Can public-service AI look stable while changing the evidence? Similar accuracy hid different labels
Eight GPT-4o configurations scored between 68.2% and 70.3% on 1,200 parliamentary questions, yet disagreed on 24.3% of them. The unreviewed UK study shows why public bodies need to test individual outputs, review routing and conclusions—not rely on one accuracy figure.
The short version
- What happened
- The experiment sampled 1,200 of 9,061 human-coded UK Prime Minister's Questions across 21 policy topics and compared eight GPT-4o configurations. Accuracy stayed in a narrow 68.2%–70.3% range, while 291 of 1,199 complete cases received different labels across configurations.
- Who it affects
- Public servants could receive different classifications or summaries from systems with nearly identical average scores, changing which issues appear prominent.
- What remains unproven
- Unreviewed single-author preprint using one classification task and one model family.
Keep in mind
Unreviewed single-author preprint using one classification task and one model family.
The human-coded labels are interpretive references and may be ambiguous or wrong for some questions.
Raw outputs from the original 17-configuration experiment were lost; eight configurations retained parsed labels and scores.
Five unchanged repeats occurred about five months later on an unpinned model alias, so time and configuration effects cannot be fully separated.
The score-extraction rule was not a validated probability of final-answer correctness, and no human reviewers were observed.
The document analysis used AI-assisted coding checked by one author, with no inter-coder agreement rate, and did not measure compliance or enforcement.
No public-service decision, allocation, individual harm or benefit was measured.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: a stable headline score concealed different accounts of the same public record
An unreviewed study finds that eight plausible ways of configuring GPT-4o produced almost the same aggregate accuracy when classifying UK Prime Minister's Questions, while disagreeing on nearly a quarter of individual questions. Accuracy against the human-coded reference ranged only from 68.2% to 70.3%. Among the 1,199 questions with complete results across all eight configurations, however, 291—24.3%—received different topic labels. For a public body using those labels to summarise political attention, the choice of prompt format, examples or output procedure could therefore change the evidence presented to decision-makers even when the dashboard score barely moves.
The result does not show that GPT-4o caused a bad public decision. The task was retrospective classification of parliamentary questions from 1997–2008, not a live benefit, policing, education or health service. Human-coded labels are also judgements rather than unquestionable truth. The study's practical contribution is narrower and useful: assurance should examine repeatability, error structure, case selection, conclusions and the operation of review queues. A single average can hide whether different system versions agree on the cases that matter and whether any agreement is actually correct.[1]
How the 1,200-question experiment was constructed
The UK Comparative Agendas Project contains 9,061 unique Prime Minister's Questions after deduplication, each assigned to one of 21 policy topics by human coders. The classes are highly imbalanced: Government Operations accounts for 17.4% of the corpus, while Culture, Foreign Trade and Technology each contribute less than 1%. The author drew a fixed-seed stratified sample of 1,200 questions—60 from each of 19 sufficiently large topics, 45 Foreign Trade questions and all 15 Culture questions—and applied weights when estimating full-period topic shares. That design gives small categories visibility without pretending the sample has the original class balance.
Eight retained configurations varied the amount of category description, few-shot examples, requested reasoning, historical and speaker context, and whether output was structured or free text. The anchor used GPT-4o, full descriptions, an explicit coding rule, six examples, JSON output and temperature zero. It was a defensible default, not a configuration selected to maximise accuracy. Labels and extracted scores survived from eight of 17 configurations run in April 2026; the raw per-call outputs were lost when the cloud environment was deleted. One free-text response was missing, leaving 1,199 complete cases for comparisons across configurations.[1]
The distinction that matters
What the headline accuracy did not show
All figures come from one exploratory GPT-4o classification study; they are not public-service outcome rates.
Headline accuracy
68.2%–70.3%
Narrow range across eight configurations
Changed labels
291 of 1,199
24.3% differed across configurations
Shared disagreement
191 questions
All eight differed from the human reference
Review workload
0.9%–41.0%
Same score threshold, different queues
Repeatability, agreement and selective evaluation produced three different warnings
The anchor configuration was run five more times over all 1,200 questions on 1 September, with the prompt, question order and settings unchanged. Pairs of runs disagreed on 2.78% of labels on average, and 65 questions—5.42%—changed at least once. Temperature zero did not make the service deterministic. Because the repeats occurred about five months after the original experiment and used a provider alias rather than a dated model snapshot, the comparison cannot separate run-to-run variation from a model update. That limitation is operationally relevant: public bodies often consume provider aliases whose behaviour can change without a new local deployment.
Agreement also failed as a shortcut to correctness. For 191 questions, all eight configurations chose the same topic but differed from the human reference. The reference could be wrong or the question could legitimately span topics, but eight outputs from one model family are not eight independent witnesses. A separate filtering exercise retained only the 908 questions on which every configuration agreed. The anchor's reported accuracy then rose from 69.6% to 79.0% without any prediction improving. The higher number described an easier selected subset and excluded half of the Public Lands questions, changing what the evaluation represented.[1]
Human review depends on the routing rule, not merely the presence of a reviewer
The author tested a hypothetical rule that would send outputs with an extracted score below 0.99 to human review. Under the anchor, 36.0% were flagged and the remaining outputs had 83.6% accuracy against the reference. Under the chain-of-thought configuration, only 0.9% were flagged and the unflagged accuracy was 68.7%. Across all eight configurations, the same nominal threshold produced review queues covering between 0.9% and 41.0% of cases. A public manager told that a human checks uncertain answers would therefore still need to ask how uncertainty is calculated and whether the queue captures the errors that matter.
