The Computer Found a Car. Police Stopped Looking.

A few weeks ago I wrote about what “human in the loop” is supposed to mean. Lindsey Isaacs’s case is a different version of the same problem, and in some ways it bothers me more because the basic verification problem is so painfully ordinary.

After a fatal chain-reaction crash on Interstate 4 in Florida, investigators were looking for a Dodge Durango. Modern automated license-plate-reader systems are built for exactly this kind of search: take a description, search a huge stream of vehicle observations, and return possible candidates much faster than a human could review cameras one by one.

The system returned a Dodge Durango associated with Lindsey Isaacs. That should have been the beginning of a question. Instead, it appears to have become the beginning of an answer.

The car was black

Isaacs owned a black Dodge Durango. Evidence ultimately pointed investigators toward a maroon Dodge Durango. A witness had reported a partial plate beginning with 458. Isaacs’s plate was different. The vehicle ultimately associated with the case had a plate matching that partial information much more closely.

This is where the story stops being primarily about artificial intelligence and starts being about ordinary investigative reasoning. A search result is not a conclusion. It is a row returned by a query. If I ask a database for vehicles matching some parameters, the database can give me a perfectly accurate answer to the question I asked while I am still asking the wrong question.

Once investigators reached Isaacs’s actual vehicle, they had something better than an algorithmic result: the physical car. Color, plate, damage, timeline and witness information could all be compared against the theory. A later specialized review reportedly concluded very quickly that her vehicle did not have damage consistent with involvement in the crash.

That should be the standard. The machine gets to propose. Reality gets to dispose.

Automation bias doesn’t require autonomous AI

We talk about dangerous AI as though the problem begins when a computer is allowed to make a decision by itself. There is another failure mode that is already here: a computer gives a human a plausible answer, and the human unconsciously changes the question from “what happened?” to “how does the evidence fit this answer?”

That is anchoring with a very expensive user interface.

Flock did not put Isaacs in jail. Human beings made those decisions. That distinction matters, because simply making the software more accurate does not fix a workflow in which a candidate can quietly acquire the authority of a conclusion.

I build systems that score, classify, search and recommend things. I like those systems. They can make people dramatically faster. But there has to be an architectural distinction between “the computer found something interesting” and “we have established that this is true.”

If your system returns a black Durango, look at the damn Durango. Check the plate. Check the damage. Check the timeline. Check the witness statements. If those facts disagree with the machine-generated lead, the correct response is not to explain the discrepancy away. Throw the lead away.

The computer should make us faster, not stupider

Computers are extraordinarily good at finding things. They can search millions of license plates, photographs, transactions, log entries and hours of video. That is useful precisely because human beings cannot practically perform those searches themselves.

But search is not judgment. A database match is not guilt. A similarity score is not identity. A model output is not evidence merely because it arrived on a screen with decimals after it.

The lesson from Isaacs’s case should not be reduced to “technology bad.” The more useful lesson is that powerful search tools require stronger verification habits, not weaker ones.

The computer found a Dodge Durango. The job of the investigators was to determine whether it was the Dodge Durango.

Those are not the same thing.