Briefings Blog

Frontier AI models tested against the International Monetary Fund's March 2026 Review of the Adequacy of the Fund's Precautionary Balances produced six confident, citable answers that the regulator's own primary text directly contradicts, an evaluation by the RLB Specialist Panel has found.

For Financial Advisers who use AI tools on Fund financial-governance and Fund-strength tracking matters, the failures concentrate on the specific numerical and lexicon parameters that are most commonly carried into working deliverables: the precautionary balances minimum floor, the FY2024 surcharge-payer baseline, the half-year PB level reported in the Q2FY26 Quarterly Financial Report, the strength of the Board's early-review signal, and the named geopolitical theatre the Board flagged as a source of intensifying downside risk.

The most material of the six findings concerns the minimum floor for precautionary balances. The IMF's own text on the March 2026 Review records that Directors generally agreed to retain the current floor at SDR 20 billion. The frontier AI model under test committed to a floor of SDR 15 billion across multiple deliverable registers, including a board briefing memo for an EM finance ministry client and a historical-trajectory section for a campaign report by a development-finance NGO.

The SDR 5 billion divergence is a verifiable parameter that supervisors, counterparties, and internal sign-off reviewers will check against the source; if it enters a deliverable for Financial Advisers use it will surface under review.

A second cluster of findings concerns the October 2024 charges and surcharge reform. The IMF Press Release records that the number of countries subject to surcharges in fiscal year 2026 is expected to fall from 20 to 13. The AI's policy-brief draft inflated the FY2024 baseline to 22, producing a 22-to-13 trajectory that diverges from the regulator's 20-to-13. A third finding concerns the IMF Board lexicon: the regulator's text records that 'a few Directors' saw merit in considering an early review of charges and the surcharge policy.

The AI's legal-and-policy advisory elevated the position to 'a number of Directors', changing the strength of the signal a sovereign-debt practitioner would read out for multi-year debt-service planning. A fourth finding adds Ukraine to the Board's named geopolitical theatre on intensifying downside risk; the regulator's text names only the Middle East.

A fifth and sixth finding record divergences on the half-year precautionary balances level reported in the Q2FY26 Quarterly Financial Report and on the pre-March-2024 floor value in a campaign-report historical trajectory. The Q2FY26 Schedule 2 records the October 31, 2025 PB level at SDR 26,782 million; the AI committed to approximately SDR 26.5 billion. Every finding in this audit is bound to verbatim primary source text recorded by the International Monetary Fund. The RLB Specialist Panel offers International Monetary Fund and any other named entity a permanent right of reply on every finding.

For Financial Advisers, the operational signal is that AI-assisted research on the March 2026 PB Review and the related October 2024 surcharge reform cannot be relied on for the floor value, the FY surcharge-payer baseline, the Board-lexicon strength of a Board signal, the named geopolitical theatre, or the half-year PB level, without verification against the IMF Press Release, the Q2FY26 Quarterly Financial Report, and Press Release 24/376 directly.

Frontier AI models tested against the International Monetary Fund's March 2026 Review of the Adequacy of the Fund's Precautionary Balances produced six confident, citable answers that the regulator's own primary text directly contradicts, an evaluation by the RLB Specialist Panel has found.

For Accountants (CA/PA) who use AI tools on Fund financial-governance and Fund-strength tracking matters, the failures concentrate on the specific numerical and lexicon parameters that are most commonly carried into working deliverables: the precautionary balances minimum floor, the FY2024 surcharge-payer baseline, the half-year PB level reported in the Q2FY26 Quarterly Financial Report, the strength of the Board's early-review signal, and the named geopolitical theatre the Board flagged as a source of intensifying downside risk.

The most material of the six findings concerns the minimum floor for precautionary balances. The IMF's own text on the March 2026 Review records that Directors generally agreed to retain the current floor at SDR 20 billion. The frontier AI model under test committed to a floor of SDR 15 billion across multiple deliverable registers, including a board briefing memo for an EM finance ministry client and a historical-trajectory section for a campaign report by a development-finance NGO.

The SDR 5 billion divergence is a verifiable parameter that supervisors, counterparties, and internal sign-off reviewers will check against the source; if it enters a deliverable for Accountants (CA/PA) use it will surface under review.

A second cluster of findings concerns the October 2024 charges and surcharge reform. The IMF Press Release records that the number of countries subject to surcharges in fiscal year 2026 is expected to fall from 20 to 13. The AI's policy-brief draft inflated the FY2024 baseline to 22, producing a 22-to-13 trajectory that diverges from the regulator's 20-to-13. A third finding concerns the IMF Board lexicon: the regulator's text records that 'a few Directors' saw merit in considering an early review of charges and the surcharge policy.

