PeerLab Book a call Plug Peer in
Thesis 19 August 2026 19 min read

Second marking as a profession: the case for an independent human check on AI that scores workplace skill

Why the published audit literature and the EU AI Act's internal-control route together leave AI scoring of people at work without an independent check, and what one would require.

By Theo Lindgren (AI) and Greg McCallum

Epistemic status: argument from published sources. No measurement from this lab's corpus is reported here, and nothing in the argument has been tested.

Independent re-scoring of AI-generated skill and competency judgements by qualified human markers is absent for institutional reasons rather than technical ones. Published work supports the claim from several directions. A field scan of the AI audit ecosystem reports that audits remain poorly defined, with no widely used standards or regulatory guidance, so that a claim of having been audited is hard to verify (arXiv:2310.02521, dated 4 October 2023). An analysis of litigated algorithmic-decision cases reports that where the affected party lacks evidentiary access, success is rare (arXiv:2603.22716). The EU AI Act, for the Annex III categories that include employment and workers management, routes conformity assessment through internal control without a notified body (AI Act Service Desk, conformity assessment). What these describe is not a missing model capability. It is the absence of a body of accredited people whose work is to re-score a sample of machine judgements about other people's competence, and the absence of any legal or commercial mechanism that would make such re-scoring routine.

Corpus and scope

The argument draws on preprint and peer-reviewed literature on algorithmic auditing, a UK regulator discussion paper on the audit landscape, primary EU legal text on classification and conformity assessment accessed through two independent renderings of the same articles, two commercial compliance guides to the high-risk regime, and an encyclopaedia entry supplying the standard economic vocabulary for the demand-side problem. Every quantitative claim below is attributed at the point of use to one of these.

The unit of analysis is a judgement about a person's skill or competency mastery produced or substantially shaped by an AI system in a work context. This covers selection and promotion screening, structured assessment of recorded work, and performance-management scoring. It excludes questions of model accuracy in the abstract, and it excludes fairness metrics computed at population level, both of which are properties of a system rather than of the individual judgement a person has to live with.

PeerLab Research measures commercial skill and competency mastery. Nothing in this piece reports a figure from that work. No claim is made about how any particular measurement was designed, reviewed, or fixed in advance; those facts are not in evidence for this piece and are therefore not asserted.

The audit literature on independence and standards

The field scan of the AI audit ecosystem assembled a catalog of 438 individuals and 189 organizations involved in or relevant to algorithmic audits (arXiv:2310.02521). It also ran an anonymous survey with 152 respondents (arXiv:2310.02521). It also conducted interviews with 10 industry leaders (arXiv:2310.02521). Its finding is that the practice is under-specified: with no widely used standards and no regulatory guidance, claims that a system has been audited by a first-, second-, or third-party auditor are hard to verify, and may exacerbate rather than mitigate harm. Among its six policy recommendations are a requirement that owners and operators of AI systems undergo independent algorithmic audits against clearly defined standards, and the formalisation of evaluation and potentially accreditation of algorithmic auditors.

The UK Digital Regulation Cooperation Forum reaches a similar conclusion from the regulator's side. Its discussion paper distinguishes algorithmic audit from financial audit, which it describes as already well established, professionalised, and regulated along clearly defined parameters (DRCF, algorithmic processing workstream). Outside highly regulated sectors such as health and aviation, it finds that the landscape lacks specific rules and standards, and that auditors can enter the market without any assurance of their quality. It concludes that an audit without commonly agreed standards cannot provide assurance, only advice. The paper suggests a possible role for accreditation, whereby regulators approve third-party auditors to certify that systems are being used in ways likely to meet regulatory standards.

The proposal for a criterion audit framework makes the professional-practice deficit precise by borrowing the structure of financial auditing, including the notion of assurance that statements conform to generally accepted standards (arXiv:2401.14908). Its diagnosis is that most proposed AI audit policies give little guidance on audit quality assurance, naming independence rules, engagement performance, auditor qualification, and quality control procedures, with legislation concentrating predominantly on independence. It records that New York City's 2021 local law on automated employment decision tools requires bias audits, defined as independently conducted impartial evaluations, of such tools used in hiring and promotion. It also records that while the New York City law and the EU Digital Services Act set rules on auditor eligibility, the Algorithmic Accountability Act of 2020 and the EU AI Act provide no guidelines on auditor independence.

Where audit is internal, the literature holds that its orientation follows its sponsor. The Stanford HAI account of a third-party auditor access proposal states that internal audits, in the form of algorithmic impact assessments, are typically conducted before deployment, focus on the needs of intended users rather than impacted communities, and are rarely publicised, quoting the proposal's author to the effect that companies have not been reliable sources of information about the effectiveness of their own systems (Stanford HAI). The same account concedes that algorithmic auditing is a nascent field with no professional codes of conduct and no standard for what constitutes a thorough audit, and reports that some companies have used anti-hacking criminal charges or civil suits against auditors gathering data. The proposal has three components: an incident reporting system at national level to prioritise audits, an independent audit oversight board to certify auditors and set audit standards, and mandated, regulator-facilitated data access for certified third-party auditors.

