Augmented: Planners in an Era of Generative AI
The rapid development of generative AI, which focuses on creating new content such as images or text, is challenging preconceptions of computers' capabilities.
Generative AI is starting to power political campaigns, redesign streets, aid public communications, disrupt education, synthesize video and audio, and integrate with business software (Microsoft 365, G-Suite, Adobe). OpenAI's ChatGPT has set adoption records: 1 million users after 5 days and 100 million after 2 months.
These advances have the potential to improve planning processes, but raise significant concerns, and their scope and pace is thus far exceeding existing educational or regulatory processes. Large Language Models (LLMs) represent a technology of significant power whose ultimate bounds are currently uncertain.
Key AI Terms
Fine-Tuning: Adapting a pre-trained foundation model for specific tasks or domains by training on a smaller dataset. This focuses the model on specific tasks or domains and introduces important policy controls.
Foundation models: Enormous, general-purpose deep learning models that can be fine-tuned for different applications. These models often are identified as the basis for a new paradigm of large-scale training in machine learning research.
Generative AI: Advanced artificial intelligence models, such as ChatGPT, that create human-like text, images, or sounds based on input data, enabling diverse applications like conversational agents, content generation, and creative tasks.
Large Language Models (LLMs): Foundational AI models trained on massive datasets containing text from diverse sources. LLMs like those powering ChatGPT, learn patterns, grammar, syntax, context, and semantic relationships to generate coherent and contextually relevant text.
Policy Constraints: Rules or guidelines controlling the AI model's behavior during text generation, ensuring content meets specific requirements and aligns with desired goals.
Training Data Corpus: A large (typically multi-terabyte data set) of samples used to train AI models, which determines their base knowledge before fine-tuning.
How Does it Work?
These advances are stunning. They seem sudden, but they stem from a paradigm shift: behind the scenes, machine learning has been shifting to large-scale training, using terabytes of data to build massive, general purpose foundation models that can be specialized for different applications.
Over the past decade, resources (compute) used for top-end deep learning models have doubled every six months — a 1,000-fold increase every five years. These models are trained to build an internally consistent "world model" of sequences of text or patterns in images.
As a second step, foundation models are "fine-tuned" — sometimes with a single powerful computer, but often with significant human interaction to evaluate results. It's a critical step to enable conversational interfaces, safety, and support specific domain knowledge. How are these answers evaluated? By whom?
OpenAI, the makers of ChatGPT, have faced significant criticism about how they may have performed this process and by what means. OpenAI has obscured their methods with their most recent LLM, partly out of an abundance of caution about its potential ramifications of an open release. This has complicated existing concerns about the use of "black box" models in public contexts.
An entire open source LLM ecosystem has rapidly arisen in response, providing interesting alternative choices. These models tend to be smaller and more specialized, and often able to provide comparable performance for certain applications. For example, a recent study used only peer-reviewed climate research evaluated by experts (ClimateBERT). This was shown to provide significantly more accurate results than ChatGPT with climate-related queries.
More is Coming
With this scale and complexity, large-scale models such as LLMs have demonstrated "emergent behavior." Some have been generalized to complete tasks they were not specifically trained to do, and developer communities are very actively pushing their capabilities for task orchestration and context aware queries. Projects include:
- HuggingGPT — chains other models available in open-source communities to solve multistep tasks that can, for example, combine image recognition and language understanding.
- AutoGPT — attempts to make Chat-GPT autonomous through recursive prompting and expands it utility by connecting it to the web. It can self-prompt its way to abstract goals like "conduct a literature review on sea level rise and roadway abandonment."
- Adding context — researchers are finding methods that enable these models to provide more accurate or contextually-aware responses by using external knowledge stores that can be used inject additional context into a query. Imagine a model that could answer questions about a specific zoning code, plan, or arbitrary PDF.
