La vérificatrice générale a conclu que le chatbot IA de l'ARC, au coût de 18 millions de dollars, fournissait des renseignements fiscaux erronés
Le chatbot fiscal de 18 M$ du Canada n'a répondu correctement qu'à deux questions sur six, selon la vérificatrice générale.
Field: narrative
Reports (sources available for citation)
[1] CBC News (2025-10-21) — In scathing report, AG finds CRA call centres are slow to answer and often inaccurate Supports: Auditor General found CRA chatbot Charlie provided accurate answers only a third of the time [2] 980 CJME (2025-10-25) — CRA must fix human responses before pursuing AI, experts say Supports: AI experts warned that AI will replicate or amplify errors from inaccurate human agents [3] iPhone in Canada (2025-12-12) — Auditor General Slams Ottawa's $18 Million CRA Chatbot 'Charlie' Supports: CRA spent $18 million on chatbot Charlie; AG found it gave wrong answers 66% of the time [4] Unpublished (2025-12-12) — The CRA spent $18M on 'Charlie,' a new tax information chatbot that is wrong most of the time Supports: Charlie met a 70% accuracy threshold in internal testing; upgraded to generative AI in November 2025
Text to add citations to
The Canada Revenue Agency announced Charlie, an AI-powered chatbot, in February 2020 and launched it in March 2020 to answer taxpayer questions about tax filing, benefits, and CRA services. By the time of the Auditor General's review in October 2025, the system had processed over 18 million questions at a cost of approximately $18 million (iPhone in Canada, 2025; Unpublished, 2025).
The Auditor General's October 2025 report found significant accuracy problems. When tested with common taxpayer questions, the chatbot answered only two out of six questions correctly (CBC News, 2025; iPhone in Canada, 2025). The Auditor General also noted that other publicly available AI tools answered five out of six of the same questions correctly.
With 18 million queries processed, the error rate identified by the Auditor General raises concerns about the accuracy of tax information provided to Canadians through the system. Taxpayers who relied on Charlie's responses may have received incorrect information about filing requirements, deadlines, or eligibility for benefits. The CRA chatbot carries the implicit authority of the federal tax agency, and users have no straightforward way to know when the chatbot's answer is wrong.
The Auditor General also noted that Charlie had met only a 70% accuracy threshold in CRA's own internal testing — meaning inaccurate responses 30% of the time (Unpublished, 2025). The Secretary of State (Canada Revenue Agency), Wayne Long, responded publicly: "We've got a lot of room for improvement. We know it, we accept the report and we're going to do better." The CRA acknowledged it could not confirm real-world accuracy without a comprehensive review of every interaction. In November 2025, the CRA upgraded Charlie to a generative AI version and reported pre-release testing showed approximately 90% accuracy, though the agency acknowledged it could not confirm real-world accuracy (Unpublished, 2025).
Matérialisé à partir de
Préjudices
Lors des tests de la vérificatrice générale avec six questions courantes de contribuables, le chatbot Charlie de l'ARC n'a répondu correctement qu'à deux d'entre elles, fournissant des informations erronées sur les obligations fiscales, les exigences de production et les procédures de l'ARC. Le chatbot avait traité un estimé de 18 millions de requêtes au cours de sa durée de vie.
L'ARC a dépensé 18 millions de dollars pour un chatbot d'IA sans tests d'exactitude adéquats ni surveillance continue de la qualité, le déployant comme source officielle d'information gouvernementale portant l'autorité implicite de l'agence fiscale fédérale.
Preuves
4 rapports
- In scathing report, AG finds CRA call centres are slow to answer and often inaccurate Source principale
Auditor General found CRA chatbot Charlie provided accurate answers only a third of the time
-
AI experts warned that AI will replicate or amplify errors from inaccurate human agents
-
CRA spent $18 million on chatbot Charlie; AG found it gave wrong answers 66% of the time
-
Charlie met a 70% accuracy threshold in internal testing; upgraded to generative AI in November 2025
Détails de la fiche
Réponses et résultats
Upgraded Charlie to a generative AI version following the Auditor General's findings; reported pre-release testing showed approximately 90% accuracy but acknowledged it could not confirm real-world accuracy
Évaluation éditoriale évalué
L'autorité fiscale fédérale a dépensé 18 millions de dollars pour un chatbot d'IA que la vérificatrice générale a jugé incapable de répondre correctement à des questions fiscales de base (iPhone in Canada, 2025; Unpublished, 2025). Le chatbot a traité plus de 18 millions de requêtes, soulevant des préoccupations quant à l'exactitude des renseignements fiscaux fournis aux Canadiens par le système (CBC News, 2025; 980 CJME, 2025).
Entités impliquées
Systèmes d'IA impliqués
AI-powered chatbot 'Charlie' deployed to answer taxpayer questions about tax filing, benefits, and CRA services; processed over 18 million questions and was found by the Auditor General to answer only two out of six test questions correctly
Fiches connexes
Taxonomieévalué
AIID : Incident #1310
Historique des modifications
| Version | Date | Modification |
|---|---|---|
| v1 | 8 mars 2026 | Initial publication |
| v2 | 11 mars 2026 | Neutrality and factuality review: corrected AG report date from December 2025 to October 2025 (report tabled October 21, 2025); fixed launch date (announced February 2020, launched March 2020); replaced inferred wrong-answer topics with AG's actual comparison to other AI tools; corrected 70% accuracy framing (AG cited this as evidence of inadequacy, not CRA's defense); clarified Secretary of State pushback was about broader report findings; removed three fabricated policy recommendation attributions (AG's formal recommendations addressed call centre operations, not chatbot-specific measures). |