AI Traffic Predictions May Be No Better Than Old School Math
Your traffic app's AI might not be smarter than a simple average.
A new academic paper audited popular AI models used for making predictions about things you encounter every day: traffic jams, disease outbreaks, and package deliveries. These models are called "graph neural networks" (GNNs) — a type of AI that learns from maps and networks, like street grids or flight routes. The researchers found something uncomfortable: these fancy AI systems often perform about the same as, or even worse than, simple, cheap math like averaging past data. That's a big deal because companies and cities rely on these predictions to deploy buses, staff hospitals, or stock warehouses.
The problem isn't just the AI — it's the way we test it. The industry tends to use a handful of standard datasets, including traffic measurements from Los Angeles (METR-LA), California highways (PEMS-BAY), chickenpox case counts, and delivery records. The researchers found that these benchmarks have hidden biases that make AI look better than it really is. For example, the common trick of comparing a day to the previous day ("differencing") can create a false impression of strong performance. When they used stricter statistical tests, simple linear models looked just as good as the AI.
What does this mean for you? Whenever an app says "arrival in 27 minutes" or a hospital predicts patient surges, there's a chance the "smart" AI behind it isn't any better than a calculator with a spreadsheet. Worse, the lack of reliable testing means billions could be wasted on complex systems that deliver no real benefit. The authors recommend that researchers compare their models against simple baselines first and use rigorous statistical checks before claiming success. They also show that combining simple math with AI can lead to better designs.
The takeaway isn't "AI is useless" — it's that we need to be more honest about what works. The next time you see a forecast, ask yourself: could this have been done with a simple average? If the answer is yes, then the fancy AI may be selling you hype.
- Fancy AI forecasting models often perform no better than simple, old-fashioned math.
- Popular test datasets for traffic, disease, and delivery predictions contain hidden biases that make AI look better than it is.
- Researchers want stricter testing; this could lead to more accurate and trustworthy predictions in your daily apps.
Why It Matters
Your daily traffic, weather, and delivery predictions may be less accurate than you think — and money may be wasted on overhyped AI.