Learn Law / Research & discussion

Living discussion paper · Version 1.0

The Legal Challenges for AI Learning Apps

Can an AI Tutor Replace a Human One?

Research cut-off: 2026-09-24 · Published 2026-09-26 · About 30 minutes

Link to or cite this paper

Durmah, 'The Legal Challenges for AI Learning Apps: Can an AI Tutor Replace a Human One?', version 1.0, research cut-off 2026-09-24.

https://www.durmah.ai/articles/ai-tutors-legal-challenges

Use this permanent link in LinkedIn summaries, newsletters and references.

Scope: UK copyright and data-protection questions, with particular attention to England and Wales for litigation, educational services and academic-integrity law. Differences affecting Scotland and Northern Ireland are identified where material. This is a selective doctrinal and policy analysis, not an exhaustive digest of every AI judgment or an empirical evaluation of tutoring products.

Interest disclosure: Durmah is developing a commercial legal-learning service and has an interest in this debate. Its experience supplies questions, not evidence that its preferred legal interpretation is correct. The product descriptions below do not imply provider involvement or endorsement.

Abstract

AI learning products increasingly offer functions associated with tutors: guided questioning, explanations, revision exercises, feedback and spoken interaction. Whether those functions can replace a human tutor is partly an empirical question about learning and partly an institutional question about responsibility. It is not answered by calling software educational. This paper examines the separate permissions and obligations involved when a student uses a commercial AI tool, particularly with university or publisher material. It distinguishes private study from commercial exploitation, inference from training, source citation from legal authority, and lawful copying from authorised assessment assistance. Recent UK decisions illuminate particular issues without supplying a general licence for AI tutoring. We argue for a function-specific approach: identify each actor, copying and processing operation, purpose, output and foreseeable reliance. We set out the strongest objections to that approach and the evidence needed to test it. The proposed conclusion is that some tutoring functions can be supported or substituted, while any claim to replace the human role requires much stronger evidence about learning, accountability and safeguarding. This is a living discussion paper intended for correction and reasoned disagreement.

Keywords: AI tutors; private study; fair dealing; copyright; legal education; generative AI; academic integrity; UK GDPR; consumer protection.

1. Start with the learner, but examine the transaction

A law student is struggling with consideration. At midnight, an AI assistant can offer an example, ask why a promise is enforceable, and adapt its explanation. That is a useful educational possibility. Now change the facts. The student uploads an entire commercial textbook, asks the assistant to rewrite an assessed answer, or relies on a fabricated judgment. The conversational interface looks the same, but the legal and educational issues change.

Our research question is therefore narrower than whether AI is good or bad for learning: under what conditions can a commercial AI service support an individual student's learning without assuming rights, competence or institutional permission it does not possess? A second question follows: if a provider markets the service as replacing a tutor, what responsibilities does that representation invite?

We use five separate inquiries throughout:

1. Source rights: who owns or licenses the input, and which acts are permitted?
2. Processing: who receives the material, for what purposes and for how long?
3. Academic permission: what assistance does this particular assessment allow?
4. Reliance: what has the provider promised, and how could errors harm the learner?
5. Accountability: who can correct, explain, refund, investigate or answer a complaint?

This framework is our analytical proposal. It is not a statutory test. Its advantage is that it prevents one favourable answer, such as a private account or a no-training promise, from being treated as a universal permission.

Method and limitations

The source review prioritises legislation, judgments, regulators, official product documentation and original educational research. Search terms included private study and commercial intermediaries; AI tutoring and learning modes; copyright and training; hallucinated authorities; consumer subscriptions; children's data; and AI in education. Product descriptions are vendor representations, not independently tested outcomes. Case discussions distinguish holdings from our application by analogy.

The search did not identify a judgment directly resolving the precise single-student commercial AI-processing scenario developed below. This is a bounded search finding, not proof that no relevant case exists. No paid citator was available. Subsequent treatment, appeals and statutory commencement must be refreshed before release. Direct retrieval of several consolidated legislation pages failed in the drafting pass; the independent reviewer subsequently retrieved the central copyright, cheating-service and automated-decision provisions. The source register distinguishes those checks from remaining access and subsequent-treatment limitations. The paper must not imply a comprehensive current-law audit.