LLMs are already affecting planners' work. During "A Planning Chat with ChatGPT" hosted by APA's research team and the APA Technology Division, planners report uses ranging from memo editing and curation to idea generation and more. A recent international roundtable of professionals attracted planning and technologists across the world. Discussions on the potential of this technology have included creating a common shared understanding of zoning codes or regulations across regions, as demonstrated by the Urban Institute and the Cornell Legal Constructs Lab. Beyond applications, many are aware of the risks of "stochastic parrots" ranging from systemic bias to environmental costs. APA has published several planning-specific publications, including a PAS QuickNotes on ChatGPT, PAS Memo on "Articial Intelligence and Planning Practice", and AI in Planning: Opportunities and Challenges and How to Prepare.
Based on these efforts and others, here's our initial SWOT analysis.
Using AI models in planning practice offers great promise and peril — and risks disrupting community economies. Reports from joint U.S. and EU research, the OECD, and Goldman Sachs explore the economic impacts of AI, noting that AI trends differ from previous automation eras, which reduced demand for routine middle-wage jobs but increased it for non-routine low and high-wage jobs. They think generative AI could replace a quarter of existing work, but this impact is likely to vary by sector and new jobs are likely to arise to replace them. This automation could boost labor productivity by 1.5 percent over a 10-year adoption period.
Who benefits from such gains in productivity is unstated, but the risks of increasing inequities are present. The OECD goes as far as to suggest governments work with stakeholders to "prepare for the transformation of the world of work and society."
Steps for Planners on Generative AI
This is a rapidly evolving topic. Everything including community engagement, how codes and policies are presented, what a plan is, and the economies of our communities could change. Near-term steps for planners:
- Procure responsibly: Ask vendors how they address bias. Familiarize yourself with data sheets for training data and emerging algorithmic auditing tools and open source alternatives. Demand accountability and transparency by providers, and start first with solutions that address existing community needs.
- Advocate for appropriate regulation: The White House's proposed AI Bill of Rights addresses privacy, algorithmic discrimination, and more. Initiating broader conversations is crucial to tackle accountability, transparency, and power imbalances.
- Plan with foresight: Planning for the future requires understanding disruptive trends and preparing to address them. Technology action plans and scenario planning already support decision-making under uncertainty.
- Plan ethically: Many aspects of the AICP Code of Ethics require planners to think before they naively integrate generative AI into workflows. We must not hide beyond models when our rules of conduct require us to provide "accurate information on planning issues." The APA Technology Division published an open letter to planners to consider the applications of generative AI in ethical practice.
- Apply the reversibility principle: To maintain urban resilience, initial applications of AI should be made in investments whose impacts are reversible. Precautions should be taken in applications that have the potential for harm or integrations are planned for important civic systems.
Much has changed, but much remains the same. We can focus on a humane vision of the future. We can plan for the needs of the communities we serve.
AI Resources from APA
- "ChatGPT: Implications for Planning," PAS QuickNotes, 2023
- Digitalization and Automation, 2023 Trend Report for Planners
- AI in Planning: Opportunities and Challenges and How to Prepare: Conclusions and Recommendations from APA's 'AI in Planning' Foresight Community, September 2022
- "Artificial Intelligence and Planning Practice," PAS Memo 111, April 2022
- "AI in Planning: Why Now Is The Time." Planning, Februrary 2022
- APA Technology Divison's Urban & Regional Planning Resources
Stanford University's 2023 AI Index Report
AI Language Models: Technological, Socio-economic, and Policy Considerations, OECD Digital Economy Papers, April 2023.
The EU and U.S. Diverge on AI Regulation: A Transatlantic Comparison and Steps to Alignment. Brookings Institution, April 2023.
"ChatGPT for State and Local Agencies? Not So Fast." GCN, March 27, 2023.
Blueprint for an AI Bill of Rights: Making Automated Systems Work For The American People. White House, October 2022.
"On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" FAccT '21: 2021 ACM Conference on Fairness, Accountability, and Transparency, Virtual Event.
Top image: Urbanist Grimoire — created with DreamStudio.