{"id":"0698c956-86ff-428a-add8-0017767393ac","arxiv_id":"2608.07452","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper rejects technical explainability as the standard for consumer credit decisions and argues for a legal justification standard it calls 'justifiable AI'.","lead":"This paper argues that technical AI explanations are not enough for consumers in credit scoring, and that the law should demand a legal justification instead. It combines a review of EU law with an argument for 'justifiable AI' as a better standard to protect consumers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that technical XAI cannot provide actionable knowledge rests on an untested empirical premise, and the paper's own proposed alternative (LLM-generated justification) is not yet operationalized; a head-to-head consumer study is needed to support the 'only path' claim.","rationale":"I read the paper as making a normative-legal argument with an empirical hinge: technical XAI explanations fail consumers because they do not provide actionable knowledge, and only a broad, justification-based reading of the right to explanation can deliver that knowledge. The legal analysis is competently assembled: the paper correctly notes the CCD's lower applicability threshold, the SCHUFA judgment, the EDPB's procedural guidelines, and the art. 86 AIA derogation for credit scoring. Those elements give the paper independent support for the narrow legal point that the content of the right is unsettled and that courts have leaned toward global, high-level disclosure. The philosophical distinction between context of discovery and context of justification is also a legitimate framework, and the paper does not claim to have implemented the proposed LLM-based justification pipeline. The most load-bearing weakness is empirical, and it is the same one the reader identified: the step from 'current XAI outputs are developer-oriented' to 'technical explanations are inherently incapable of empowering consumers' is not established. The paper's own evidence includes the author's prior studies with lawyers and tax professionals, plus anecdote; it does not test consumers in a credit-scoring setting, which is the very setting the conclusion generalizes to. Moreover, the conclusion in Section V that justifications can be generated by LLMs from low-level XAI output means the distinction is not between technical and non-technical explanation but between presentation formats; therefore the superiority of the proposed format is exactly what needs empirical demonstration. A controlled user study comparing plain-language technical explanations with the proposed justification letter would settle whether the central claim holds. Because this concern is substantial but not decisive enough to reject the paper's contribution, the reader's CONDITIONAL verdict remains appropriate; no adjustment is needed.","tokens_in":14612,"tokens_out":3748,"duration_ms":38930,"concrete_test":"Run a pre-registered between-subjects experiment with N≈300 consumers who receive a mock credit denial generated by the same underlying model (e.g., a logistic regression or a black-box model with SHAP values). Randomize three explanation conditions: (A) current technical XAI output (feature weights / SHAP values), (B) a plain-language translation of the technical explanation, and (C) the paper's proposed justifiability-style letter (data categories, why pertinent, procedural context, contestation rights). Measure pre- and post-explanation ability to state the actual grounds, identify a possible error or discrimination, and draft a substantive appeal; also measure perceived comprehension. If (C) does not significantly outperform (B) on those measures, the asserted superiority of justifiability over technical XAI is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is in Section IV: from 'contemporary XAI outputs are alien to laypeople' to the conclusion in Section V that 'technical explanations alone constitute regulatory fiction' and that justifiability 'is the only path to genuine transparency.' This inference requires two unestablished claims: (i) the purpose of the right to explanation is to give consumers actionable knowledge for contestation, as opposed to transparency, auditability, or institutional accountability; and (ii) legal-format justifications actually supply that knowledge while technical XAI (including plain-language translations of SHAP/LIME output) cannot. The paper's evidence for (ii) is thin: it cites the author's own lawyer/tax-domain studies and anecdote, not a controlled comparison with consumers in credit decisions. There is also an internal tension: Section V concedes justifications can be produced by LLMs fed with 'information obtained from low-level explanation-generation systems,' so the operative comparison is not technical-versus-legal content but presentation format. The superiority claim is therefore exactly the empirical claim that is not tested. The paper's own limitation statement acknowledges that the LLM justification layer is future work. If a well-designed consumer-facing legal justification does not increase consumers' ability to identify and contest erroneous or discriminatory grounds relative to a plain-language technical explanation, the claimed superiority of justifiability would be unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that technical XAI explanations are insufficient in business-to-consumer contexts such as credit scoring, and that a broader interpretation of the right to explanation as a legal justification ('justifiable AI') is the only way to provide consumers with actionable knowledge for contestation. The paper reviews relevant EU law, identifying art. 18(8) CCD as the central provision after the art. 86(3) AIA derogation, and draws on distinctions from legal theory between discovery and justification. It concludes that technical explanations