2. What are the products actually offering?

2.1 ChatGPT Study mode

OpenAI describes Study mode as a learning experience that works through problems using questions, structured explanations and checks of understanding. Its documentation also describes using uploaded material and, where available and enabled, memory for personalisation. It acknowledges that the system can give direct answers and make mistakes, and expressly says it does not replace teachers, tutors, course materials or academic requirements. These are important qualifications to any description of a ChatGPT tutor replacement. OpenAI, Study mode documentation.

The service should be distinguished from the underlying model and from a third-party app using the OpenAI API. A study interface does not establish identical contracts, data settings or retention across all three. OpenAI's documentation distinguishes personal-workspace training controls from the default exclusion of Business, Enterprise and Edu content from model training. A user can keep history while opting out of training. Consequently, neither Study mode nor a no-training setting should be described as device-only processing or automatic deletion. OpenAI, data controls; OpenAI, history and training.

2.2 NotebookLM and Google's source-based learning tools

Google's September 2025 NotebookLM announcement describes source-based flashcards, quizzes, tailored reports, Audio Overviews and a Learning Guide using open questions and adapted explanations. These functions combine document processing with tutorial interaction. The important legal distinction is between being grounded in a supplied source and having permission to process that source. A citation demonstrates a proposed connection to material; it does not grant rights or establish that the interpretation is correct. Google, NotebookLM student features.

Product-name caveat: on this review date, the NotebookLM help URLs retrieved by our research tool redirected to pages headed Gemini Notebook. We retain NotebookLM when discussing the dated launch material and the product the reader may recognise. The timing and extent of any branding transition have not been independently established and should be checked before publication. Google, current help page.

Google's current help distinguishes personal use from qualifying Workspace and Workspace for Education use. Its feedback notice describes collection of associated content and human review for ordinary feedback, while stating stronger exclusions for Workspace and Education users. The practical research question is the actual account, service and feedback route, rather than an unqualified statement that Google never uses uploaded data. This is a description of published terms, not an audit of Google's systems. Google, privacy and feedback.

2.3 A wider market, not a two-product contest

Google's Guided Learning positions Gemini as an interactive learning companion using questions, visual explanations and quizzes. Anthropic's Claude for Education describes Learning mode in Projects, including Socratic questions and support for independent thinking. Both are relevant examples of providers presenting learning as a process rather than merely supplying completed answers. Their descriptions do not show that every user interaction follows that design. Google, Guided Learning; Anthropic, Claude for Education.

This comparison is our issue map, not a product ranking. Mentioning providers invites scrutiny of their public positions; it does not imply an invitation has been sent or accepted.

3. Does the evidence support replacing tutors?

One important randomised study compared a specially designed AI tutor with active classroom learning in an undergraduate physics setting. It reported stronger immediate learning outcomes for the AI condition. It did not compare every commercial chatbot with expert one-to-one law tutors, nor establish long-term professional judgment or pastoral competence. The intervention's design and the comparator matter. Greg Kestin and others, Scientific Reports 15, 17458 (2025).

Contrary evidence is equally important. A randomised field experiment with nearly 1,000 high-school mathematics students found that a general GPT-4 interface improved assisted practice performance but reduced subsequent unaided performance relative to the control group. A tutor designed with teacher-informed safeguards substantially mitigated that harm; this did not demonstrate a positive unaided-examination effect. The study concerns a particular population, subject and system, not today's Study mode or Durmah. Read alongside the physics study, it supports attention to instructional design and independent learning rather than a single verdict about AI tutors. Hamsa Bastani and others, PNAS 122(26), e2422633122 (2025). The published correction concerns an author affiliation, not the reported results. Correction, August 2025.

Our inference is that targeted substitution of some instructional tasks is plausible, while wholesale replacement is a much larger claim. A law tutor may help a student notice ambiguity, question a source's status, prepare oral argument, understand assessment expectations or seek human help. Those functions should be tested separately. Fluency, satisfaction, time saved and a high score immediately after assistance are not interchangeable with retained understanding.