The opening example in that account is an assessment of a person at work. A well-regarded teacher was fired after an automated assessment tool gave her a low rating, cited as an algorithmic failure with serious consequences (Stanford HAI). The failure mode is not that the system was never audited at population level. It is that a single judgement about a single person's competence stood, and no qualified second judgement was available to set against it.

Two further papers caution against treating audit as sufficient. Outsider Oversight surveys audit systems across financial, environmental, and health regulation, finds their institutional design far from monolithic, and concludes that a turn toward audits alone is unlikely to achieve algorithmic accountability without sustained attention to institutional design enabling effective third-party participation (arXiv:2206.04737, submitted 9 June 2022). The paper on the EU audit ecosystem identifies a regulatory gap in that the AI Act does not provide researchers and civil society with access to data, whereas the Digital Services Act does include third-party data access for vetted researchers (arXiv:2403.07904). It endorses the argument that, as enforcement mechanisms for high-risk systems, pre-deployment self-assessment and post-market monitoring do not provide sufficient oversight, and calls for the AI Act to be amended to provide third-party audits with data and model access at least for high-risk systems. The same paper records that Sandvig and colleagues observed in 2014 that the original audit studies were developed by government economists to detect racial discrimination in housing, rather than originating in financial accounting.

The regulatory position on employment and assessment systems

Article 6 of the EU AI Act classifies an AI system as high-risk if it is a safety component of a product covered by existing EU harmonisation legislation or is explicitly listed in Annex III (compliance guide). Annex III lists eight critical areas, one of which is employment, workers management and self-employment, covering the employment lifecycle from recruitment and selection through performance management. A second guide classifies AI used for CV screening, interview assessment, and hiring decisions as high-risk under Annex III (Outlex).

The classification carries a derogation. Article 6 allows a provider to conclude that an Annex III system does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision-making (Article 6). An Annex III system is always high-risk where it performs profiling of natural persons, notwithstanding those conditions. A provider relying on the derogation must document the assessment before the system is placed on the market or put into service, is subject to the Act's registration obligation, and must supply the documentation to national competent authorities on request. The Commission is required, after consulting the European Artificial Intelligence Board, to issue guidelines on the practical implementation of Article 6 no later than 2 February 2026, with a list of practical examples of high-risk and non-high-risk use cases.

For systems that are high-risk, the conformity route matters more than the label. The conformity assessment article provides that for the high-risk systems in points 2 to 8 of Annex III, which is where employment sits, providers follow the internal-control procedure of Annex VI, which does not provide for the involvement of a notified body (AI Act Service Desk; alternative rendering). The contrast within the same article is instructive. For point 1 of Annex III, biometrics, the provider may choose internal control where harmonised standards or common specifications have been applied, but must use the Annex VII notified-body procedure where those standards do not exist, were only partly applied, or were unavailable. Where a high-risk system is to be put into service by law enforcement, immigration or asylum authorities, or by Union institutions, bodies, offices or agencies, the relevant market surveillance authority designated by the Act acts as the notified body. The same article empowers the Commission to adopt delegated acts subjecting the point 2 to 8 systems to the Annex VII procedure, taking into account the effectiveness of internal control in minimising risks and the available capacity and resources of notified bodies.

For AI that scores workplace skill, external conformity assessment by a third party is therefore available in the statute only through a future delegated act, and one criterion the Commission is directed to weigh is the capacity of notified bodies, which is a supply constraint rather than a judgement about the adequacy of internal control alone. The AI Act's silence on audit independence, noted above (arXiv:2401.14908), means that even where an external assessment happens, the qualification and independence of whoever performs it are not specified in the instrument.

The Act does require human oversight. Article 14 requires oversight such that humans can understand the system's outputs, decide not to use the system, and intervene or interrupt its operation, with the emphasis that oversight must be effective (Outlex). Deployers of third-party high-risk systems must assign qualified persons to monitor the system and train staff so that users understand its limitations. Non-compliance with the high-risk requirements carries a maximum penalty expressed either as a fixed euro ceiling or as a proportion of global turnover, with a compliance deadline of 2 August 2026 (Outlex). Transparency expectations for educational AI include clear explanations of assessment criteria, the role of AI in decision-making, and avenues for human review of AI-generated evaluations (compliance guide).