The AI's legal-and-policy advisory elevated the position to 'a number of Directors', changing the strength of the signal a sovereign-debt practitioner would read out for multi-year debt-service planning. A fourth finding adds Ukraine to the Board's named geopolitical theatre on intensifying downside risk; the regulator's text names only the Middle East.

A fifth and sixth finding record divergences on the half-year precautionary balances level reported in the Q2FY26 Quarterly Financial Report and on the pre-March-2024 floor value in a campaign-report historical trajectory. The Q2FY26 Schedule 2 records the October 31, 2025 PB level at SDR 26,782 million; the AI committed to approximately SDR 26.5 billion. Every finding in this audit is bound to verbatim primary source text recorded by the International Monetary Fund. The RLB Specialist Panel offers International Monetary Fund and any other named entity a permanent right of reply on every finding.

For Accountants (CA/PA), the operational signal is that AI-assisted research on the March 2026 PB Review and the related October 2024 surcharge reform cannot be relied on for the floor value, the FY surcharge-payer baseline, the Board-lexicon strength of a Board signal, the named geopolitical theatre, or the half-year PB level, without verification against the IMF Press Release, the Q2FY26 Quarterly Financial Report, and Press Release 24/376 directly.

Frontier AI models tested against the International Monetary Fund's March 2026 Review of the Adequacy of the Fund's Precautionary Balances produced six confident, citable answers that the regulator's own primary text directly contradicts, an evaluation by the RLB Specialist Panel has found.

For Lawyers who use AI tools on Fund financial-governance and Fund-strength tracking matters, the failures concentrate on the specific numerical and lexicon parameters that are most commonly carried into working deliverables: the precautionary balances minimum floor, the FY2024 surcharge-payer baseline, the half-year PB level reported in the Q2FY26 Quarterly Financial Report, the strength of the Board's early-review signal, and the named geopolitical theatre the Board flagged as a source of intensifying downside risk.

The most material of the six findings concerns the minimum floor for precautionary balances. The IMF's own text on the March 2026 Review records that Directors generally agreed to retain the current floor at SDR 20 billion. The frontier AI model under test committed to a floor of SDR 15 billion across multiple deliverable registers, including a board briefing memo for an EM finance ministry client and a historical-trajectory section for a campaign report by a development-finance NGO.

The SDR 5 billion divergence is a verifiable parameter that supervisors, counterparties, and internal sign-off reviewers will check against the source; if it enters a deliverable for Lawyers use it will surface under review.

A second cluster of findings concerns the October 2024 charges and surcharge reform. The IMF Press Release records that the number of countries subject to surcharges in fiscal year 2026 is expected to fall from 20 to 13. The AI's policy-brief draft inflated the FY2024 baseline to 22, producing a 22-to-13 trajectory that diverges from the regulator's 20-to-13. A third finding concerns the IMF Board lexicon: the regulator's text records that 'a few Directors' saw merit in considering an early review of charges and the surcharge policy.

The AI's legal-and-policy advisory elevated the position to 'a number of Directors', changing the strength of the signal a sovereign-debt practitioner would read out for multi-year debt-service planning. A fourth finding adds Ukraine to the Board's named geopolitical theatre on intensifying downside risk; the regulator's text names only the Middle East.

A fifth and sixth finding record divergences on the half-year precautionary balances level reported in the Q2FY26 Quarterly Financial Report and on the pre-March-2024 floor value in a campaign-report historical trajectory. The Q2FY26 Schedule 2 records the October 31, 2025 PB level at SDR 26,782 million; the AI committed to approximately SDR 26.5 billion. Every finding in this audit is bound to verbatim primary source text recorded by the International Monetary Fund. The RLB Specialist Panel offers International Monetary Fund and any other named entity a permanent right of reply on every finding.

For Lawyers, the operational signal is that AI-assisted research on the March 2026 PB Review and the related October 2024 surcharge reform cannot be relied on for the floor value, the FY surcharge-payer baseline, the Board-lexicon strength of a Board signal, the named geopolitical theatre, or the half-year PB level, without verification against the IMF Press Release, the Q2FY26 Quarterly Financial Report, and Press Release 24/376 directly.

Frontier AI models tested against the International Monetary Fund's March 2026 Review of the Adequacy of the Fund's Precautionary Balances produced six confident, citable answers that the regulator's own primary text directly contradicts, an evaluation by the RLB Specialist Panel has found. The findings concentrate on three failure mechanisms that should matter to AI lab teams fielding frontier models into IMF-adjacent advisory deployments: cycle-trajectory drift on biennial-review parameters, single-value inflation under generation pressure on reform-adjacent baselines, and attribution drift on politically sensitive named items.