A credible evaluation for legal education would compare unaided later performance, transfer to new facts, identification of incorrect authorities, willingness to question confident advice, accessibility and effects across student groups. It would report the model, prompts, source materials and permitted interventions. Until such evidence exists for the particular service and population, a provider should frame improved grades or human-equivalent tutoring as hypotheses requiring evidence, not established benefits.

4. Copyright: the learner's purpose does not answer every copying question

4.1 Identify the work, owner and act

UK copyright protects qualifying expression rather than legal doctrines or facts as such. A student may write an original explanation of consideration without a university owning the doctrine. The lecturer's wording, diagrams and a publisher's case commentary may nevertheless attract protection. Original notes can also contain third-party material. Tuition fees and lawful access do not by themselves transfer ownership. WIPO, copyright FAQ; Bodleian Libraries, copyright for students.

The statutory starting points are CDPA 1988 sections 16 and 17: restricted acts include copying, and electronic storage can be a form of reproduction. An analysis should separately consider uploading, extraction, indexing, transmission to a model provider, retention, generated quotations and onward sharing. Not every technical representation necessarily reproduces a substantial part of protected expression. Whether an embedding, extract or output does so requires evidence, not assumptions about its label. CDPA, sections 16 and 17.

4.2 Private study is a real exception with real limits

Section 29 distinguishes research for a non-commercial purpose and private study. Subsection (1C) provides the private-study fair-dealing route; the statutory definition in section 178 excludes study directly or indirectly for a commercial purpose. The non-commercial research route includes sufficient acknowledgement, subject to the statutory qualifications. Subsection (4B) makes contractual restrictions unenforceable to the extent they prevent qualifying acts under section 29. An exception is a legal entitlement when its conditions are met, not merely permission a provider can choose to recognise. CDPA, section 29; section 178.

Fair dealing has no universal word-count or percentage safe harbour. The amount and significance taken, the purpose and market effects matter. IPO guidance explains that copying a whole work would not generally qualify. That is not the same as a mechanical rule that every complete work, however short, is forbidden in every setting. A narrowly selected passage presents different facts from a substitute library of commercial textbooks. IPO, exceptions to copyright.

Section 29(3) also addresses copying by someone other than the researcher or student, including specified circumstances involving substantially the same material being supplied to multiple people at substantially the same time for substantially the same purpose. Not falling within that exclusion is not an affirmative licence: the rest of the fair-dealing analysis still matters. Conversely, the existence of a paid intermediary does not justify skipping the actual statutory analysis. CDPA, section 29(3).

4.3 The unresolved commercial-tool problem

Consider a student who lawfully accesses a short passage, asks a paid AI service to explain it solely to that student, and does not permit reuse, sharing or model training.

The strongest argument for permissibility is functional: the student remains the initiator and beneficiary; the tool is an instrument of private study; the dealing is proportionate; and no substitute product is supplied to others. Many ordinary study activities use paid equipment or services. Treating the supplier's commercial status as decisive risks making private-study rights depend on whether the student owns sophisticated computing resources.

The strongest objection is actor-specific: the provider and its suppliers may make distinct copies for their own business, and the student's purpose may not extend to those acts. A provider may retain or reuse more than the student expects. At scale, individual requests can support a service that competes with licensed educational material, even if accounts are separate. Characterising the service as merely an instrument may conceal meaningful independent control.

There is also a textual objection. The Act expressly regulates third-party copying in section 29(3), with a library-specific rule and a separate multiple-recipient rule for other copiers. A critic can argue that this careful structure, coupled with the non-commercial private-study definition, leaves no basis for a broad implied exception covering a provider's own dealings. The best reply is narrower: section 29(3) expressly contemplates other copiers and does not say that every non-library intermediary is prohibited; who legally makes each automated copy, and for whose purpose, still requires analysis. As a textual argument, this structure suggests that third-party copying is not excluded as such, while leaving the other fair-dealing requirements in place. Neither the library provisions nor an isolated account establishes the answer. In particular, the multiple-recipient wording concerns substantially the same material and the copier's knowledge, not merely whether users can see each other's files. CDPA, sections 29(3) and 42A.

