Text original publicat per No es época para tontos.
What happens when a justice system demands rigorous testing, traceability and permanent verification from artificial intelligence, yet largely assumes that a human judge remains cognitively fit to exercise enormous coercive power?
This investigation examines one of the most uncomfortable questions facing modern justice: should judicial decision-making remain exclusively human when ageing, cognitive decline, fatigue, inconsistency and institutional pressure are measurable risks?
The episode reviews medical data cited in the investigation on mild cognitive impairment, with prevalence rising from around 6.7% at ages 60–64 to 10.17% at 70–74, 14.8% at 75–79 and substantially higher levels at more advanced ages. The point is not that every older judge is cognitively impaired. The question is whether a system that can imprison people, decide custody, destroy fortunes and determine fundamental rights should rely only on presumed cognitive fitness instead of continuous objective verification.
The investigation then turns to Andorra. It contrasts Spanish judicial retirement limits with an Andorran framework that allows judicial activity to continue to 75 and permits substitute magistrates from neighbouring countries who may already be retired from their original judicial systems. The episode asks directly whether Andorra should assume a risk that neighbouring systems have already chosen to limit.
The alternative proposed is not a commercial chatbot. It is a Sovereign AI Judge operating on controlled infrastructure and using a closed, certified and versioned legal corpus. Every decision would have to follow a traceable chain:
FACT → EVIDENCE → LAW → PRECEDENT → REASONING → DECISION.
Multiple independent AI systems would challenge one another: one proposes the ruling, another tries to destroy it, another checks fundamental rights, another audits evidence, another searches for contrary precedent, and another tests for statistical bias.
The proposal also includes counterfactual testing, identity-swap tests, full decision reproducibility and systematic comparison with human judges. The objective is not to claim that AI is infallible. It is to replace part of the opaque, biological and non-reproducible error of human judgment with a system whose failures can be detected, replayed, audited and corrected at scale.
The suggested roadmap begins with a closed Andorran legal corpus, followed by a 12–24 month AI Shadow Judge experiment on anonymised real cases. Human and machine decisions would remain blind to each other and later be compared for legal accuracy, precedent use, omitted arguments, consistency, speed, bias and appeal outcomes. Only after demonstrated performance would the debate move toward binding low-risk decisions and, eventually, possible autonomous jurisdiction.
The final question is simple:
If we demand that AI prove it is fit to judge, why should a human decision-maker be exempt from proving the same?
