Accountability by document: model cards, audits and oversight duties

A claim about a model cannot be checked by looking at the interface. Whether a classifier holds up on a subgroup lives in the pipeline, and the screen shows only the consequences. The field's response was documentary rather than declaratory. It produced model cards, data cards and audits that write a system's claims down for another person to check, and then a law that makes part of the practice a duty. This seventh of the ten units takes how accountability is written down. Artefacts, procedures and enforceable duties build it, rather than principles. One part of the construction is missing. The document addresses the supply chain of providers, deployers and auditors. The person a system decides about meets only a thin and contested information duty.

Convergence in wording, divergence in meaning

Jobin, Ienca and Vayena (2019) mapped the published principles and guidelines for ethical artificial intelligence and identified 84 documents. They found global convergence around five principles, transparency, justice and fairness, non-maleficence, responsibility and privacy, with substantive divergence over how each is interpreted. A shared vocabulary is not a shared specification.

Mittelstadt (2019) explains why the comparison with medical ethics fails. AI development lacks common aims and fiduciary duties, a professional history and norms, proven methods for translating principles into practice, and robust legal accountability mechanisms. That is the same corpus read for a different conclusion. Munn (2023) goes further and calls the principles useless. They are contested or incoherent and hard to apply, isolated in an industry that largely ignores ethics, and toothless because they carry no consequences. The field's answer was not a better principle but a document.

A card makes a performance claim state its conditions

Mitchell and colleagues (2019) proposed the model card, a short document released with a trained model. It records intended use, evaluation conditions, performance broken down by demographic and intersectional groups, and known limitations. The template forces a performance claim to state the conditions it holds under.

Pushkarna, Zaldivar and Kjartansson (2022) applied the logic to datasets. A data card summarises sources, collection and annotation methods, evaluation methods and intended use. The authors report lessons from over twenty deployed cards. Arnold and colleagues (2019) proposed FactSheets, modelled on supplier's declarations of conformity. A FactSheet records purpose, performance, safety and provenance for examination by consumers. All three make a claim checkable by writing it down, and none reports a measured effect on what a reader then does. Disclosure itself has since been scored from outside. Bommasani and colleagues (2023) define 100 indicators of transparency and score ten leading developers against them, finding a mean of 37 out of 100 and a top score of 54, with none disclosing its model's downstream impact. The index was repeated six months later, when fourteen developers scored a mean of 58, most of the rise from firms reporting information that had not been public, while copyright status, data access, data labour and downstream impact stayed opaque.

The document serves the supply chain, not the person decided about

The commitment leaves open whose reading the document is meant to survive. Article 13 of the Artificial Intelligence Act (2024) requires a high-risk system to carry instructions for use. Those instructions must inform the deployer of its intended purpose and conditions of use, a business-to-business disclosure that matches the audiences the papers write for. Arnold and colleagues (2019) address buying organisations, and Mitchell and colleagues (2019) address those releasing models.

A different route reaches the person the system decides about. Articles 13 to 15 of the General Data Protection Regulation give data subjects two rights: to be informed about automated decision making, and to receive meaningful information about the logic involved. Its scope is disputed. Wachter, Mittelstadt and Floridi (2017) read it as a limited right to be informed. Selbst and Powles (2017) read the same text as supporting a right to explanation. Either way the apparatus serves the supply chain, leaving the affected person at the end of an information duty two leading readings describe differently. A card published as an image or in legal English is one an older reader, a screen reader and a reader with low literacy cannot use.

Audit is a procedure, not a document

A card on its own is a static claim. Raji and colleagues (2020) make documentation a procedure, setting out an end-to-end framework for internal algorithmic auditing across the development lifecycle. The framework has five stages: scoping, mapping, artefact collection, testing and reflection. Each stage yields documents that form the report, which draws on the organisation's values to assess whether the decisions taken fit them. External audits of deployed systems are hard to trace to a cause, while an internal audit can find harms before deployment.

Mokander and Floridi (2021) name the conditions under which the procedure becomes accountability rather than a form. It must be continuous and constructive, approach ethical alignment from a system perspective, and align with public policies. An audit compares a system against its declared values, and the declaration belongs to the organisation audited.