Our provisional position is that both the commercial status and the technical design are relevant, but neither determines the answer alone. Limiting amounts, purposes, access and reuse improves the facts; it does not create an exemption. The instrument argument is an argument for examination, not settled doctrine. A rights-holder licence covering the actual operations can resolve uncertainty that a user declaration cannot.

The existing Durmah page invokes Sillitoe v McGraw-Hill Book Co (UK) Ltd [1983] FSR 545. Because the full report and subsequent treatment were not obtained in this first pass, we do not use it as a verified holding supporting our conclusion. Further research should obtain the report, identify its treatment of commercial study notes under the earlier legislation, and examine its relevance and limits under today's Act. A commercial publisher's prepared study notes and a responsive tool processing one student's passage should not be treated as identical without analysis.

4.4 Training, inference, temporary copies and teaching

Training changes a model; inference uses a model to produce a response. A no-training arrangement removes one possible purpose of reuse. It does not mean that uploads, retrieval stores, logs or response generation involve no copying. Section 29A concerns computational analysis for non-commercial research by a person with lawful access, subject to conditions including restrictions on other uses and transfers. It is not a general permission for commercial model training or every student-facing AI operation. CDPA, section 29A; IPO, research exceptions, pp 6-9.

Section 29A also places restrictions on transferring the analysis copy to another person and using it for another purpose without the copyright owner's authority. Sending a copy to an external model supplier therefore requires attention to the transfer condition, not simply a student's non-commercial intention. Section 29A(2).

Section 28A's temporary-copy conditions are separate. For covered works, the copy must be transient or incidental, integral and essential to a technological process, serve the specified transmission or lawful-use purpose, and have no independent economic significance. The provision excludes computer programs and databases from this particular route. A copy's role, duration and separate value need examination; charging for a service does not by itself answer the question of independent economic significance. Public Relations Consultants Association Ltd v Newspaper Licensing Agency Ltd [2013] UKSC 18 and the subsequent CJEU decision, Case C-360/13, concern ordinary browsing copies. They do not establish that a retained AI document store or an upload deleted after processing automatically qualifies. The CJEU ruling predates Brexit; its present domestic relevance must be assessed within the UK's post-Brexit case-law framework. Section 28A; UK Supreme Court case; CJEU judgment, particularly paras 24-63.

Section 32 requires fair dealing for the sole purpose of illustration for instruction, a non-commercial purpose, dealing by a person giving or receiving instruction or preparing to do so, and sufficient acknowledgement unless impractical or otherwise impossible. Subsection (2) includes setting and answering examination questions. A provider cannot simply borrow a student's position: identify who carries out the dealing and whether each condition is met. For the provider's own copying, the non-commercial-purpose condition is a particularly strong objection requiring analysis of that dealing. Neither calling an app a tutor nor calling its supplier commercial conclusively resolves those questions. Section 30's quotation and criticism/review provisions may matter to particular outputs, but also have their own requirements. The correct exception must be matched to the act; exceptions are not interchangeable labels. CDPA, section 32; section 30.

4.5 Outputs and additional rights

A summary can convey unprotected ideas or reproduce protected expression. A source citation does not immunise the latter. Source databases, lecture performances, confidentiality and contractual access controls may raise separate issues from literary copyright. Public access also does not make every judicial document, editorial headnote or database feature unrestricted for bulk reuse. Our operational recommendation is to identify the specific source licence and avoid treating a law-report database as interchangeable with an official judgment.

For original AI-assisted output, the student's own creative contribution, any copied expression, and the unsettled application of computer-generated-work provisions should be kept separate. CDPA sections 9(3) and 178 address works without a human author; they do not establish that every prompt automatically gives the user copyright in every output. The government's March 2026 report discusses policy uncertainty here. CDPA, section 9; government report, section I.

5. What recent cases do, and do not, decide

Getty Images v Stability AI

In Getty Images (US) Inc and others v Stability AI Ltd [2025] EWHC 2863 (Ch), the court recorded abandonment of the training/development, output and associated database claims described at paragraph 9, including the absence of evidence establishing training in the UK. The secondary-infringement analysis distinguished model weights from copies of protected works on the facts considered, particularly at paragraphs 597-603. There were also findings on trade marks. It is therefore misleading to cite the result as a universal declaration that AI training or generated outputs are lawful. A system that stores students' documents presents different facts from the model-weights issue. This is our distinction, not a holding about tutors. Judgment, 4 November 2025.

