The Day I Became a Case Study

Accountable Cognition, Academic Integrity, and AI as Accessibility Infrastructure

Dr Danie Adendorff, DSc, MSc

There is a moment when a technology stops being a convenience and becomes part of the infrastructure that allows a person to function. For me, artificial intelligence crossed that line not because it replaced my scholarship, but because it helped me make my scholarship visible. I am dyslexic and neurodivergent. I have lived with the gap between what I can think, analyse and connect, and what the page sometimes allows me to express cleanly. AI did not give me my subject knowledge. It did not give me my professional experience. It did not give me my ethical position. It helped me convert those things into disciplined, readable form.

That is why I became a case study. Not because my situation is exceptional, but because it exposes a blind spot in the current academic debate. Too much of that debate asks whether AI use is a threat to integrity. That is an important question, but it is not the only question. The equally important question is whether prohibiting or stigmatising responsible AI use may exclude disabled and neurodivergent scholars from full participation in academic life.

The better framework is not technological permissiveness. Nor is it academic purity. The better framework is accountable cognition: the use of tools in ways that preserve human explanation, evidential discipline, transparent material assistance, data protection and responsibility for final judgement. A researcher may use AI, synthetic data, open data, transcription tools, language tools, coding tools, reference tools and assistive technologies; but the researcher must remain able to explain the method, defend the evidence and accept accountability for the work.

This article therefore narrows the earlier title. It is not a full treatise on open science or data governance. It is an ethics-of-AI-use article with a personal case study at its centre, and with open data and synthetic data treated where they matter most: as tests of provenance, disclosure, privacy and epistemic trust.

1. From suspicion to access

Academic debate often begins with suspicion: AI may fabricate sources, conceal plagiarism, weaken originality, leak confidential information, and allow students or researchers to submit work they do not understand. These risks are real, but starting with them alone creates a distorted ethical field. It places responsible disabled users in the dock before the accessibility argument has even been heard.

The access argument should come first. Academia has never been unaided. Scholarship depends on libraries, indexes, statistical packages, laboratory instruments, collaborative editing, peer review, screen readers, dictation software, translation aids and reference managers. The relevant question is not whether a tool has assisted the scholar. The relevant question is whether the scholar remains intellectually accountable for the work produced with that tool.

For dyslexic and neurodivergent academics, AI can function as a cognitive ramp. It can reduce friction in reading, sequencing, drafting, checking, reformulating and organising complex material. This does not lower the scholarly standard. It changes the route to the standard. A wheelchair does not make movement inauthentic. A screen reader does not make reading inauthentic. Speech-to-text does not make writing inauthentic. In the same way, AI-mediated writing support does not make knowledge inauthentic when the scholar retains ownership of the argument.

2. The case study: what AI actually did for this article

The personal case must be operational, not decorative. In this article, AI was used as an assistive research and drafting environment. It helped convert a broad ethical concern into a clearer article architecture; identify where the argument was front-loaded; reorganise the thesis around accountable cognition; generate alternative section sequencing; strengthen transitions; and convert peer-review criticism into specific revision tasks.

AI was also used to support source discovery and source discipline. It helped identify current governance sources on responsible generative AI in research, research-integrity policy, synthetic data risk, and disability-related AI use in higher education. Those suggestions still required human checking. Sources had to be verified, unsuitable sources rejected, and claims restrained where the evidence did not support stronger wording. The tool was not treated as an authority; it was treated as an aide requiring supervision.

The most important corrections in this rewrite came from human editorial judgement. The earlier version over-promised on open data relative to its real emphasis. It introduced accountable cognition too late. It allowed the personal case study to retreat after the opening. It did not sufficiently anchor the accessibility claim in disability-specific literature. These weaknesses were not solved by asking AI to make the text smoother. They were solved by treating the review as a diagnostic instrument.

