Nov 2, 2025
AI Incidents Forecasting
Chamod Kalupahana, Ahmed Elbashir, Christopher L. Lübbers
This research develops a framework for forecasting AI incidents to help predict future risks. We have developed two models that forecasts incidents which include calibrated 90% prediction intervals with backtests. These models predicts a large growth in the total number of incidents over the next five years. This has generated forecasts with quantified uncertainty, enabling policymakers to shift from reactive to proactive risk mitigation, supporting evidence-based regulation and responsible AI deployment.
I think this line of research is quite exciting, and trying to build models on the AI incidents database sounds like a great idea. I like how you included backtesting.
One failure mode this may have is that it assumes ceteris paribus, i.e. that all conditions (e.g. safety regulations, regulations and conditions of mandatory reporting, etc.) stay the same over the near future, which is unlikely.
* Great call using a preexisting work for data preprocessing!
* Why did you make the design decisions you did for the model (trend-change hinge @ 2021; spline flexibility constraint)? Make sure all decisions like this are thoroughly justified, and that this is conveyed to the reader.
* Figure 1 uses a scale which makes it quite difficult to evaluate how well your model does on already logged data. Additionally, I think it is unlikely that incidents will continue on the upward trend indefinitely; as some point, like any technological adoption, there will be an inflection point and the rate of change will reduce. This should be taken into account, or at least mentioned, when presenting this kind of plot.
* You mention that incidents are distinct from reports, which definitely should be mentioned, but I think this could also be used as another feature to broaden the context provided. Did number of reports have similar behavior to number of incidents? What inferences might we be able to draw from this additional source of information?
* It would also be interesting if an analysis could be conducted on the number of incidents that happens in a given year vs. the number of incidents reported here. Although we can't know this for certain, one could attempt to approximate this with comparisons to other crowd-sourced reporting endeavors and historic analysis. This is similar to the "future work" you present.
Cite this work
@misc {
title={
(HckPrj) AI Incidents Forecasting
},
author={
Chamod Kalupahana, Ahmed Elbashir, Christopher L. Lübbers
},
date={
11/2/25
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