alone constitute 'regulatory fiction' and proposes that justifiability, possibly generated by LLMs with expert rules, should be the standard.","tokens_in":14838,"tokens_out":3579,"duration_ms":37035,"significance":"The paper addresses a timely and important question: whether the EU's right to explanation is satisfied by technical XAI or requires a context-aware, procedurally grounded justification. It correctly identifies the key legal provisions (art. 86 AIA and its derogation, art. 18(8) CCD) and the relevant case law, and it offers a useful conceptual distinction between explanation and justification. The credit-scoring example is well chosen and keeps the argument focused. The paper does not provide machine-checked proofs or a new empirical dataset, but it advances a clear, falsifiable normative thesis and acknowledges the need for future empirical work on LLM-generated justifications.","major_comments":[{"comment":"The central claim that justifiable AI is 'the only path to genuine transparency' (Section V) is not supported by the evidence presented. The paper relies on studies involving lawyers and tax-domain users (footnotes 57 and 58) and on anecdotal experience, but it does not provide a head-to-head comparison between legal-format justifications and plain-language technical explanations with consumers deciding about credit. Because the paper explicitly concedes that justifications can be produced by LLMs fed with 'information obtained from low-level explanation-generation systems' (Section V), the operative difference appears to be presentation format, not the underlying content. Without a controlled consumer study, the superiority claim is a hypothesis, not an established conclusion.","section":"Section IV and Section V"},{"comment":"The paper asserts that the purpose of the right to explanation is to provide actionable knowledge for contestation, but this normative premise is not defended against competing interpretations. The legal texts (GDPR, CCD, AIA) also support transparency, auditability, and institutional accountability as purposes, and the paper does not engage with authors who favor these views. This premise is load-bearing: if the right is primarily about auditability, technical XAI may be sufficient, and the argument for justifiability weakens. The paper should explicitly address and rebut alternative readings, beyond labeling some interpretations 'creative'.","section":"Section III"},{"comment":"There is an internal tension about whether technical XAI can ever provide actionable knowledge. Section II describes 'actionable recourse explainers' that tell the explainee what action to take and notes that explanations can provide 'recourse (actionable knowledge)'. Section IV, however, states that technical explanations are 'alien' and that even source code or weights would 'still be insufficient to present end user with an actionable knowledge'. The paper should clarify whether the claim is that technical XAI is insufficient in practice due to design and implementation, or that it is impossible in principle. The current wording overstates the case and makes the 'regulatory fiction' rhetoric vulnerable to counterexamples.","section":"Section II and Section IV"},{"comment":"The paper begins by saying that justifiability is 'hypothesized' as a superior standard (Section I) but concludes with the unqualified statement that it 'is the only path' (Section V). This escalation is not justified by the arguments in between. The paper should either soften the conclusion to a conditional claim or provide sufficient evidence for the strong claim. As written, the mismatch between the hypothesis language and the final categorical assertion weakens the paper's credibility.","section":"Section I and Section V"}],"minor_comments":[{"comment":"The phrase 'safeguard the creditors rights' appears to be a typo for 'consumers' rights' (or 'debtors' rights'), which would align with the paper's argument about consumer protection.","section":"Abstract"},{"comment":"'introductory remaks' should be 'introductory remarks'.","section":"Section I"},{"comment":"The text contains a typo 'GPPR' where 'GDPR' is intended.","section":"Section III"},{"comment":"The phrase 'a capite ad calcem' is not translated; a brief gloss (e.g., 'from head to heel') would help readers unfamiliar with Latin.","section":"Section IV"},{"comment":"The Busuioc, Curtin, and Almada article appears twice in the reference list under identical titles; one entry should be removed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's reliance on the author's own prior empirical work is not inherently disqualifying, but the referee report should encourage the authors to present those findings more transparently and to seek independent replication. The paper may be a better fit for a law-and-technology or AI-and-society venue than for a technical AI journal, but it does make a contribution to the policy debate. The main concern is the overclaim in the conclusion, which is fixable in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a useful legal synthesis: it argues that the right to explanation in EU consumer credit law should be read as a right to justification, not as a duty to hand over technical XAI output. The author identifies the right provisions - art. 18(8) CCD as the operative rule, with art. 86(3) AIA's derogation - and applies the discovery/justification distinction to credit scoring more concretely than prior work. That application is genuinely new, and the legal analysis is largely accurate.\n\nWhat the paper does well: it takes the 'right to explanation' debate out of the abstract and anchors it in a specific regulatory scheme. The conclusion that consumers need actionable, contestable reasons rather than SHAP values or LIME masks is defensible and matches the EDPB's procedural reading. The author is also honest about limits: the LLM-based justification layer is explicitly future work.