The later form-of-order judgment, [2025] EWHC 3343 (Ch), 16 December 2025, granted Getty permission to appeal the secondary-copyright decision, identifying a novel and important question of statutory construction. It also made clear that dismissal of the abandoned primary claims should not be mistaken for adjudication of their merits. Permission is not reversal. This review has not established the final subsequent appellate position, so no assertion of finality or an appeal still pending at the cut-off is made. Form-of-order judgment, paras 4-11.

Ayinde and Al-Haroun

R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin) addresses false legal authorities, verification and responsibilities to the court. The duty discussion is at paragraphs 6-8. The court considered alternative explanations in Ayinde at paragraph 68, rather than finding AI use proved; Mr Al-Haroun acknowledged using AI tools and other online sources at paragraph 76. The judgment warns against treating generated citations as research. It is not a prohibition on AI-assisted learning, nor does it establish that every erroneous student answer produces professional disciplinary liability. Our educational implication is to teach source verification as part of legal method. Judgment, 6 June 2025, paras 6-8, 68 and 76.

Emotional Perception AI

Emotional Perception AI Ltd v Comptroller General of Patents, Designs and Trade Marks [2026] UKSC 3 concerns the patentability analysis for a system involving an artificial neural network. It matters to businesses seeking protection for AI inventions. It does not answer copyright permission for training data, liability for inaccurate tutoring or permission to submit AI-assisted work. We include it to show why a recent AI judgment must be matched to the legal question, not cited merely because it concerns AI. Judgment, 11 February 2026.

Judicial guidance and pending reform

The judiciary's October 2025 AI guidance is professional guidance, not a judgment creating a new copyright exception. Similarly, the government's 18 March 2026 copyright-and-AI report evaluates policy options and evidence; it is not itself enacted permission. An announced proposal, an enacted section and an operative rule on the relevant date must be distinguished. Judicial guidance; government report.

6. Academic integrity is not simply a copyright question

A student can own an essay and still breach assessment rules by obtaining prohibited assistance. Conversely, permitted discussion with a tutor does not license copying a third party's material. Institutions may distinguish brainstorming, explanation, feedback, translation, proofreading and generation. Permission should be checked against the actual assessment, not inferred from generic institutional enthusiasm for AI.

The Skills and Post-16 Education Act 2022, Chapter 1 of Part 4 (sections 26-30), creates an additional issue for commercial cheating services connected to the students and assignments defined in the Act. These provisions came into force on 28 June 2022 under section 36(2). The explanatory notes distinguish completing a student's work from generally accepted tutoring, proofreading and teaching. That does not make everything branded tutoring exempt: whether the service completes work that should be the student's, and the statutory elements and defences, require close analysis. These provisions must not be casually described as a uniform UK-wide ban on all AI assistance. Act, Chapter 1 of Part 4; section 36; official explanatory notes, paras 147 onwards.

The defined England connection matters: section 26 covers students undertaking a relevant course at a post-16 institution or sixth form in England, and certain other examination candidates over compulsory school age taking a regulated-qualification examination in England. A relevant assignment is one required to be completed personally, taking permitted assistance into account. It would therefore be inaccurate to make nationality or the provider's foreign incorporation the sole scope test. The precise application to a generative service remains an interpretative issue, not a decided AI-tutoring rule identified by this paper. Section 26(6)-(8), official Act text.

A difficult boundary case is an assistant that never supplies a complete essay but successively replaces the reasoning, structure and language until little remains of the student's contribution. Our view is that an output-length restriction alone is an inadequate integrity policy. The relevant questions include the task's permitted assistance, the student's actual contribution, the intended submission and the service's knowledge and design.

The statutory detail matters here. Section 26(3)-(4) addresses provision of material in connection with an assignment, including material that has not been published generally, rather than only delivery of a complete ready-to-submit essay. Section 27 provides a defence tied to knowledge and reasonable diligence, but a written standard contract term about permitted assistance or the student's intended use is not sufficient evidence by itself. A checkbox is therefore not a substitute for considering the defined service and assignment. Section 28 separately addresses advertising; section 29 concerns corporate and other organisational responsibility; section 30 supplies further definitions. These are statutory starting points, not a conclusion that every personalised explanation is an offence. Section 27; official explanatory notes, especially paras 148-159.

