The technologies in this chapter are new, the marketing is confident, and the failure modes are the ones already catalogued in Part 4.
VeriPol was developed with Spain's Policía Nacional and deployed from 2018. It is a text classifier trained on robbery reports, some of which had been established as false, using linguistic features to flag reports likely to be fabricated. Reported accuracy in validation was high, and its deployment was credited with a rise in withdrawn false reports. It was withdrawn from operational use by 2025 following sustained criticism, including a legal and technical analysis arguing that the training data was small, that the ground truth labels were themselves the products of police processes rather than independent verification, and that a system trained on the outcomes of police discretion learns to reproduce that discretion.
iBorderCtrl was an EU-funded Horizon 2020 pilot run between 2016 and 2019 in Hungary, Greece and Latvia. Travelers answered questions from an avatar while a system analyzed facial micro-expressions for deception. It attracted extensive criticism on scientific and legal grounds — the underlying premise is chapter 56's, which failed at a billion dollars of scale — and litigation over disclosure of its documents. It did not proceed to deployment.
EyeDetect, from Converus, uses ocular-motor measures — pupil dilation, blink, fixation, reading behavior — during a computerized test. The physiological basis is chapter 47's: pupil response indexes cognitive load. The company reports accuracy in the high eighties; independent replication is limited and largely conducted by parties with an interest.
The underlying science is not empty. Recent work has found that fine-tuned language models can classify deceptive from truthful statements in text corpora at rates exceeding human performance, which is not surprising given the 54 percent baseline. There is genuine signal in language and in ocular behavior.
The failures are in deployment, and there are three, each of which has appeared already in this part of the guide.
Base rates. A classifier applied to a low-prevalence population produces mostly false positives regardless of its accuracy on a balanced test set. This sank SPOT and screening polygraphy and it sinks every border-scale application, as a matter of arithmetic rather than engineering.
Training data. A model learns the labels it is given. If the labels come from police determinations, the model learns to predict police determinations, including whatever demographic patterns those contain, and then returns them as objective output.
Automation bias. People over-trust machine outputs, particularly under time pressure and particularly when the output is numeric. An officer with a score in front of them will weight it above their own judgment, and — the crucial asymmetry — the accused cannot examine, cross-examine, or reproduce it. A human interviewer's reasoning can be challenged. A proprietary model's cannot.
Which is why the closing observation of Part 4 is not about technology at all. The methods in this part that work — strategic evidence disclosure, verifiability, cognitive load, the concealed information test, PEACE — produce artifacts that a third party can examine: a discrepancy, a checkable list, a recognition pattern, a recorded account.
The ones that fail produce a verdict about a person, from a process nobody outside can inspect.
Cross-ref: Part 11 algorithmic defense