\n\nThe soft spots are real but not fatal. First, the empirical premise that technical XAI cannot provide actionable knowledge rests on a small set of studies, including the author's own tax-domain work and a personal anecdote. There is no controlled comparison with consumers in a credit decision. Second, the paper concedes that LLM-generated justifications would be fed with 'information obtained from low-level explanation-generation systems.' That means the operative contrast is not technical content versus legal content; it is presentation format. The claim that justifiability is 'the only path to genuine transparency' is therefore stronger than the evidence supports. Third, the paper does not seriously engage with the possibility that plain-language technical explanations could be designed for laypeople; it assumes current XAI outputs are the only technical option.\n\nNone of that sinks the central normative argument. The broad interpretation of the right to explanation is well argued, and the policy recommendation - that compliance should focus on narrative justification with procedural safeguards - stands on independent legal grounds. The overclaim is in the packaging, not in the core.\n\nWho should read it: people working on EU AI regulation, consumer credit, and XAI governance. It would benefit from a head-to-head consumer study before the 'only path' claim is pushed, but as a doctrinal contribution it deserves a serious referee.\n\nI would send it out.","headline":"A solid legal synthesis arguing for a broad right to justification in EU credit scoring, but the 'only path' claim outruns the thin empirical evidence.","tokens_in":15381,"tokens_out":2149,"would_cite":true,"duration_ms":20797,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that technical explainability fails consumers in credit scoring; only a broad right to explanation as legal justification—justifiable AI—can make the right operative.","keywords":["Explainable AI","XAI","credit scoring","right to explanation","justifiable AI","EU consumer credit law","automated decision-making","regulatory fiction"],"falsifier":"A controlled study comparing how well credit applicants can formulate a challenge after receiving either a feature-weight explanation of the model's decision or a written procedural justification for the same rejection: if applicants given the technical explanation contest decisions just as effectively as those given the justification, the paper's central claim is refuted.","tokens_in":14388,"feed_emoji":"⚖️","tokens_out":9803,"duration_ms":90098,"temperature":0.7,"pith_summary":"This paper argues that the technical explanations produced by explainable-artificial-intelligence tools do not, and cannot, fulfil the legal right to explanation for consumers whose credit applications are decided by machine-learning systems. In a business-to-consumer setting such as credit scoring, the author contends, presenting model weights, formulas, or saliency maps to laypeople delivers no actionable knowledge and amounts to 'regulatory fiction'—formal compliance without substantive protection. Drawing on EU law and on empirical studies of how lawyers and laypeople actually receive explanation outputs, the paper proposes that the right to explanation should be read broadly, as a requirement of legal justification grounded in procedural, legal, and social norms. The author names this standard 'justifiable AI' and argues it is the only one that lets consumers understand and contest automated credit decisions. The credit-scoring example is offered as the first concrete application of this justifiability standard.","feed_headline":"Credit scoring needs justified AI, not explainable AI","feed_subtitle":"Technical explanations leave consumers unable to contest decisions; only legal justification empowers them.","key_machinery":"The load-bearing distinction is between the context of discovery and the context of justification, borrowed from philosophy of science and legal realism. The paper defines 'justifiable AI' as the standard under which a decision is presented not through model internals but through a persuasive, legally grounded argumentative chain—like a court judgment that gives reasons rather than a trace of the judge's mental processes. The machinery doing the legal work is Art. 18(8) of the 2023 Consumer Credit Directive, read together with Art. 86 of the AI Act and Art. 22 GDPR, which together require a clear, comprehensible explanation of the logic, significance, and effects of automated creditworthiness assessment; the author argues these provisions are only satisfied when the explanation gives the consumer actionable grounds for contestation. On the technical side, the machinery includes the empirical finding that explanation-generation methods are built for developers and are unintelligible to lay users, which is what turns the legal requirement into a fiction.","core_discovery":"The central claim is that explainability and justifiability are different things, and the law has been relying on the wrong one. Explainability addresses the context of discovery—the internal, often mathematical process by which a model produces an output—whereas justifiability addresses the context of justification, the externally grounded reasons, rooted in legal, procedural, and social norms, that make a decision defensible to the person it affects. In business-to-consumer credit scoring, the author argues, handing consumers technical explanations such as feature weights or saliency maps neither matches how humans justify decisions nor gives consumers the knowledge they need to exercise their rights, so it constitutes regulatory fiction. A broad interpretation of the right to explanation—encompassing a legal justification that sets out the categories of data used, why they are pertinent, how any profile was built, and how the profile bears on this