There are competing institutional interests: reliable assessment, fair access to support, clear notice and a fair opportunity to answer an allegation. AI-detection scores should not be presented in this paper as conclusive proof of misconduct. This draft does not evaluate any detector or claim a decided legal rule about a particular disciplinary procedure. Those questions need their own evidence.

7. Privacy, memory and the student under 18

Purpose and responsibility

AI tutoring may process names, drafts, voices, learning difficulties and inferences about performance. An apparently ordinary essay can contain information about other people. The ICO distinguishes developing an AI system from deploying it and requires the purposes and lawful bases to be considered for the relevant processing. A commercial service should not assume that all processing falls under educational research or that one consent covers every later use. ICO, lawfulness in AI.

The UK GDPR framework raises questions of territorial scope, controller/processor roles, transparency, lawful basis, special-category data, minimisation, retention, security, rights, impact assessment and international transfers. For a Singapore operator serving UK students, establishment abroad does not by itself remove UK obligations. The actual targeting and processing arrangements matter; so does whether an Article 27 representative is required. These are issues to resolve against facts, not conclusions about Durmah's completed compliance. UK GDPR, Articles 3, 5, 6, 9, 13-14, 27-28, 32, 35 and Chapter V.

Significant automated decisions

Feedback on a practice paragraph should not automatically be equated with a decision determining admission, progression or access to an educational opportunity. The Data (Use and Access) Act 2025 section 80 replaced Article 22 with Articles 22A-22D. The amended scheme permits a wider range of solely automated significant decisions, while requiring safeguards including information, representations, human intervention and contesting a decision, and retaining stricter conditions concerning special-category data. Meaningful human involvement and legal or similarly significant effects are central questions. It would be misleading to reproduce the former Article 22 regime as unchanged. Section 80; official explanatory notes on section 80.

The government's commencement record places the main data-protection tranche on 5 February 2026; the associated instrument includes section 80 on automated decision-making and section 81 on children's higher protection. That date helps distinguish operative reform from proposals. It does not remove the need to examine the amended provisions, safeguards and any relevant transitional rules for a particular use. Government commencement record; SI 2026/82.

Children and sensitive inferences

A minimum age of 16 would still admit children under the data-protection framework. The ICO's Children's code guidance distinguishes services likely to be accessed by children and different edtech arrangements. A direct-to-consumer app cannot assume it enjoys the same position as a service used solely under a school's instructions. Actual likely use matters. ICO, Children's code and edtech.

Our design questions include whether a tutor should remember a student's disclosed disability, infer distress from voice, or optimise engagement late at night. Helpful personalisation and intrusive profiling can arise from the same feature. The research agenda should examine what is necessary for learning, how a student can correct or delete an inference, and whether the interface nudges children to disclose more than they intended. A comforting voice is not evidence of clinical competence or a substitute for human safeguarding.

8. Consumer law: promises are part of the product

An AI tutor may be supplied as a service, digital content or a mixed arrangement. The classification affects the relevant Consumer Rights Act 2015 provisions. Services attract the reasonable-care-and-skill requirement; relevant information about the service can become binding. Digital-content quality provisions and unfair-term controls also deserve attention. A disclaimer that AI can make mistakes does not automatically displace statutory rights or cure a contradictory headline promise. CRA 2015, particularly sections 34-36, 49-50, 57 and 62; official explanation of section 57.

The Digital Markets, Competition and Consumers Act 2024 unfair-commercial-practices regime applies to practices from 6 April 2025. Claims such as guaranteed grades, official university approval, universal legal accuracy or replacement of expert tutors require careful substantiation and presentation. The paper's own public framing can also affect consumers' understanding of Durmah, so calling it research does not excuse promotional overstatement. CMA, unfair commercial practices guidance, para 1.15 and relevant substantive chapters.

