> ## Content Index > Fetch the complete content index at: https://www.zhaojinhua.com/llms.txt > Use this file to discover other available public pages before exploring further. # Two Types of AI in Transportation: Fantastic AI and Grounded AI - URL: https://www.zhaojinhua.com/two-types-of-ai/ - Published: 2026-09-20T19:45:47.000Z - Updated: 2026-09-28T01:47:44.000Z - Description: What Transportation Professionals Should Learn and Do Now - Author: Jinhua Zhao - Tags: AI for Transportation, Memos There are two types of AI in transportation: **Fantastic AI** and **Grounded AI**. They ask different questions, move at different speeds, and belong in different parts of our job. **Fantastic AI asks: What may become possible?** Physical world models; AI for social simulation; Trustworthy and verifiable AI; Multimodal AI; AI for transportation science; Recursive self-improvement (RSI). This is where the scientific frontier is moving, sometimes week by week. It is driven by the pursuit of AGI, not by any DOT's problems. Fantastic AI neglects implementation frictions and is evaluated by abstract benchmark. But most transportation professionals wake up to a very different reality. **Grounded AI asks: What actually works?** What works when: your data is incomplete, your infrastructure is only partially monitored, your organization uses legacy systems, and many processes are still done manually on paper, union rules matter, and contract negotiation outcomes still defy predictability, institutional memory exists in some staff’s minds; much of the organizational knowledge is implicit, drivers and riders have behavioral quirks that AI models do not expect. Grounded AI is oriented toward tangible value and evaluated in the field. Its projects move with the budget cycle, slowly. Grounded AI respect behavioral and organizational reality. That's the world transportation agencies actually operate in. The question isn't simply: “Can AI do this?” It's: Can we make it work in our organization? Can we trust it? Will it deliver measurable benefit? Can we deploy it at scale? Know which question you're asking. That's where clarity starts. Fantastic AI expands the possibility. Grounded AI produces value in the real world. The two are not rivals. Transportation Professionals must watch the frontier, and build on the ground. --- **Seven Fantastic AI Frontiers worth watching** **1\. Generative world models for mobility systems.** Large language models absorbed much of the internet but have limited physical and spatial intelligence. A world model learns space, time, physical dynamics and causal structure, so it can simulate alternative futures. For mobility, that means learning how traffic, demand, infrastructure and weather evolve and respond to an intervention (close a lane, retime a corridor, reroute a bus) and generating many plausible futures. Unlike a classic simulation or digital twin, its rules are learned rather than engineered. The frontier is to keep the conservation laws and learn everything else. **2\. AI for behavioral, social, and multi-agent simulation.** Agents can now be given personality, memory and social behavior, making it possible to simulate heterogeneous human behavior, social interaction, and community dynamics at much higher fidelity. This could answer the behavioral and social realism gap in today's models. The multi-agent part is growing: drivers, riders and operators increasingly interact with AI agents (routing apps, automated vehicles, booking assistants), and those interactions need simulating too. This is a research frontier. We can give agents personality and memory, but we cannot yet validate the resulting behavior against a real community, and early studies find language-model agents more average and more rational than the people they stand in for. **3\. Multimodal foundation models for transportation.** Economists describe a city through numbers, transportation engineers through networks, designers through images, historians through narrative. An agency holds the same variety: sensor time series, probe trajectories, camera video, LiDAR, satellite imagery, GTFS feeds, scanned as-built plans, inspection reports and work orders. A foundation model pretrained across these modes can be adapted to many tasks instead of one model built per task, and can begin to hold all these modes in one analytical frame, with the potential to enable interdisciplinary study of cities. **4\. AI for decision, control, and large-scale optimization.** Transportation is ultimately a decision problem: timing signals, scheduling buses and crews, pricing roads and parking, programming maintenance, routing an evacuation. Operations research solves these well when the problem is well specified and small enough. The frontier has three parts: learning to optimize problems too large or too dynamic for classic solvers; network-wide, real-time control instead of one intersection or one route at a time; and language models that turn a plain-language question into a formal model a solver can run. **5\. Trustworthy and verifiable AI.** The question is whether an output can be evidenced well enough for a licensed professional to stand behind it: legally compliant, ethically sound, and technically verified. Verification comes in grades: traceable sources, reproducible results, quantified uncertainty, and, for narrow cases such as a controller's safety limits, formal proof. Proof holds for the model, not for the world; the rest is earned through testing and monitoring. **6\. AI for transportation science.** Scientists integrate knowledge, form hypotheses and test them. Hypothesis generation has always been regarded as distinctly human. AI systems can now read a literature, propose hypotheses, design the analysis and run it. In transportation, that could mean mining years of fare-card or probe data for behavioral regularities no one thought to look for. The new risk is false discovery at machine scale: test thousands of hypotheses and some will look significant by chance. Human judgment stays essential but shifts to choosing the question, guarding against false discovery, and judging whether a finding is relevant and desirable for society. **7\. Recursive self-improvement.