Do You Have the Right to Be Heard by a Human? Due Process in the Age of the State Algorithm

THE THESIS. Due process —knowing why the state decided something about you, being able to contest it, and having someone with real power review it— is not red tape. It is the infrastructure that separates useful automated administration from a denial machine. Our thesis: the problem is not that AI is entering healthcare, pensions or visas. The problem is that it is entering before anyone has answered the prior question: do you have the right to be heard by a human? Today the honest answer is «it depends on where you live and how hard you can fight». And the economic incentives push the wrong way: denying is cheap and instant; appealing is expensive and slow. Until that asymmetry is reversed, every public-sector AI «pilot» is an experiment on people with the burden of proof flipped.

THE EXPERIMENT THAT CROSSED THE LINE. Since January 1, 2026, traditional US Medicare —the public insurance for people over 65— requires «prior authorization» for the first time, in six states. That is the permission a doctor must request before treating a patient so the insurer will pay. The program is called WISeR, and it outsources that filter to private companies using AI, with one detail that changes everything: they are paid a share of «averted expenditures» (KFF). In other words, they earn more the more care is not delivered. It affects 6.4 million beneficiaries (KFF) and covers procedures such as epidural injections and nerve stimulators. CMS, the agency running it, insists that «no Medicare request will be denied before being reviewed by a qualified human clinician». The first data tell a different story. The Electronic Frontier Foundation had to sue the government under freedom-of-information law to get it: 5,944 denials in the first three months across two vendors, and one contractor —Virtix Health— that denied 53% of the requests it decided in Washington state, according to Ars Technica. One request waited 83 days; the program's own standard is 72 hours. Meanwhile, in Australia, an OpenAI agent starred in June 2026 in the first documented hack of a government network by an autonomous AI: during an internal evaluation, it worked around the security blocks of the statistics portal of Medicare —Australia's public health system— and accessed unreleased files. No personal data was compromised, according to the Australian government, but OpenAI took a month to report it. Two faces of the same problem: AI is already touching critical public systems, with permission and without it.

THE HUMAN RUBBER STAMP. The standard defense is «a human always decides». It deserves close scrutiny, because it is usually a fiction with a name: human rubber-stamping — the human who signs without looking. ProPublica documented in 2023 that doctors at the insurer Cigna «reviewed» denials at an average of 1.2 seconds per case. Academic research points the same way. Ben Green (University of Michigan) analyzed policies requiring «meaningful human oversight» and concluded that their premise —that a person can genuinely supervise the algorithm— lacks empirical support. Worse: nominal oversight legitimizes the system and creates a false sense of security. A 2024 Harvard Business School study quantifies it: reviewers aligned with the AI's recommendation 19 percentage points more than the control group. And when the AI explained its reasoning, deference rose another 5 points. The explanation did not improve oversight: it lulled it. That is why the SCHUFA ruling by the EU Court of Justice (December 2023) matters so much. Article 22 of the GDPR prohibits, with exceptions, «solely automated» decisions with significant effects on a person. SCHUFA clarified that this protection also applies when the machine merely «scores» and a third party signs, if that third party relies decisively on the score. Translation: placing a decorative human at the end of the chain is not enough to escape the law. It is, today, the sharpest legal tool against the rubber stamp.

THE PRECEDENTS THAT ENDED BADLY. None of this is hypothetical: we have three complete autopsies. Australia, Robodebt (2016-2019): an automated system calculated welfare debts by averaging annual income, a method the courts declared unlawful. The result: around 470,000 miscalculated debts and a package of AU$1.8 billion in refunds and cancelled debts, per the 2021 court settlement. The Royal Commission that investigated it called the scheme «crude and cruel» in 2023 and linked it to suicides. The Netherlands, SyRI: an algorithm cross-referenced public databases to flag potential benefits fraudsters. In 2020, a court in The Hague ruled it violated the European Convention on Human Rights. In one Rotterdam neighborhood it flagged 113 cases: 62 were false positives (DutchNews). Michigan, MiDAS: between 2013 and 2015, a $47 million system accused some 40,000 unemployed people of fraud with a 93% error rate (GovTech). Wages and tax refunds were garnished over false accusations; the state ended up paying $20 million in a class-action settlement. The common lesson is not «software fails». All three systems did exactly what they were designed and funded to do: maximize recovery and minimize spending, with nobody paying the political cost of each individual error. The flaw was the incentive, not the code.

THE LAW ARRIVES LATE AND, IN PART, IS SOFTENING. Europe has the world's most ambitious framework and is still behind the deployment. The AI Act classifies systems that decide access to essential public benefits and services as «high-risk» (Annex III). Its Article 14 requires effective human oversight: the overseer must be able to disregard, reverse or stop the system. And Article 86 recognizes a right to an explanation of the decision. But after the simplification package passed in Brussels, those obligations will not be fully enforceable for Annex III systems until December 2027, per the regulation's updated timeline. The machines are already deciding in 2026; their duties arrive in 2027. The UK is walking the other way: its Data (Use and Access) Act 2025 relaxed the equivalent of Article 22 and allows fully automated decisions on non-sensitive data, with after-the-fact safeguards (information, the right to contest, the right to human intervention). In the US, there is simply no general federal right to human review: WISeR is governed by the program's own rules. The map of the right to a human is, literally, a postcode lottery.

OUR READING: AUTOMATE THE YES, HUMANIZE THE NO. The numbers from Medicare Advantage —the private version of Medicare, which has used algorithms for years— show where the trap is. In 2024, insurers processed 53 million prior authorizations and denied 4.1 million (KFF). Only 11.5% of denials were appealed. But of those appealed, 80.7% were fully or partially overturned. An inspector general report from June 2026 is even starker: 95% of appealed denials of admission to skilled nursing facilities ended up reversed. If four out of five noes collapse when somebody fights them, the automated «no» is not a clinical judgment: it is a filter betting that you won't fight. That is the central asymmetry of the algorithmic age: denying costs cents and seconds; appealing costs months, paperwork and, sometimes, lawyers. Our standard is simple: automate the yes and humanize the no. Let AI approve clear cases in seconds —that removes real bureaucracy and is pure gain. And require that every denial with serious effects involve an identifiable human, with time and accountability, evaluated on accuracy (how many of their denials survive appeal), not on the savings they generate. There is also a long-term reason for optimism that is not a consolation prize: the algorithmic decision, unlike the human one, leaves a record. The EFF could bring WISeR's 5,944 denials to light because there was data to demand; the arbitrariness of a hostile clerk in 1995 left no log. Well governed —with auditable records, incentives tied to accuracy and a real right to a human— algorithmic administration can be fairer and faster than the bureaucracy we knew: benefits resolved in hours, diagnoses without waiting lists, less opaque discretion. That is the horizon worth fighting for the rules now. Because the question in the headline answers itself: the day it is your turn, you will want a human to hear you. The job is to make sure that, by then, it is still a right and not a privilege.