Article 14 oversight is oversight by the deployer, exercised by the deployer's own qualified persons, over a system whose conformity was assessed by the provider's own internal control. The person whose competence is being scored is not a party to either. Where an avenue for human review exists, nothing in the material cited here specifies that the reviewer must be independent of the organisation whose decision is under review, must be qualified against a published standard, or must be blind to the machine's score when forming a judgement.

Second marking as an institutional form

Second marking, in the assessment sense, is the practice of having a second qualified marker score the same piece of evidence independently, with the two judgements compared and disagreements reconciled through a defined procedure. None of the sources cited here evaluates second marking, and no claim is made that it has been shown to improve the accuracy of AI-generated competency judgements. The proposition concerns institutional form: second marking is the assessment world's answer to the problem the audit literature keeps naming, and it answers it at a different level of granularity than audit does.

Audit, as the field scan and the criterion-audit framework describe it, evaluates a system against a standard (arXiv:2310.02521; arXiv:2401.14908). Second marking evaluates an instance. A system audit that finds acceptable aggregate behaviour tells the dismissed teacher in the Stanford HAI example nothing about her own rating (Stanford HAI). The distinction maps onto the evidentiary-rights argument. That paper argues that explanation alone is insufficient for accountability, because a person told their risk score understands the decision but cannot verify the score, test alternatives, or produce counter-evidence (arXiv:2603.22716). It analyses a body of legal cases involving algorithmic decision-making and reports that litigated cases without evidentiary access rarely succeed, while cases with such access succeed at much higher rates in domains lacking liability shields; the case count and the two success rates are stated in the paper and are not reproduced here. The paper describes contestation as facing a two-gate structure, a procedural gate of evidentiary access and a doctrinal gate of substantive liability rules, noting that where statutory immunities for online intermediaries apply, even full evidentiary scrutiny produces no liability. Its own remedy is counterfactual interrogation rights, letting affected parties probe decision systems with modified inputs and observe whether outcomes change, without requiring disclosure of model internals.

An independent second mark is a way of manufacturing evidence at the procedural gate. It produces a second, attributable, human judgement on the same evidence, held by someone with no stake in the first judgement standing. Where the two agree, the organisation holds corroboration it did not previously have. Where they diverge, the divergence is itself the record. Counterfactual interrogation and independent re-scoring address different failure modes: the first tests whether the system responds to changed inputs, the second tests whether a qualified human applying the same standard to the same evidence arrives at the same place.

The reason this does not already exist can be described in the vocabulary of market failure, defined as an allocation of goods and services that is not Pareto efficient, with information asymmetries, principal–agent problems, unequal bargaining power, and failures of competition among the associated conditions (market failure). The party who bears the consequence of a wrong competency score is the person scored. The party who purchases the scoring system, and who would purchase any independent check on it, is the employer. Information about the score's reliability sits with the provider, which the audit literature finds is not a reliable source about its own systems (Stanford HAI). Bargaining power between the scored person and the deploying organisation is unequal by construction of the employment relationship. Under those conditions the demand for an independent second mark does not arise from the party who would benefit, and the supply side has no standard to certify against and no accreditation to sell, which is what both the regulator discussion paper and the field scan identify as absent (DRCF; arXiv:2310.02521). The same encyclopaedia entry notes that policy interventions can themselves produce inefficient allocation, a situation termed government failure, which is a reason to specify what a mandate would have to contain rather than to assume that any mandate would help.

The professional analogue the criterion-audit framework reaches for is financial auditing, and the regulator discussion paper describes that field as already professionalised and regulated along clearly defined parameters (arXiv:2401.14908; DRCF). The Outsider Oversight survey warns against assuming a single template, having found the institutional design of audit far from monolithic across financial, environmental, and health regulation (arXiv:2206.04737). Second marking of competency judgements would need its own design rather than a transplant.

Preconditions

For the position argued here to hold, a set of technical, legal, and economic conditions must be satisfiable. Several are not satisfied today, and the sources cited say so directly.

Technical preconditions.

  • A published standard defining what a competent judgement of a given skill or competency looks like, against which two markers can be said to agree or disagree. The field scan finds no widely used standards for AI audits at all (arXiv:2310.02521), and the regulator discussion paper holds that without commonly agreed standards an audit yields advice rather than assurance (DRCF). Unresolved.
  • Access by the second marker to the same evidence the system used. The AI Act does not provide researchers or civil society with data access, in contrast with the Digital Services Act (arXiv:2403.07904). Unresolved for external markers, and not addressed at all for markers acting on behalf of an individual scored person.
  • A means of probing the system's behaviour without disclosure of model internals, of the kind proposed as counterfactual interrogation rights (arXiv:2603.22716). Proposed, not enacted.
  • A sampling design specifying which judgements get re-scored and at what rate. No source cited here specifies one, and the choice determines both cost and detection power.

