Filed under — EU AI Act · Article 13 · Transparency · High-Risk AI · Compliance · Instructions for Use
Prefer this source on Google →EU AI Act Article 13: Transparency Obligations for High-Risk AI
Article 13 requires high-risk AI systems to be transparent and to come with instructions for use for deployers. Learn the six points those instructions must cover and how to document them.
Free check · no account
Check your own system against what this article describes.
EU AI Act Article 13: Transparency Obligations for High-Risk AI
Updated 2 October 2026 — correction. Earlier versions quoted Article 13 as written for "users" (the adopted text says deployers), gave a list of six information categories that does not match Article 13(3), read "changes" as update notifications, and named accuracy metrics the Article does not name. The guide now follows Article 13(3), points (a) to (f), and no longer refers to a paid audit.
Updated 31 July 2026 — deadline changed. The Digital Omnibus (adopted by the European Parliament on 16 June 2026 and the Council on 29 June 2026) deferred the Annex III high-risk obligations from 2 August 2026 to 2 December 2027. The Article 50 transparency obligations were not deferred and still apply from 2 August 2026. This article has been corrected accordingly. If your AI system is classified as high-risk under the EU AI Act, Article 13 mandates that it must be "sufficiently transparent to enable deployers to interpret a system's output and use it appropriately" (Regulation (EU) 2024/1689, Article 13(1)). This is not a soft recommendation — it's an enforceable obligation, with a statutory maximum of €15 million or 3% of global annual turnover, whichever is higher.
Most companies underestimate Article 13. They assume transparency means "add a disclaimer" or "show confidence scores." In reality, Article 13(3) lists six points of information that the instructions for use must contain at a minimum, each with its own documentation work.
This guide breaks down what Article 13 actually requires, common compliance gaps, and how to build transparency into your high-risk AI system before the 2 December 2027 deadline.
What Article 13 Actually Requires
Article 13 binds the provider. It requires that high-risk AI systems be accompanied by instructions for use that give deployers information that is:
- Concise, complete, correct, and clear — no jargon, no ambiguity
- Relevant, accessible and comprehensible to deployers — tailored to their role and technical literacy
- Sufficient to enable deployers to interpret the output — they must understand what the system is telling them and why
- Sufficient to enable deployers to use the system appropriately — they must understand when to trust the output and when to override it
Article 13(3) sets out, in six points, what the instructions for use must contain at a minimum:
- (a) Identity and contact details of the provider — and of its authorised representative, where there is one
- (b) Characteristics, capabilities, and limitations of performance — including the intended purpose; the level of accuracy, including its metrics, robustness and cybersecurity; known or foreseeable circumstances that may lead to risks; and, where applicable or appropriate, information that helps explain the output, performance for specific groups of persons, and the input and training data
- (c) Changes pre-determined by the provider — changes to the system and its performance fixed at the initial conformity assessment, if any
- (d) Human oversight measures — those referred to in Article 14, including the technical measures that help deployers interpret the output
- (e) Computational and hardware resources, expected lifetime, and maintenance — including the frequency of care measures and software updates
- (f) Logging mechanisms — where relevant, how deployers can collect, store and interpret the logs required by Article 12
The provider must supply this information. The deployer must then use the system in accordance with the instructions for use (Article 26(1)). A system placed on the market without these instructions appears not to meet Article 13.
The Information in Detail
The sections below take the points in the order most teams tackle them. Accuracy and foreseeable risks are both part of point (b), but they get their own sections because that is where most gaps are.
1. Identity and Contact Details of the Provider — point (a)
This is the simplest requirement: deployers must know who built the system and how to contact them.
What auditors look for:
- Provider name, address, and contact email in the instructions for use
- Clear identification of the legal entity responsible for compliance, and of the authorised representative where there is one
Common failure mode: Shipping a system with no provider identification or burying contact details in a 50-page terms-of-service document.