The law turns oversight into a duty

The Artificial Intelligence Act (2024) is the strongest enforcement in this literature, and it is deliberately narrow. It prohibits certain practices, imposes requirements on high-risk systems and sets transparency rules for general-purpose models. For high-risk systems, Article 9 requires a documented, lifecycle-wide risk management system. Article 14 requires the system to be designed with appropriate human-machine interface tools, so that natural persons can effectively oversee it. Article 99 attaches penalties of up to EUR 35 million or 7 per cent of worldwide annual turnover for the prohibited practices, and up to EUR 15 million or 3 per cent for the remaining obligations. For that class of system, interface design is part of compliance.

Two readings qualify the force of the duty. Veale and Borgesius (2021), analysing the draft Act, find that differentiation by risk level makes sense. They also find that several provisions carry surprising legal implications and others are likely to be ineffective, singling out the enforcement regime. Tabassi (2023) offers the voluntary alternative, the NIST AI Risk Management Framework. It is non-prescriptive and law-agnostic, and it is organised around governing, mapping, measuring and managing risk. A binding duty and a voluntary framework can describe the same practice, but only one can be breached.

Contested question: is a person entitled to be told?

The live disagreement is whether the affected person holds a legal right to an explanation: two readings of the same text. Wachter, Mittelstadt and Floridi (2017) conclude that no such right exists in the General Data Protection Regulation as the commentary claims, Articles 13 to 15 mandating meaningful but properly limited information about the logic involved, a right to be informed, while Article 22, from which the claimed right is said to stem, is ambiguous. Selbst and Powles (2017) reach the opposite conclusion from the same articles. They accept that no provision is labelled a right to explanation, and hold that rights to meaningful information about the logic involved are a right to explanation, to be interpreted functionally so that a data subject can exercise it.

What would settle the question is a decision with authority. It could be an enforcement action or court ruling requiring a controller to explain a specific decision's logic to the person affected. It could be a documented case in which a data subject obtained that information and acted on it.

What to do while the dispute is open

Both sides read the same articles, so the disagreement is interpretive and neither rests on a measurement of what a data subject receives. The procedural literature is the counterweight: Raji and colleagues (2020) supply an audit framework with five stages and Mokander and Floridi (2021) set the conditions under which an audit becomes accountability, but neither reports a case in which a person obtained an explanation or acted on one. The strongest objection to this unit's thesis is that documents may be the wrong instrument. Munn (2023) calls AI principles toothless because they carry no consequences, and the same charge can be made of a card no regulator reads. While the dispute is open, a designer should treat the affected person as a reader of the system's claims, and record for each automated decision what that person was told and what they could do with it.

Where the argument is heading

The direction is towards duties attached to a role in a value chain rather than to a technology. Hacker, Engel and Mauer (2023) distinguish developers, deployers, users and recipients of output and propose three layers of obligation: minimum standards, high-risk obligations, and value-chain collaboration, arguing that regulation should target concrete high-risk applications rather than the pre-trained model itself. Mokander and Floridi (2021) push auditing towards a continuous process, and Raji and colleagues (2020) supply its stages.