particular decision—is therefore the only reading that makes the right operative. The paper answers its three research questions by concluding that current explanation methods are developer-oriented and fail consumers, that the right to explanation must be interpreted broadly in the existing legal framework, and that justifiability should replace technical explainability as the standard.","pith_inferences":["The same justifiability standard would naturally extend beyond credit scoring to any high-stakes consumer decision made by AI, such as hiring, insurance pricing, or benefit eligibility, shifting the regulatory demand from model transparency to contestable reasons.","If regulators adopt this standard, the trade-secrecy conflict softens: firms can protect model internals while disclosing the procedural rationale, so the legal battle moves from opening the black box to judging whether the stated reasons are adequate.","A testable extension follows: studies comparing consumer contestation after technical explanations versus procedural justifications should show that only the latter improves a consumer's ability to challenge an adverse decision, a prediction future work could measure directly.","The discovery/justification split implies a division of labour in which technical explainability remains valuable for developers, auditors, and supervisors, while a separate justification layer becomes the only legally relevant interface with consumers."],"forward_implications":["Lenders would satisfy their explanation duties under Art. 18(8) CCD by supplying procedural justifications—categories of data, why they are pertinent, how a profile is built, and its role in the specific decision—rather than exposing model weights or code.","Regulators and courts would treat developer-oriented explainability outputs as non-compliant in business-to-consumer contexts, since they do not confer actionable knowledge on consumers.","The derogation in Art. 86(3) of the AI Act means the CCD's broader explanation right governs credit scoring, so recent case law on automated decisions supports an expansive reading of that right.","Consumers' right to present their point of view and contest the decision (Art. 18(8)(c) CCD) becomes the test of adequacy: an explanation is legally sufficient only if it equips the consumer to challenge the outcome.","Generating justifications with large language models, constrained by a logical rule layer encoding expert knowledge, is presented as the paper's concrete route to operationalizing justifiability."],"supporting_citations":[{"why":"Supplies the idea that transparency requirements can be satisfied formally while lacking substantive effect, the 'regulatory fiction' the paper argues against.","marker":"n 5"},{"why":"Grounds the legal concern by showing that opaque decisions can deprive persons of rights of defence under EU law.","marker":"n 6"},{"why":"Supplies empirical evidence that legal professionals find technical explanation outputs alien and unusable in practice.","marker":"n 21"},{"why":"Supplies a study showing lawyers struggle to make sense of LIME- and SHAP-style explanations presented to them.","marker":"n 23"},{"why":"Provides the distinction between explanation and justification and the categories of understanding, contestability, and recourse that structure the argument.","marker":"n 24"},{"why":"Supplies survey evidence that banks use explainability mainly for internal model improvement rather than for communicating with consumers.","marker":"n 46"},{"why":"Supports the claim that current explanation-generation methods are aimed at technical staff, not end users.","marker":"n 55"},{"why":"Supplies the recent judicial expansion of automated decision-making that supports a broad reading of the right to explanation.","marker":"n 38"},{"why":"Establishes the Consumer Credit Directive's art. 18(8) explanation requirement as the central legal provision for credit scoring.","marker":"n 36"}],"fun_headline_variants":["Justifiability beats explainability for credit scoring","Credit scoring: legal justification over technical explainability","Explainability fails consumers; justifiability empowers them","Rethink AI in credit: justifiable, not just explainable","For credit decisions, justifiability is the operative standard"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that the legal point of an explanation is to give the consumer usable grounds to challenge the decision, and that no technical description of how the model works can provide those grounds; if the right to explanation is really about transparency or auditability, or if technical explanations ever become genuinely usable by ordinary people, the case for justifiability loses its footing.","fun_headline_variants_meta":{"raw":{"variants":["Justifiability beats explainability for credit scoring","Credit scoring: legal justification over technical explainability","Explainability fails consumers; justifiability empowers them","Rethink AI in credit: justifiable, not just explainable","For credit decisions, justifiability is the operative standard"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000134,"raw_usage":{"total_tokens":1096,"prompt_tokens":858,"completion_tokens":238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":173}},"tokens_in":474,"tokens_out":238,"duration_ms":2855,"temperature":1.0,"reasoning_tokens":173,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:24:32.374842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled study comparing how well credit applicants can formulate a challenge after receiving either a feature-weight explanation of the model's decision or a written procedural justification for the same rejection: if applicants given the technical explanation contest decisions just as effectively as those given the justification, the paper's central claim is refuted.","supporting_citations":[],"review_version":1}