2. Characteristics, Capabilities, and Limitations of Performance — point (b)
Deployers must understand what the system can and cannot do. This includes:
- Intended purpose — what the system is designed for
- Performance characteristics — accuracy, robustness, and, when appropriate, performance for the specific persons or groups the system is meant to be used on
- Known limitations — tasks the system cannot perform reliably
What auditors look for:
- A written specification of intended purpose and out-of-scope use cases
- Performance benchmarks (e.g., "92% accuracy on validation set")
- Documentation of known failure modes (e.g., "performs poorly on handwritten text")
Common failure mode: Providing only marketing claims ("state-of-the-art accuracy") without quantitative performance data or documented limitations.
3. Pre-determined Changes — point (c)
Point (c) asks for the changes to the system and its performance that the provider fixed in advance, at the moment of the initial conformity assessment, if any. This matters most for systems that keep learning: changes pre-determined there and recorded in the technical documentation do not count as a substantial modification, while other substantial modifications require a new conformity assessment (Article 43(4)). Point (e) separately asks for maintenance and care measures, including software updates.
Article 13 does not require update e-mails or release notes. Keeping deployers informed when the system changes is still good practice, and makes points (c) and (e) easier to evidence.
What auditors look for:
- A description of the pre-determined changes, consistent with the technical documentation
- Performance comparison before and after updates
- A maintenance and update schedule for deployers
Common failure mode: Changing a model in ways the instructions for use never described, without checking whether the change is a substantial modification.
4. Level of Accuracy, Robustness, and Cybersecurity — point (b)(ii)
Article 13 requires the instructions to state the level of accuracy, including its metrics, robustness and cybersecurity referred to in Article 15, against which the system has been tested and validated, and any known circumstances that may affect it. It does not name the metrics. Depending on the task, they might be:
- Accuracy — precision, recall, F1, or domain-specific measures
- Robustness — performance under adversarial inputs or distribution shift
- Cybersecurity — resistance to data poisoning, model extraction, or adversarial attacks
What auditors look for:
- Test set performance reports with confidence intervals
- Robustness benchmarks (e.g., performance on out-of-distribution data)
- Cybersecurity audit reports or penetration test results
Common failure mode: Reporting only aggregate accuracy without breaking down performance by demographic group, edge case, or adversarial scenario.
5. Known or Foreseeable Circumstances That May Lead to Risks — point (b)(iii)
Deployers must be warned about situations, in intended use or reasonably foreseeable misuse, that may lead to risks to health, safety or fundamental rights.
What auditors look for:
- A documented list of edge cases and failure modes
- Risk mitigation guidance (e.g., "Do not use this system for medical diagnosis")
- Evidence that the people using the system are trained on these limitations
Common failure mode: Providing no failure mode documentation or assuming deployers will "figure it out."
6. Human Oversight Measures — point (d)
Article 14 mandates human oversight for high-risk AI systems. Article 13 requires that the instructions for use describe those oversight measures, including the technical measures that help deployers interpret the output.
What auditors look for:
- Documentation of the human overseer's role (e.g., "Review all flagged cases before final decision")
- Training materials for human overseers
- Evidence that the system supports oversight (e.g., explainability features, override mechanisms)
Common failure mode: Deploying a fully automated system with no documented human oversight role.
7. Resources, Maintenance, and Logs — points (e) and (f)
The instructions must also state the computational and hardware resources the system needs, its expected lifetime, and the maintenance and care measures that keep it working, including software updates and how often they happen. Where relevant, they must describe how deployers can collect, store and interpret the logs the system records under Article 12.
What auditors look for:
- A hardware and infrastructure specification
- A stated expected lifetime and maintenance schedule
- Instructions for accessing and retaining logs
Common failure mode: Leaving deployers to discover the system's logging and maintenance needs after go-live.
