: Categorise images, videos, or sounds (from medical diagnosis to general object recognition)

• Use Language: Generate and respond to human language for more natural technology interaction

• Optimise: Find best solutions among multiple options (like balancing electricity supply and demand)

• Generate: Create new ideas, designs, or art based on learned patterns

• Suggest: Provide recommendations by observing user behaviour and offering coaching

• Monitor: Track systems and alert users to unusual activities or patterns

• Control: Connect to and operate other systems (like home heating)

• Combine: AI systems can integrate multiple capabilities (example: self-driving cars use classification, prediction, optimisation, and control together)

We want to focus on realistic, useful applications of existing technology rather than speculative future developments.

Large Language Models

LLMs are advanced AI systems trained on massive text datasets that can interact using normal human language through prompts. LLMs are powerful but imperfect tools that require thoughtful application in civic contexts.

Language serves as vehicle for information, and LLMs trained on content from books, newspapers, and websites pick up knowledge about the world in the process.

Training on diverse text sources creates powerful, versatile tools.

LLMs are "Completion Machines". They predict the most likely (statistically plausible) information based on input provided. By doing so, they can answer questions and explain complex topics in accessible, understandable ways. We can ask them questions and receive responses.

Machine Learning

Most new AI is based on machine learning (ML), which extracts patterns from large datasets and captures them in models.

Most AI projects begin as data collection and management initiatives (e.g., collecting scans from breast cancer screening). We need to clean and prepare data for training, a process that often requires humans to create categories for model training. For example, by labelling data with correct categories such as whether a picture from a microscope is 'cancer' or 'not cancer'.

We feed labelled data into ML algorithms to create a model that can then classify new, unlabelled data.

The Large Language Models use the same principle, but they are much more sophisticated . These are trained on massive datasets including books, newspapers, and websites. They generate new text based on captured patterns from training data.

A key challenge in AI projects is collecting sufficient relevant data for new applications. Models need enough diverse, well-labelled data to perform accurately on new, previously unseen inputs.

Ethical Challenges

Artificial intelligence presents both opportunities and challenges for cities and society that require careful consideration. Some of the key questions surrounding the ethical challenges are summarised below.

Will AI eliminate existing jobs, or change how jobs are performed?

AI models can be biased. For example, AI trained on current salary data might conclude women should be paid less than men. Biased AI systems can worsen existing inequalities and social divisions.

There are challenges in identifying biases in AI systems.

• Who should be responsible for spotting AI biases?

• Who should be accountable for fixing identified problems?

• What rights do people have when treated unfairly by AI systems?

• How practical will it be to challenge AI-based decisions?

And as AI automation makes it easier to process personal data, how can we ensure protection of data and privacy rights?

What happens when an AI system says 'no'? Is there a need for human intervention capabilities when AI systems fails?

Technical Challenges

AI isn't magic, doesn't always work, and achieving even basic functionality requires significant effort after decades of research. It is important to understand that AI has limitations to implement it responsibly and effectively.

Bias and misinformation

The "Garbage in, garbage out" principle applies to all AI algorithms. Errors or biases in training data directly transfer to the resulting AI models. It's difficult to collect unbiased data due to existing societal biases and misinformation in books, newspapers, and online sources.

Power and Resources

AI requires extremely powerful computers for model creation and operation. They consume substantial amounts of electricity for processing and cooling, and demand significant physical materials for electronics and physical infrastructure.

Truth

AI has a difficult relationship with the truth. LLMs capture language patterns that often include truthful information (e.g., "the sky is blue") although they are only statistically likely word sequences rather than "knowing". AI can generate grammatically correct output, but often the output is not factually accurate. AI has no inherent drive for accuracy or truth-seeking, this differs from humans, who have beliefs based on experience.

Explainability

AI models often function as "black boxes" - we can assess if they work but not how they work. AI tools cannot provide reliable explanations for their decision-making processes. Even when asked "Why did you say that?" AI responses are just statistically probable text, not actual explanations of what happened!

Challenges don't make AI useless, but require careful application. For example, avoiding certain AI types in safety-critical settings can mitigate risks, or implementing human review for AI-made decisions. We want to focus on finding safe and effective applications with appropriate safeguards.

Case study: Traffic

Congested streets and long waits are daily frustrations. AI traffic management has the potential to ease urban transportation challenges.

Using real-time sensor data from cameras, GPS, and other monitoring devices AI can process vehicle flow information across an entire city to create a more comprehensive understanding of traffic flow.

Replace fixed timer systems with AI-driven dynamic adjustments that respond to real-time traffic conditions.

AI could encourage sustainable transport by prioritising pedestrians and cyclists, or accommodate emergency vehicles for improved safety and better response times.

Such systems need constant and reliable data streams, and raise privacy concerns about monitoring and data collection. So we need appropriate safeguards and protections.

Case study: Engaging Communities

AI can be used as a tool to make local government more responsive and inclusive while maintaining transparency in civic processes. The system can predict which community initiatives will have the greatest positive impact and amplify citizen voices to strengthen democracy at the local level.

AI could make resident participation in local decision-making more accessible and include more people.

It could create more responsive government-citizen interactions - for example by instantly processing community input from multiple communication channels and overcoming traditional resource limitats on volume and depth of analysis.

AI can analyse thousands of comments to identify key community concerns to help officials understand what matters most to residents, organisations and businesses.

It has the potential to breaks down language barriers through multilingual interfaces, and offering multiple ways for residents to participate.