Subscription law illustrates the importance of dates. The government's April 2026 response anticipates the new subscription-contract regime commencing in spring 2027. That anticipated date is not proof of commencement. Existing consumer-contract requirements still need analysis, including cancellation and the distinctions between starting services and supplying digital content. A free trial should disclose conversion and cancellation accurately. DBT, subscription consultation response; Consumer Contracts Regulations 2013.

If a learner suffers loss after relying on incorrect guidance, contractual remedies and potentially negligence questions require facts about promises, duty, reasonable reliance, causation and recoverable loss. This draft does not claim an automatic entitlement to compensation for a poor grade or identify a decided AI-tutor negligence precedent. That is an open research question, particularly where a provider knows the student cannot reliably detect an error.

9. Accessibility, safety and jurisdiction

AI could make explanation more accessible through varied pace, text and audio. It could also disadvantage students through inaccessible interfaces or unreliable interpretation of speech and language. In Great Britain, Equality Act 2010 duties and reasonable-adjustment obligations are relevant according to the actor and service. Northern Ireland has a different equality-law framework and must not be folded into an undifferentiated Equality Act claim. A commercial app's service duties and a university's education duties require separate analysis. EHRC, services code, chapter 11; chapter 7, reasonable adjustments.

Online Safety Act coverage is feature-dependent. Ofcom describes relevant user-to-user, search and pornography-service definitions, and explains that some standalone chatbots fall outside them. Adding public sharing or multi-source search can change the analysis. A private-tutor label is not a reliable legal scope assessment. The position should be refreshed if Parliament changes the rules or the service adds features. Ofcom, AI chatbots and online regulation.

That December 2025 explainer is not the end of the policy story. The government's July 2026 response, updated in August, announces further chatbot safeguards, including breaks for users under 18, and describes action to close the illegal-content coverage gap. It leaves details of scope and breaks to further work and regulations. An educational app admitting 16-17 year olds should follow these developments; the announcement alone does not establish that every proposed duty is already operative. This paper does not assert an educational exemption or a current universal chatbot-break requirement. Government response, chapter 3.1.

The Department for Education's generative-AI product safety standards are also relevant design material. Their role, audience and legal status should be described accurately; following guidance is not the same as obtaining a regulatory certification. DfE, product safety standards.

Cross-border expansion introduces further questions. The EU AI Act contains education-related high-risk categories tied to specified uses, including certain assessment and access decisions. It is not UK domestic law, and an ordinary study aid should not be declared high-risk solely because it concerns education. Territorial scope, intended purpose, exceptions, amendments and application dates need a separate review before an EU launch. Regulation (EU) 2024/1689, Articles 2 and 6 and Annex III.

For legal-learning apps specifically, moving from hypothetical education to advice about a real person's dispute raises additional competence, confidentiality and legal-services questions. We do not equate all legal information with reserved legal activity, or treat an educational disclaimer as sufficient for every service. This boundary should be assessed before offering case-specific representation or procedural assistance.

10. Durmah's experience: questions encountered, not a court dispute

Durmah's records document a small student pilot that ended in September 2026 by founder decision. The member product was then closed while future provision was reconsidered. It would be inaccurate to present this simply as a product that had never launched. Its public pages describe material-use boundaries, provider checks and a distinction between individual assistance and training. Those statements and the pilot closure provide context for this paper; they are not evidence that a court has approved or condemned the design. No private correspondence is reproduced here, and we do not attribute the closure to correspondence we have not examined. Responsible Study and Student Rights; service availability.

The documented challenges were how to recognise students' own rights without overstating their ability to authorise third-party copies; how to avoid turning a prudent safeguard into a claim that all other study is unlawful; and how to distinguish intended no-training commitments from verified provider arrangements. The practical case study on the existing page already invites opposing arguments. Existing case study; private-study guide.

This paper does not establish any court ruling, regulatory determination or external legal clearance for or against Durmah. It also makes no claim that no university raised concerns. The questions arising from the pilot and subsequent redesign must be examined on their merits, without treating either closure or continued development as a legal finding. We do not identify students or characterise unseen correspondence.

Our proposed position is deliberately contestable: make source use proportionate; isolate individual content; avoid secondary training and shared-library reuse; explain provider processing accurately; honour credible rights complaints; and preserve the student's own reasoning. These are design commitments to evaluate, not proof of legal entitlement. An operator should be prepared to revise a feature where the necessary rights or safeguards cannot be established.

