Academia is obsessed with solving real-world violence with software. The latest iteration of this flawed trend comes from Lethbridge Polytechnic, where researchers secured $3.5 million in grant funding to build a system called DORA—Dynamic Operational Risk Assessment. The premise sounds logical on paper: front-line police officers have a tight 24-hour window after an arrest to make bail recommendations. They lack data. So, let us build a tool that uses decades of risk assessment factors to scientifically back up police decisions and prioritize compliance checks.
It is a comfortable consensus. It satisfies politicians demanding a fix to recent federal legislative shifts on bail. It satisfies police agencies wanting a scientific shield for their choices.
It is also dangerous, misguided, and destined to fail.
I have watched public sector agencies dump millions into predictive risk tools for over a decade. The story is always identical: a massive injection of cash, a pilot program using historical data, a flashy press release about evidence-based governance, and a complete failure to recognize how human systems actually react to automation.
By attempting to turn the deeply flawed, high-stakes human calculation of bail into a standardized data problem, we are not fixing the justice system. We are giving it a high-tech excuse to stop thinking.
The Myth of "Minimal Information" Prediction
The foundational flaw of the DORA project—and tools like the Public Safety Assessment or COMPAS before it—is the belief that you can extract a reliable predictive risk score from limited, low-quality data.
Think about what happens during a standard 24-hour arrest window. The information available to a front-line officer is messy, incomplete, and highly subjective. It consists of immediate arrest details, a cursory look at local police databases, and a criminal record wrap sheet that may or may not be updated.
When you feed this highly fragmented, low-quality input into a predictive algorithm, you do not magically get a precise output. You get amplified bias.
Statisticians call this the garbage-in, garbage-out principle. If the underlying data is sparse, the algorithm must lean heavily on static historical proxies to generate a score. What are those proxies?
- Previous failures to appear in court.
- Number of past arrests (not convictions).
- Socio-economic markers masquerading as "stability variables."
A tool designed to operate on minimal information cannot calculate dynamic risk. It calculates systemic history. It punishes an individual not for the likelihood of what they will do, but for the chaotic reality of what their demographic or zip code has already endured.
Imagine a scenario where a person misses a court date because they hold two minimum-wage jobs and lack access to reliable transit. To an algorithm, that is a data point indicating a failure to comply, instantly elevating their risk score. The tool cannot parse the difference between defiance and poverty. It simply flags the user as a threat, driving up detention rates or triggering aggressive, unannounced police surveillance.
The Rubber-Stamp Trap: How Science Blinds Discretion
The developers of these tools always include a boilerplate disclaimer: "The software is just there to support human decision-making, not replace it."
This is an administrative fantasy. I have sat in rooms with decision-makers who are handed a software score. Do you know what they actually do? They look at the color-coded risk flag—red, yellow, or green—and they match their recommendation to the box.
This is driven by a powerful psychological phenomenon known as automation bias. When a human operator is presented with an automated recommendation, they default to trusting it because the software is framed as objective and scientific.
More importantly, it offers absolute bureaucratic cover. If a police officer or a Crown prosecutor recommends releasing a defendant who later commits a violent offense, the human takes the blame. Their career is on the line. But if they follow the recommendation of a $3.5 million AI tool, the blame disappears into the software architecture.
By providing a tool to "scientifically back up" police opinions, projects like DORA do not enhance discretion; they eliminate it. They turn front-line officers and judges into rubber stamps for statistical models. The moment you introduce a mathematical score into a bail hearing, the nuanced, qualitative observation of the human being standing in the dock is replaced by an unyielding number.
The Hidden Cost of Prioritizing Compliance Checks
One of the advertised features of Lethbridge's DORA tool is its ability to prioritize unannounced bail compliance checks based on offender risk level rather than chronological order. On the surface, targeting resources at high-risk individuals sounds like a smart optimization trick for underfunded police departments.
In reality, it creates a self-fulfilling feedback loop that breaks the integrity of criminal justice data.
Consider the mechanics. If the algorithm flags an individual as high-risk, the police conduct frequent, unannounced compliance checks at their home or workplace. Because they are being watched constantly, any minor technical breach—like being five minutes late for a curfew or missing a phone call—is instantly caught and logged.
Meanwhile, an individual flagged as low-risk receives zero compliance checks. They could be violating their conditions daily, but because no one is looking, their record remains pristine.
When the algorithm is updated with new data six months later, what does it see? It sees that the individuals it flagged as high-risk had high rates of non-compliance, while the low-risk individuals had low rates. The software developers throw a party and declare their tool highly accurate.
It is not accurate. It is an echo chamber. The tool did not predict the risk; it dictated the policing behavior that manufactured the data point. This loop deepens systemic disparities while doing absolutely nothing to stop actual, violent re-offending.
The Flawed Premise of the "Bail Crisis"
The push for tools like DORA is a direct reaction to recent federal legislative changes in Canada, which require certain repeat offenders to demonstrate why they should be granted bail, effectively reversing the onus of proof.
The political narrative is that the bail system is broken because it is too lenient, and that data tools will help us separate the truly dangerous from the safe. This premise is entirely wrong.
The real crisis in our bail system is not a lack of predictive software; it is an absolute collapse of institutional capacity.
- Legal aid systems are chronically underfunded, meaning accused individuals cannot get proper representation during the critical 24-hour window.
- Courthouses are bottlenecked, turning what should be a swift assessment into days of limbo.
- Mental health and addiction resources—the actual drivers of erratic bail behavior—are virtually non-existent on the front lines.
Throwing millions of dollars at an algorithmic risk assessment tool is an exercise in resource misallocation. It is an attempt to use software to paper over structural decay. We are spending millions to build a more efficient thermometer instead of treating the underlying infection.
Stop Scoring Risk, Start Funding Infrastructure
If we want to fix the bail process, we need to abandon the obsession with predicting human behavior via software. No algorithm can tell you with certainty what a desperate person will do in the next 72 hours.
Instead of building predictive tools, that $3.5 million in grant funding and the structural focus of our polytechnics should be directed toward building concrete operational infrastructure that removes the volatility from the bail window.
First, invest in real-time coordination systems that connect front-line officers with mental health diversion teams at the moment of arrest, rather than relying on historical wrap sheets. Second, fund immediate court-navigation services that address the logistical reasons people miss bail hearings, such as text-message reminder systems and subsidized transport, which have been proven to reduce failures to appear far more effectively than aggressive police surveillance.
Admitting that software cannot solve systemic human crises is uncomfortable. It means admitting that the solution requires slow, expensive, unglamorous structural work rather than a sleek user interface. But until we stop trying to turn human desperation into a data point, we will continue to build tools that protect institutions while leaving communities exposed. Turn off the software, step away from the predictive scoring models, and start fixing the real physical infrastructure of the justice system.