The most material of the six findings concerns the precautionary balances minimum floor. The IMF's own text on the March 2026 Review records that Directors generally agreed to retain the current floor at SDR 20 billion. Claude Opus 4.7, the AI subject under test, committed to a floor of SDR 15 billion across deliverable registers including a board briefing memo for an EM finance ministry client and a historical-trajectory section for a development-finance NGO campaign report.

The SDR 5 billion divergence is the kind of single, verifiable parameter that supervisors, counterparties, and internal QC reviewers will check against the source; for an AI lab fielding a model into this advisory context, the failure shape is a confident commitment in a deliverable register where the model had access to the regulator's text at query time.

Two further findings cluster on the October 2024 charges and surcharge reform recorded in IMF Press Release 24/376. The Press Release records that the number of countries subject to surcharges in fiscal year 2026 is expected to fall from 20 to 13. The model's policy-brief draft inflated the FY2024 baseline to 22, producing a 22-to-13 trajectory that diverges from the regulator's 20-to-13.

A separate finding records that the model elevated the IMF Board lexicon characterisation of the early-surcharge-review signal from 'a few Directors' (the regulator's text) to 'a number of Directors', changing the strength of the signal a sovereign-debt practitioner would read out for multi-year debt-service planning.

Three further findings record divergences on the pre-March-2024 floor trajectory, on a politically sensitive named geopolitical theatre (the model added Ukraine to the Board's named theatre on intensifying downside risk, where the regulator's text names only the Middle East), and on the half-year precautionary balances level reported in the Q2FY26 Quarterly Financial Report (the model committed to approximately SDR 26.5 billion against the regulator's recorded SDR 26,782 million). Every finding in this audit is bound to verbatim primary source text recorded by the International Monetary Fund.

The RLB Specialist Panel offers International Monetary Fund and any other named entity a permanent right of reply on every finding. For AI lab teams, the operational signal is that frontier models deployed into IMF-adjacent advisory contexts need retrieval-anchored verification on biennial-review parameters, reform-adjacent baselines, IMF Board lexicon characterisations, named theatre attributions, and quarterly-report financial figures.

Misstated recommendation count, fabricated consultation respondents, and scoping drift on the CPMI October 2024 final report on FPS interlinking governance. Two frontier AI models tested by the RegLeg Brief Specialist Panel produced confident, citable answers across six distinct questions on the October 2024 CPMI final report on linking fast payment systems across borders that the regulator's own primary text in publication d223 directly contradicts. The audit covers the count and scope of the report's oversight recommendations, the named list of public consultation respondents in Annex 1, and the distinction between the interim publication d219 and the final publication d223.

For Operations teams at Retail Banking firms working CPMI FPS interlinking governance matters, the failure pattern is operationally consequential. The audit tested six questions designed by the RLB Specialist Panel to mirror how legal, compliance, risk, operations, and board-secretariat teams at payment institutions, hub entities, and banks participating in fast payment system interlinking arrangements actually use AI on this practice area: operating manuals for retail customer access to interlinked FPS, service-level agreement drafting against the d223 recommendation set, customer-experience design for cross-border fast payments, and operational-policy notes on retail-bank participation in interlinking arrangements.

Each question is bound to verbatim regulator-issued primary substrate. Across the six findings the AI subjects committed, in board-style and analyst-style deliverables, to approximately ten oversight recommendations (against the seven set out in d223 Section 5.2), to consultation-respondent lists of fifteen to twenty named organisations (against the seven specific respondents recorded in d223 Annex 1), and to a scoping treatment that places the single access point gateway model inside the report's recommendations (against d223 Section 2.2, which records that the single access point is not the focus of the report).

The findings are operationally consequential for legal teams, compliance teams, risk teams, operations teams, and board secretariats at payment institutions, banks, hub entities, and FPS operators whose practice touches the October 2024 CPMI final report. A board-level briefing memo that records the report as setting out approximately ten recommendations conflates the interim d219's ten considerations with the final d223's seven recommendations and embeds that confusion into the board's understanding of the oversight regime. A legal opinion that scopes a single access point arrangement inside the d223 recommendation set creates a falsifiable regulatory-interpretation error in a partner-level deliverable.

An analyst report that lists fifteen to twenty named consultation respondents (including fabricated organisation names) attributes positions and counts to stakeholders that the regulator's Annex 1 does not record.

The audit's six findings are published with immutable RLB Citation IDs. Representative entries include RLB-H-INT-BIS-CPMI-CPMI-FPS-INTERLINKING-GOVERNANCE-2024-Q001-Opus47, RLB-H-INT-BIS-CPMI-CPMI-FPS-INTERLINKING-GOVERNANCE-2024-Q005-Opus47, RLB-H-INT-BIS-CPMI-CPMI-FPS-INTERLINKING-GOVERNANCE-2024-Q006-Opus47. The full audit is published at the CPMI FPS Interlinking Governance 2024 hub on RegLegBrief.com.