11. Four problems for a commercial-law seminar

Problem A: the paragraph. A student copies a short, relevant passage from a lawfully accessed article into a paid tool, asks for an explanation, and receives a non-reproductive answer. The provider retains the passage for thirty days. Identify the acts and actors. Does retention change the fairness analysis? What evidence supports treating the provider as an instrument of the student?

Problem B: the course pack. Five hundred students separately upload substantially identical teaching materials. The service isolates their accounts but generates similar revision resources. Does individual delivery resolve section 29(3), fairness or market substitution? What would change if the provider marketed the service as a replacement for buying the prescribed textbook?

Problem C: the marked essay. A student owns the draft, the provider excludes training, and the university permits spelling correction only. An assistant rewrites the argument paragraph by paragraph. Which issues concern ownership, which concern academic rules, and which could concern the commercial-cheating provisions? What information would a provider need to avoid relying solely on a checkbox?

Problem D: the confident authority. A tutor app produces a plausible but false citation. The student uses it in a practice exercise, an assessed submission, or a real legal dispute. How do reliance and possible loss differ? What would count as adequate verification, and who was represented as responsible for it?

These are fictional fact patterns. None alleges unlawful conduct by a named provider. Their purpose is to expose facts capable of changing an answer rather than supply a model answer for assessment submission.

12. Provisional conclusions and the strongest remaining doubts

Our conclusion is function-specific. AI can support, and may substitute for, particular tutoring activities. The evidence reviewed does not establish that a general-purpose AI service replaces the full educational and accountable role of a human tutor. Nor does using a tutorial interface create a special legal category exempt from ordinary obligations.

The private-study question remains particularly significant. Student rights should not be erased merely because software has a commercial supplier; equally, those rights should not be stretched to cover every provider-controlled copy or reuse. The most valuable next legal contribution would identify the doctrinal basis, and limiting principles, for attributing an intermediary's operations to the student's purpose.

We invite disagreement on five propositions:

1. A proportionate, individual request may present materially different facts from commercial publication of study notes, but no automatic instrument exemption has been established here.
2. A no-training commitment narrows one risk without resolving input rights, retention, provider access or assessment permission.
3. Source grounding can improve traceability without guaranteeing legal correctness.
4. Claims of replacing tutors should be assessed against the particular function and evidence, including the responsibilities the product omits.
5. A responsibly uncertain explanation can be more useful to a law student than an apparently definitive answer that suppresses contested points.

The strongest objection to our approach is that a list of safeguards may create an appearance of legitimacy before the central legal issue is resolved. That objection should be taken seriously. The answer must come from the applicable rights, facts and authorities, not from the number of safeguards or the reputations of the models used to write this paper.

13. An invitation to challenge and improve this paper

We welcome contributions from students, lecturers, librarians, copyright specialists, privacy and consumer lawyers, accessibility researchers, tutors, rights holders and AI providers. A useful contribution identifies the passage challenged, supplies an accessible authority or evidence, explains the competing interpretation and states what facts would change the conclusion.

Please use the contribution form below and identify this paper and its version. Do not send restricted teaching materials, confidential client matters or student records. We will seek explicit permission before publishing a contributor's name, quotation or substantive contribution. Contacting us does not imply endorsement or agreement to publication.

We will distinguish corrections of fact, competing interpretations and policy proposals. A dissent need not disappear because the editor prefers another view. Where a dispute cannot be resolved, the paper should retain the competing positions and identify the evidence still missing.

14. Version and correction policy

Each released version should show its research cut-off, publication date, version number, editor, review status, source changes and material corrections. Earlier versions should remain identifiable. A new model or product feature, judgment, appeal, statute, commencement order, regulator publication or substantive reader correction should trigger review of the affected claims.

The editorial cadence is a monthly source review plus prompt review of material corrections. A monthly task has been configured to prepare private change reports; this is not continuous verification and does not automatically update the public paper. Updates should pass source checking, independent challenge, reconciliation and human approval before publication. A two-model process is a quality-control method, not legal peer review.