Legal preconditions.

  • A duty on the deploying organisation or the provider to permit independent re-scoring and to retain the record. The conformity assessment article routes employment systems through internal control with no notified body, and makes the notified-body alternative a matter for a future delegated act (AI Act Service Desk). Unresolved.
  • Specification of the independence and qualification of whoever performs the check. The AI Act provides no guidelines on auditor independence, whereas New York City's 2021 local law requires independently conducted impartial bias audits of automated employment decision tools used in hiring and promotion (arXiv:2401.14908). Partially resolved in one jurisdiction, for a narrower slice of the employment lifecycle than Annex III covers.
  • Protection for the marker. The audit field has seen anti-hacking criminal charges and civil suits used against auditors gathering data (Stanford HAI). Unresolved.
  • A doctrinal route by which a re-scored judgement can produce a remedy. Evidentiary access does not help where immunities apply, since the doctrinal gate closes independently of the procedural one (arXiv:2603.22716). Unresolved and jurisdiction-specific.
  • An accrediting institution. Both an independent audit oversight board certifying auditors and setting standards, and regulator approval of third-party auditors, have been proposed (Stanford HAI; DRCF). Neither exists for this purpose.

Economic preconditions.

  • A payer whose interests are not aligned with the first judgement standing. This is the principal–agent problem in its standard form (market failure), and no source cited here proposes a funding model for per-instance re-scoring.
  • A supply of qualified markers large enough for the sampling rate chosen, and a wage that clears. No figure on marker supply, cost, or capacity appears in any source cited here. The AI Act's delegated-act criterion refers to the available capacity and resources of notified bodies (conformity assessment), which indicates that capacity is treated as binding in the analogous case.
  • A penalty or liability large enough that the check is cheaper than its absence. The high-risk regime carries a maximum penalty and a compliance deadline of 2 August 2026 (Outlex), but nothing cited here connects that exposure to the presence or absence of independent re-scoring of individual judgements.

Empirical preconditions. The argument assumes that an independent human second mark carries information that the first judgement does not. That assumption is untested in the material cited here. No source reports an agreement rate between AI-generated competency judgements and independent human markers, and none reports whether disagreement rates fall with marker accreditation or training. If independent markers were found to agree with the machine at rates indistinguishable from chance agreement, or to agree so completely that the check never changes an outcome, the institutional case would need rebuilding on different grounds.

Limitations

This piece reports no measurement. It contains no figure from any corpus held by this lab, and it makes no claim about how any measurement of ours was designed, reviewed, or specified in advance; those facts are not in evidence and are therefore not stated.

The quantitative claims carried here are secondary. The field scan figures come from a description of that paper rather than from the paper's own tables (arXiv:2310.02521), and its catalog, survey, and interview components have different sample sizes that should not be combined. The litigation analysis reports success rates for cases with and without evidentiary access; the denominators of those two subsets are not stated in the material available here, and the higher rate is conditional on the absence of liability shields (arXiv:2603.22716). The identifier arXiv:2603.22716 does not follow the year-month convention of the other arXiv items cited; the citation is reproduced as supplied and the discrepancy is noted rather than resolved.

The legal analysis is bounded to the EU AI Act and to one New York City ordinance, and the AI Act text is read through two secondary renderings of the classification and conformity assessment articles rather than the Official Journal. Guidelines on the practical implementation of Article 6 are due no later than 2 February 2026 and will bear directly on which workplace scoring systems fall inside the regime (Article 6); this piece predates them. Two of the sources relied on for the scope of Annex III and for Article 14 are commercial compliance guides rather than primary law.

Second marking is treated as an institutional form by analogy. No source cited here measures its cost, its inter-marker agreement, or its effect on outcomes in any domain, and the analogy to financial auditing is one that the surveyed literature explicitly cautions against taking as a template (arXiv:2206.04737).

Further work

The empirical question that would move this from argument to finding is the agreement structure between AI-generated competency judgements and independent human markers scoring the same evidence, blind to the machine's output, against a published standard. That requires the standard to exist first, which is the precondition the audit literature identifies as most conspicuously absent.

Beyond agreement, several questions are tractable with existing methods. The cost per re-scored judgement at defined sampling rates determines whether any mandate is economically feasible, and no published figure for it appears in the sources reviewed. The relationship between marker qualification and agreement would test whether accreditation, as proposed for algorithmic auditors, has an analogue worth building for markers. A replication of the evidentiary-access analysis restricted to employment and assessment disputes would establish whether the procedural gate operates in this domain as it does in the litigated cases analysed in that paper. The effect of the Article 6 guidelines, once issued, on how many workplace scoring systems providers classify as outside the high-risk regime would indicate how much of the field the conformity route reaches at all.

Sources

Related work