The sources this unit rests on

• Arnold, M., Bellamy, R. K. E., Hind, M., et al. (2019). FactSheets: Increasing trust in AI services through supplier's declarations of conformity. IBM Journal of Research and Development. doi:10.1147/jrd.2019.2942288 Proposes FactSheets, modelled on supplier's declarations of conformity, recording an AI service's purpose, performance, safety, security and provenance for consumers. It offers the template without evaluating how readers use it.
• European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/1689/oj The binding instrument the unit measures other proposals against: requirements on high-risk systems and transparency rules for general-purpose models. Article 14 requires high-risk systems designed for effective human oversight, and Article 99 sets penalties of up to EUR 35 million or 7 per cent of turnover.
• Hacker, P., Engel, A. and Mauer, M. (2023). Regulating ChatGPT and other Large Generative AI Models. 2023 ACM Conference on Fairness, Accountability, and Transparency. doi:10.1145/3593013.3594067 Works through how the law should reach large generative models and proposes three layers of obligation: minimum standards, high-risk obligations and value-chain collaboration. Regulation should target concrete high-risk applications rather than the pre-trained model.
• Jobin, A., Ienca, M. and Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence. doi:10.1038/s42256-019-0088-2 Mapped the published principles and guidelines on ethical AI and identified 84 documents, finding convergence around five principles with substantive divergence over interpretation. It is the evidence that principles converge in wording but not in meaning.
• Mitchell, M., Wu, S., Zaldivar, A., et al. (2019). Model Cards for Model Reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency. doi:10.1145/3287560.3287596 Proposes model cards, short documents released with a trained model that record intended use, evaluation conditions, subgroup performance and known limitations. It shows the artefact without measuring its effect on deployment decisions.
• Mittelstadt, B. (2019). Principles alone cannot guarantee ethical AI. Nature Machine Intelligence. doi:10.1038/s42256-019-0114-4 Argues that convergence among at least 84 public-private initiatives on principles resembling medical ethics does not deliver ethical practice, because AI development lacks medicine's aims, professional norms, translation methods and robust accountability mechanisms. Principles are therefore not the mechanism; enforcement is.
• Mokander, J. and Floridi, L. (2021). Ethics-Based Auditing to Develop Trustworthy AI. Minds and Machines. doi:10.1007/s11023-021-09557-8 Argues that ethics-based auditing can improve decision quality and enable law-making, and sets out its conditions: a continuous and constructive process, a system perspective, and alignment with public policies. It is a proposal rather than an evaluation.
• Munn, L. (2023). The uselessness of AI ethics. AI and Ethics. doi:10.1007/s43681-022-00209-w Argues that AI principles are useless as formulated: contested or incoherent, hard to apply, isolated in an industry that ignores ethics, and toothless because they carry no consequences. It is the sharpest statement of the unit's scepticism.
• Pushkarna, M., Zaldivar, A. and Kjartansson, O. (2022). Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI. 2022 ACM Conference on Fairness, Accountability, and Transparency. doi:10.1145/3531146.3533231 Extends documentation from models to datasets with data cards, structured summaries of sources, collection and annotation methods, evaluation methods and intended use, with lessons from over twenty deployed cards.
• Raji, I. D., Smart, A., White, R. N., et al. (2020). Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. doi:10.1145/3351095.3372873 Introduces an end-to-end framework for internal algorithmic auditing across the development lifecycle, with five stages, scoping, mapping, artefact collection, testing and reflection, each yielding an audit report. Internal audits can catch harms before deployment.
• Selbst, A. D. and Powles, J. (2017). Meaningful information and the right to explanation. International Data Privacy Law. doi:10.1093/idpl/ipx022 Argues that the General Data Protection Regulation does support a right to explanation, Articles 13 to 15 giving data subjects rights to meaningful information about the logic involved, interpreted functionally. It is the position this unit sets against Wachter and colleagues.
• Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, National Institute of Standards and Technology. doi:10.6028/nist.ai.100-1 Offers a voluntary, non-sector-specific resource for organisations using AI systems, organised around functions for governing, mapping, measuring and managing risk. It is the voluntary standard the unit contrasts with the AI Act's binding duties.
• Veale, M. and Borgesius, F. Z. (2021). Demystifying the Draft EU Artificial Intelligence Act. Computer Law Review International. doi:10.31235/osf.io/38p5f Analyses the draft AI Act through EU product safety regulation: differentiation by risk level makes sense, while several provisions carry surprising legal implications and others may be ineffective. It is the critical reading that qualifies the Act's enforcement claim.
• Wachter, S., Mittelstadt, B. and Floridi, L. (2017). Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation. International Data Privacy Law. doi:10.1093/idpl/ipx005 Concludes that no right to explanation exists in the General Data Protection Regulation as the commentary claims, Articles 13 to 15 providing only a right to be informed about automated decision making and the logic involved. It supplies half of the contested question.
• Bommasani, R., Klyman, K., Longpre, S., et al. (2023). The Foundation Model Transparency Index. arXiv preprint arXiv:2310.12941. doi:10.48550/arXiv.2310.12941 The authors define 100 indicators of transparency and score ten leading developers against them, finding a mean of 37 out of 100 and a top score of 54. A second edition, the 2024 Foundation Model Transparency Index (arXiv:2407.12929), scored fourteen developers at a mean of 58 after inviting them to report, leaving copyright status, data access, data labour and downstream impact opaque.