The score is not a calibrated probability that the final label is correct. It is the highest probability among valid topic codes at a selected token, and when reasoning or free text comes before the answer that token can precede the final classification. No human reviewer was observed, so the experiment does not estimate whether review would repair mistakes, introduce new ones or overwhelm available staff. Its value is to expose a missing denominator: oversight capacity must be compared with the number, type and consequence of outputs actually routed to people, not recorded as a generic checkbox.[1]
What the 37 UK and EU documents require—and what they leave to organisations
The documentary analysis covered 19 UK-level and 18 EU institutional documents assessed as current by 26 October 2026. The author coded passages about decision-making, delivery, coordination and accountability, with each divided into design and resources. AI assisted the coding, but every entry was checked against the source. The analysis found that both UK and EU frameworks expect testing and oversight. UK guidance explicitly mentions repeated prompts, inconsistent outputs, trained reviewers and override mechanisms; the EU AI Act requires appropriate testing, consistency for intended purpose and effective oversight for relevant high-risk systems.
Those expectations do not consistently prescribe a common repeatability statistic, reference standard, acceptance threshold or report of how individual classifications alter substantive conclusions. UK material places strong emphasis on public bodies checking their own use. EU rules allocate substantial duties to providers while retaining deployer obligations, and a public authority can itself become a provider if it puts a system into service under its name. The paper does not determine compliance, enforcement or institutional capacity. It shows why delegated technical choices need a documented justification that affected people, auditors and decision-makers can scrutinise.[1]
What public bodies can do now—and what evidence would change the assessment
A proportionate assurance case should begin with the claim the system is supposed to support, the source population and the consequence of a wrong or unstable output. Teams can repeat unchanged inputs, compare credible configurations or versions, inspect shared errors against an independently produced reference, report exclusions, and test whether a human-review queue is both accurate and operationally manageable. They should preserve model and prompt versions and trace which reports or decisions relied on them. These steps do not prove a system is safe, but they make the basis for bounded reliance more visible than a single accuracy figure.
Confidence would rise with pinned models tested in the same period, multiple model families, preserved raw outputs and independent coding of the policy documents. The human reference should be audited for ambiguity, and real reviewers should be observed so error correction, workload and escalation can be measured. Most importantly, prospective studies should test whether configuration differences change public-service recommendations, resource allocation or outcomes. Until then, the defensible conclusion is that similar aggregate performance can conceal material differences in AI-produced evidence; it is not that every public-service AI system fails or that one threshold fits every use.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring public services: compare the finding and its evidence limits.
What must Australian governments prove before using AI?
A national framework published on 16 October asks federal, state and territory bodies to document purpose, assess impacts, govern data, test systems, monitor performance and keep humans accountable. It is a shared assurance baseline—not proof that every agency has implemented those controls.
Read the next reportExplore the full source and evidence record
What this means for people
- Public servants could receive different classifications or summaries from systems with nearly identical average scores, changing which issues appear prominent.
- People affected by a service need a route to challenge the evidence and learn who accepted the remaining uncertainty, not only confirmation that a human was nominally involved.
- Review teams need measured queue size and error coverage; a threshold that flags 0.9% of one pipeline and 41.0% of another is not the same operational control.
Global context
The experiment is British and the policy comparison is UK–EU, but the assurance problem travels with general-purpose models used by governments worldwide. Legal duties, administrative capacity and review rights differ by jurisdiction. The reusable lesson is methodological: document what the system is allowed to support, test stability and consequence-relevant errors locally, and make the justification contestable rather than transferring the study's thresholds to another service.
What the evidence does not yet show
- Unreviewed single-author preprint using one classification task and one model family.
- The human-coded labels are interpretive references and may be ambiguous or wrong for some questions.
- Raw outputs from the original 17-configuration experiment were lost; eight configurations retained parsed labels and scores.
- Five unchanged repeats occurred about five months later on an unpinned model alias, so time and configuration effects cannot be fully separated.
- The score-extraction rule was not a validated probability of final-answer correctness, and no human reviewers were observed.
- The document analysis used AI-assisted coding checked by one author, with no inter-coder agreement rate, and did not measure compliance or enforcement.
- No public-service decision, allocation, individual harm or benefit was measured.
What to watch next
- Replication with pinned versions, multiple providers and preserved raw outputs.
- Independent review of the human reference labels and documentary coding.
- Prospective tests of real public-service reviewers, queue capacity and correction rates.
- Reporting standards that connect model variation to changed conclusions and affected people.
- Evidence that UK and EU assurance requirements are applied effectively in practice.
Living evidence record
Impact record IAI-065V3EH
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
Whether aggregate accuracy is sufficient to justify public-sector reliance on AI-produced classifications and whether current UK and EU governance documents address the observed failure modes
Topics: Public services · AI assurance · Government evidence · Model stability · Human review · EU AI Act
Regions: United Kingdom · European Union · Europe
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.