An AI decides asylum, sentences, or parole unreviewed

Where things stand (13 Oct 2026)

I found no publication since the 16 Sep 2026 opening date in which an OECD government says an AI system makes asylum, sentencing or parole decisions without someone reviewing each case. Governments using AI in these areas all say the opposite:

• United Kingdom. The Home Office says asylum decisions "are made by trained decision-makers, not artificial intelligence." Its GPT-based case-summary and policy-search tools only assist caseworkers (Independent, ~9 Oct 2026; Open Rights Group, Oct 2026).
• Netherlands. The immigration service (IND) says staff may not let its two AI assistants make asylum decisions (NL Times/NRC, 18 Sep 2026).
• Germany. The cabinet's July 2026 bill on AI in migration administration says individual decisions remain caseworkers' responsibility (Tech Policy Press, 1 Sep 2026).
• Canada. The immigration department's AI strategy says its tools "do not refuse" applications.
• United States. The Justice Department AI inventory (page updated 7 Oct 2026) lists PATTERN, the federal prisoner recidivism-risk tool, only as producing risk scores. The question explicitly excludes risk scoring and eligibility screening.
• Tasmania, Australia. The parole-board case involved AI-drafted material that humans then used, and it is a state matter, not a national one.

Reference class

Over roughly ten years and 38 OECD governments, I count zero qualifying statements. The consistent pattern is explicit denial. Several structural barriers keep the chance low:

• EU data-protection rules restricting fully automated decisions, and the EU AI Act's high-risk status for AI in migration and justice;
• the separate EU rules on automated decisions in law enforcement, and due-process requirements;
• the question's rule that even rubber-stamp case-by-case approval counts as human review.

Paths to YES

A. A lenient reading of rule-based automation (about 4–5%, concentrated early).

• Norway's immigration directorate (UDI) already makes fully automated decisions on Ukrainians' applications for temporary collective protection. Norway processes these as asylum applications. Digi.no (8 Dec 2025) reported that about 90% of these decisions run through the automation, which UDI describes as a rule engine plus software robots. A 2018 Dagens Perspektiv article describes it as fixed objective checks with no discretion.
• A Norwegian government consultation paper of 17 Sep 2026 describes collective protection as a simplified asylum procedure. It proposes fully automated decisions, but only for a new work-based permit.
• New Norwegian publications will keep appearing, such as UDI's 2026 annual report, likely around Feb–Mar 2027. A resolver could count the rule engine as "logic-based inference" under the inclusive OECD definition.
• I think that reading is unlikely. The question excludes "fixed administrative workflows" and requires an explicit statement, and these documents never call the system AI.
• A similar edge case: the US immigration courts' planned AI "automated acceptance or rejection of case filings", which is still pre-deployment.

B. A genuine machine-learning or language-model system stated to decide cases with no prior human review (about 5% by 2031).

• The hazard rises over time with capability and adoption.
• Signs pointing that way: the US administration has rebranded AI as "Super Intelligence" (SI) on its agency AI pages and is adopting it aggressively. Norway's immigration IT modernisation programme runs to 2031, considers full automation, and has a Sept 2026 consultation on using applicants' data to build AI systems. Asylum backlogs remain large.

Paths to NO

Governments keep a human formally in the loop, or never say otherwise explicitly. That covers nearly every scenario I can construct.

Strongest case against my estimate

Capability gains by 2028–31, combined with the question's broad terms, could push the chance higher, perhaps to 15–20%. Those terms include the inclusive OECD definition, a limited category of cases counting, live pilots counting, and unlawful operation counting. I give this some weight, but every explicit statement so far points the other way.

Method

I modelled quarterly hazards for the two paths and combined them, giving about 1.9% by end-2026, 4.4% by end-2027 and 9.7% by end-2031. The calculation was done in Python. The Justice Department inventory came from justice.gov/media/1426076/dl (sheet "Reportable AI Use Cases"). The Norwegian consultation PDF of 17 Sep 2026 came via a municipal public-records mirror, because regjeringen.no was blocked.

Question. By the last day of each quarter, has the national government of any OECD country officially stated that an AI system decides asylum claims, criminal sentences, or parole without a human reviewing each decision?

Definitions.

• OECD country: a full OECD member on the publication date of the qualifying statement, as recorded on the OECD’s Members and partners page or in its accession announcements; candidates and partner countries do not count.
• National government: the country's central or federal executive, including its ministries, departments, agencies, and head of government speaking officially. A national-government statement about a qualifying system operated at a subnational level within that country counts.
• AI system: a machine-based system that infers from inputs how to generate outputs using machine learning or logic- or knowledge-based inference, consistent with the OECD's November 2023 definition and March 2024 Explanatory memorandum on the updated OECD definition of an AI system. Ordinary database retrieval, fixed administrative workflows, or arithmetic implementation of a human-selected outcome alone do not count. Systems from any developer are eligible.
• Decides: selects an outcome for an actual individual case that becomes an operative official determination without requiring a human to examine and approve that outcome first. At least one of the three decision classes suffices, and a limited category of cases within one class suffices. Deployment must have begun; authorization or future plans alone do not count.
• Asylum claims: decisions granting or refusing asylum or refugee protection, including dismissing a claim as inadmissible. Other visa decisions, border screening, interview summaries, and case prioritization do not count.
• Criminal sentences: imposition or alteration of a criminal punishment, including its type, duration, or amount. Risk scoring, sentencing recommendations, bail, and administrative calculation of a release date from a human-imposed sentence do not count.
• Parole: granting, refusing, or revoking conditional early release from a criminal custodial sentence. Eligibility screening or a recommendation to a parole board does not count.
• Without a human reviewing each decision: the stated operational process permits individual AI-selected outcomes to take effect without prior case-specific human examination of the proposed outcome. Sampling, exception-only review, later audits, optional appeals, or later judicial review do not disqualify it. Human data entry, setting general policy, and purely clerical transmission do not constitute decision review. Required case-by-case human approval counts as review, even if brief or routinely agreeing with the system, unless the government expressly states that approval occurs without examination of individual outcomes.

Resolves YES if a qualifying national-government publication explicitly states that an operational system meets all of the definitions above. A live pilot counts if its decisions have real official effect. The statement may describe unlawful or unauthorized operation; legal validity is not required. It must establish both that the AI determines outcomes and that prior individual human review is not universal; silence about review is insufficient.

Resolves NO / Exclusions. No qualifying statement by the horizon; proposed deployments; simulations; advisory systems; and systems whose outcomes all receive prior human review do not qualify. Claims that officials rely excessively on AI, without confirming the absence of individual review, do not qualify. For example, the UK government's Asylum Case Summarisation description, which specifies human control over all asylum decisions, does not itself qualify.

Resolving source. Official publications on the national executive's government domains, including agency reports, AI-use inventories, press releases, and official transcripts; or formal government answers or written ministerial statements published in the national legislature's official record. This is a narrower source requirement than Public confirmation (A2): independent audit-office reports, court findings, legislative findings, vendor statements, and press reporting cannot independently resolve YES unless the national executive explicitly adopts their qualifying findings in an eligible publication. Apply the Resolution window and Resolution date (A3) conventions.