** An AI system autonomously analyzes, rewrites, and optimizes itself or its training pipelines to increase its own capability. This creates a compounding, exponential feedback loop that could lead to an "intelligence explosion." RSI draws much of Silicon Valley's attention and is seen by many as the path to AGI. AI already speeds up parts of AI research, such as writing code and running experiments; end-to-end self-improvement has not been shown. No DOT will procure it soon, but every DOT will feel it: models will improve faster than contracts run. Own your data and your workflow; treat the model as a replaceable part. Fantastic AI is not driven by DOT problems, yet it has the promise to solve the frictions that ground us: — Generative world models could fill in the infrastructure we only partially monitor, and test an intervention before we build it. — Behavioral and multi-agent simulation could anticipate the behavioral quirks of drivers and riders that today's models miss. — Multimodal foundation models could read the paper processes and help capture the knowledge that lives only in staff's minds. — Decision and control AI could coordinate the signals, routes and schedules we still tune one at a time. — Trustworthy and verifiable AI could give a licensed engineer something to sign. The last two frontiers solve no friction directly. They set the pace for everything above. Each frontier is still unproven where it matters most: on real roads, with real data. **Grounded AI in Transportation Differentiators** A general-purpose AI strategy is not enough for DOTs. Because transportation has at least four characteristics that make AI substantially harder. 1\. **We operate in the physical world**. LLMs are remarkably good at language: reading, writing, summarizing, and finding patterns across text and symbols. Transportation decisions play out in a physical world, where weather, geometry, vehicle dynamics, human attention, and behavior interact in real time. An AI can speak fluently about making an intersection design and still be wrong about what happens during a snowy, unprotected left turn. Language is forgiving. The physical world is not. 2\. **We are safety-critical**. In transportation, 99.9% accuracy is not good enough. The failures are long-tailed and the 0.1% is what keeps us awake at night. For significant decisions in transportation, someone must still verify the output, own the decision, and sign. 3\. **We are half infrastructure, half human system**. Transportation models love infrastructure. But people are often the most consequential, and far less predictable. Drivers. Riders. Operators. Communities. Institutions. Their behaviors, incentives, trust, and constraints often determine whether a transportation project succeeds. 4\. **Our data is multi-channeled**, Numbers. Networks. Images. Video. Text. Sensor streams. Inspection reports. Customer complaints. Transportation AI needs to absorb all these forms as meaningful data and interpret them in a common representation. **We need an Grounded AI playbook built for transportation.** --- **What transportation professionals should do?** Transportation professionals cannot wait for a perfect AI strategy or the next generation of tools. Start building the professional judgment and capability to use AI now. I suggest 3 levels of action. **Level 1 individual productivity boost.** 1\. AI literacy: everyone engages; 2 hours of learning saves an hour the next day. 2\. Role change: from creator to critic, communicator, negotiator. 3\. Judgement and taste: domain expertise matters more. **Level 2 teams and projects.** 1\. Accountability > productivity. Producing an artifact is getting easier. Owning the decision is not. Someone must verify the output, accept responsibility and sign. 2\. Examine the business process before using AI: remove, simplify and then automate. 3\. Design human-AI workflows. 4\. Build in-house versus procure from outside: AI, particularly vibe coding and AI agents, is moving that boundary. **Level 3 executives and C-suite**. 1\. Building trust is the leader’s primary job. 2\. The apprenticeship bargain is quietly breaking: the tedious work that trained junior staff is the work AI does well. But it is through tedious work that one builds judgment and taste. 3\. Benchmark augmentation not just automation. --- In fall 2026 Mobility Forum has 3 upcoming webinars on AI in Transportation: Oct 2: **AI Frontier in Transportation Research: What Happens When AI can Understand, Simulate, Generate, and Reason about Transport Systems?** Registration link: [https://luma.com/jinhua-s6df](https://luma.com/jinhua-s6df?ref=zhaojinhua.com) Oct 9: **Grounded AI in State DOTs: Caltrans, TxDOT and UtahDOT** Registration Link: [https://luma.com/jinhua-z0bl](https://luma.com/jinhua-z0bl?ref=zhaojinhua.com) Dec 4: **AI-Native Operations for Public Transportation with Amos Haggiag CEO of Optibus and Jonny Simkin CEO of Swiftly** Registration Link: [https://luma.com/jinhua-3qik](https://luma.com/jinhua-3qik?ref=zhaojinhua.com) Secondly, My team at the JTL Urban Mobility Lab, with support from Google, is launching a major project, , with the Washington Metropolitan Area Transit Authority (WMATA)—one of America’s largest transit agencies, operating Metro rail and bus services across the nation’s capital Washington, D.C. region. Together, we will develop an **open-source agentic AI platform for public transportation monitoring, control, and rider communications**. This work is an important testing ground for my work on Grounded AI. As my team and I work alongside leading transit agencies in Washington, Chicago, and elsewhere, I’ll share what we are learning from the ground—what works, what doesn’t, and what it really takes to make AI useful in the messy reality of public transit in my memos. Third, I am considering to offer workshops on AI fundamentals for Transportation Professionals. Register your interest here: [zhaojinhua.kit.com/ai](https://zhaojinhua.kit.com/ai?ref=zhaojinhua.com) AI can generate options. It cannot decide what transportation future we should build. That remains our job. *–* Jinhua