Article 13 Compliance Checklist
| Information (Art. 13(3)) | Documentation Needed | Common Gap |
|---|---|---|
| Provider identity — (a) | Name, address, contact email in the instructions for use | No provider identification |
| Characteristics, capabilities, limitations — (b) | Intended purpose, performance benchmarks, failure modes | Marketing claims without quantitative data |
| Accuracy, robustness, cybersecurity — (b)(ii) | Test set reports, robustness benchmarks, security audits | Aggregate accuracy only, no edge case breakdown |
| Known risks and failure modes — (b)(iii) | Edge case list, risk mitigation guidance | No failure mode documentation |
| Pre-determined changes — (c) | Changes fixed at the initial conformity assessment | Undescribed changes to the model |
| Human oversight measures — (d) | Overseer role, training materials, override mechanisms | No documented oversight role |
| Resources, lifetime, maintenance — (e) | Hardware needs, expected lifetime, update schedule | No maintenance information |
| Logging mechanisms — (f) | How to collect, store and interpret logs | Logs exist but deployers cannot reach them |
How Article 13 Interacts with Other Requirements
Article 13 does not exist in isolation. It intersects with:
- Article 9 (Risk Management) — the risks you identify in Article 9 must be disclosed to deployers under Article 13
- Article 10 (Data Governance) — the data quality metrics you document under Article 10 inform the accuracy disclosures required by Article 13
- Article 14 (Human Oversight) — the oversight measures you design under Article 14 must be explained to deployers under Article 13
- Article 50 (Transparency for Certain AI Systems) — if your system is also subject to Article 50 (e.g., chatbots, emotion recognition), you have additional transparency obligations, and those have applied since 2 August 2026
A complete compliance strategy addresses all of these together, not as isolated checklists.
Concrete Example: Credit Scoring System
Suppose you've built an AI-powered credit scoring system. Under Annex III.5(b), this is a high-risk system. Here's what Article 13 compliance looks like (the company below is invented):
- Provider identity: The instructions for use state "Provided by FinTech Corp, 123 Main St, Dublin, Ireland. Contact: compliance@fintechcorp.eu"
- Characteristics, capabilities, limitations: You document that the system is designed for consumer credit decisions up to €50,000, achieves 89% accuracy on validation data, and performs poorly for applicants with thin credit files (fewer than 3 tradelines).
- Pre-determined changes and maintenance: You state which changes to the model and its performance were fixed at the initial conformity assessment, and the update schedule. When you update the model, you also send deployers release notes showing the new accuracy (91%) and changes in false positive/false negative rates.
- Accuracy, robustness, cybersecurity: You provide a performance report showing precision, recall, and F1 by demographic group, plus robustness testing results showing performance under adversarial inputs (e.g., applicants who deliberately misreport income).
- Known risks: You document that the system may underestimate risk for self-employed applicants and overestimate risk for recent immigrants. You provide guidance: "Manually review all self-employed and recent immigrant applications."
- Human oversight: You document that loan officers must review all applications flagged as "borderline" (score 600–650) and have the authority to override the system's recommendation.
- Resources and logs: You state the infrastructure the system needs, its expected lifetime, and how the lender can retrieve and keep the decision logs.
All of this is packaged into the instructions for use that go to the lender, which passes them to every loan officer who uses the system. When an auditor asks for Article 13 evidence, you hand them this document plus training records showing that loan officers have been trained on it.
What Happens If You Don't Comply
Non-compliance with Article 13 can trigger:
- Administrative fines — a statutory maximum of €15 million or 3% of global turnover under Art. 99(4) — for SMEs, whichever is lower
- Market surveillance actions — national authorities can order you to withdraw your system from the market or suspend its use
- Liability exposure — if a deployer misuses your system because you failed to provide adequate information, you may be liable for resulting harms
The operative date for Annex III high-risk systems is 2 December 2027. A provider placing a high-risk AI system on the EU market from that date will need Article 13 instructions for use ready when it does.
Common Anti-Patterns
These Article 13 anti-patterns come up most often:
- No deployer-facing documentation — the system has no instructions for use explaining its purpose, limitations, or performance
- Marketing claims without quantitative data — the system claims "high accuracy" but provides no test set metrics
- No failure mode documentation — deployers are not warned about edge cases or situations where the system is likely to fail
- No human oversight guidance — deployers are not told what oversight measures they are expected to apply
- Undescribed changes — the system changes in ways the instructions for use never pre-determined, with no maintenance or update information
- No provider identification — deployers do not know who built the system or how to contact them
This article is for informational purposes only and does not constitute legal advice. Consult a qualified EU AI Act attorney for guidance on your specific situation.
Related dispatches