The claim that AI can detect “90% of crimes before they happen” is misleading. After tracing the article back to the original research, I found that the AI does not predict specific crimes before they happen in the way the headline suggests. What it actually does is analyze past crime data and forecast where certain types of crime are more likely to occur within roughly 1,000-foot areas over the following week. The “90%” figure refers to the model’s average performance in that hotspot-style forecasting in Chicago, not to detecting exact crimes, exact people, or exact future events.
The main primary source is the original peer-reviewed study published in Nature Human Behaviour, “Event-level prediction of urban crime reveals a signature of enforcement bias in US cities.” The researchers say they created a model that forecasts crime by learning patterns from reported crime events, and the paper says the model achieved a mean area under the ROC curve of about 90% in Chicago for crimes predicted per week within about 1,000 feet. That is very different from saying it can detect 90% of crimes before they happen. The study is about forecasting patterns in place and time, not identifying specific future criminal acts. Link: https://www.nature.com/articles/s41562-022-01372-0
Another useful primary-source-style source is the University of Chicago’s writeup on the study. It explains that the algorithm predicts future crimes one week in advance using public data on violent and property crimes and that it was tested in Chicago and other major US cities. This source helped confirm what the model actually does and also made clear that the system relies on historical crime reports, which matters because reported crime data is not the same thing as all crime that actually occurs. Link: https://biologicalsciences.uchicago.edu/news/algorithm-predicts-crime-police-bias
The main secondary source connected to the claim is the IFLScience article by Jack Dunhill. This is the article that frames the story with the headline saying AI predicts 90 percent of crime before it happens. After comparing it to the original study, I found that the article is based on real research, but the headline and wording make the technology sound more powerful and precise than it actually is. Link: https://www.iflscience.com/ai-predicts-90-percent-of-crime-before-it-happens-creator-argues-it-wont-be-misused-65025
I also found a Bloomberg summary of the study. That coverage is more careful because it explains that the model divides cities into approximately 1,000-square-foot tiles and predicts future crime events based on historical data patterns. That helped me see that the stronger “before it happens” framing is more of a media simplification than what the research itself claims. Link: https://www.bloomberg.com/news/articles/2022-06-30/new-algorithm-can-predict-crime-in-us-cities-a-week-before-it-happens
The Nature Human Behaviour study is the strongest source because it is the original research, but the authors still have an interest in presenting their work as important and innovative. The University of Chicago writeup is useful, but it is also promotional since it comes from the institution connected to the researchers and naturally highlights the study’s impact. IFLScience is a science news site that uses attention-grabbing headlines, so it has a clear incentive to make the story sound dramatic and clickable. Bloomberg is a news outlet, so while it is more measured, it is still simplifying technical research for a general audience.
There is some truth behind the claim because the model really did show strong performance in predicting where crime was more likely to occur based on past reported incidents. The original paper reports about 90% performance in Chicago using its evaluation measure, and both the University of Chicago summary and other coverage confirm that the model could forecast general crime patterns about a week ahead. So the claim is not made up from nothing. It is based on real research about predicting higher-risk areas and time windows.
What undermines the claim is that the AI is not actually detecting individual crimes before they happen. It is forecasting probabilities in geographic zones using past crime reports. That is a much narrower and less dramatic claim than the headline suggests. The original paper itself describes spatio-temporal forecasting from event reports, not mind-reading or identifying exact future criminal acts. The study also discusses enforcement bias, which matters because crime data reflects what gets reported and what gets policed, not necessarily all crime equally. That means the system’s predictions can be shaped by gaps in reporting and existing policing patterns.
I was not able to document a direct response from the author or IFLScience before finishing this fact-check. I also did not find a correction attached to the article during my review. Because of that, I relied on the published article itself, the original study, and additional reporting to trace what the claim was actually based on.
Overall, I would rate this claim as misleading. The AI did not prove it can detect 90% of crimes before they happen. What the research actually shows is that a model can use past crime data to predict where crime may be more likely to occur in the near future, and that is a very different claim.