This workflow illustrates the ethical distinction that matters. AI assisted with formulation, structure, language, source discovery, revision and formatting. It did not decide what I believe. It did not create my lived experience. It did not relieve me of responsibility for the claims. The final article remains mine because I can explain why each section is here, what each claim means, what the limits are, and where the evidence is strongest or weakest.

3. Assistance is not substitution

The distinction between assistance and substitution is the most practical ethical line in AI-supported academic work. Assistance occurs when AI helps the researcher perform tasks under human direction. Substitution occurs when AI replaces the researcher’s intellectual labour while concealing that replacement.

Assistance includes clarifying a research question, improving readability, generating search terms, summarising material that is then checked, identifying weaknesses in an argument, supporting translation, helping disabled researchers organise information, producing draft structures, or generating code that is inspected and tested. Substitution includes submitting AI-generated analysis as if it were independently produced, using fabricated sources, inventing data, presenting synthetic material as observed evidence, or allowing the AI system to determine the argument without human understanding.

The objection is predictable: critics will say that this line is impossible to police. They are partly right. No integrity system can perfectly police intention, hidden assistance or intellectual dependency. But difficulty of enforcement does not invalidate the distinction. Universities already police distinctions that are imperfect: collaboration versus collusion, editing versus authorship, legitimate proofreading versus ghost-writing, statistical support versus fabricated analysis, and supervisory guidance versus student substitution.

A workable standard is therefore not detection-centred but accountability-centred. The researcher must be able to answer a simple question: can I explain, defend and verify every substantive claim without appealing to the authority of the AI system? If the answer is no, the work is not ready. If the answer is yes, then AI use may be disclosed, governed and defended as part of a responsible workflow.

4. Why the disability evidence matters

The accessibility claim cannot rest on my case alone. My experience is evidence of a lived problem, but responsible argument needs wider support. Recent research on disabled students in higher education has found that generative AI is already being used to overcome barriers in academic writing, especially through chatbots, rewriting applications and translation tools (Zhao, 2025). The same research also records the risks: inaccurate AI answers, academic-integrity concerns, subscription-cost barriers and the danger that disabled students may be left to manage complex tools without sufficient institutional guidance.

This is precisely the balance required. AI can widen access, but access without training produces new vulnerabilities. Disabled and neurodivergent scholars should not be told simply to use AI and hope for the best. Nor should they be told that AI assistance is presumptively suspect. They need clear institutional rules, discipline-specific guidance, disclosure norms, privacy protection, and assessment models that distinguish intellectual ownership from mechanical polish.

Assistive-technology research has long recognised that tools can support writing, learning and independence for people with learning disabilities, although effectiveness varies by tool, user, context and training. Generative AI should be placed within that assistive-technology tradition, but not absorbed uncritically into it. It is more powerful than a spellchecker and more opaque than a word processor. That means it requires stronger disclosure and stronger accountability.

5. Open data: transparency without recklessness

Open data is usually defended on strong grounds: transparency, reproducibility, public value, reuse, error detection and efficient science. Those aims remain important. But open data is not open season. The ethical commitment is not maximum exposure; it is responsible access.

The same principle applies to AI-supported research. If a researcher uses open datasets, AI tools may help clean, classify, summarise or visualise them. But the provenance of the dataset still matters. Licensing conditions still matter. Consent still matters. Sensitive attributes still matter. The fact that data is publicly accessible does not automatically make every reuse ethical, lawful or contextually safe.

This is particularly important in disability, health, education, migration, policing, conflict, employment and security research. Data can be technically open but socially dangerous. Re-identification can occur through linkage. Small populations can be exposed through rare combinations of traits. Geolocation, timestamps, institutional affiliation and behavioural traces can carry risk even when names are removed.

Responsible open-data practice therefore requires layered judgement: what may be public, what should be controlled, what should be aggregated, what should be synthetic, what should be withheld, and what should be available only through trusted access conditions. AI does not remove these judgements. It makes them more urgent because it increases the speed and scale at which data can be recombined.

6. Synthetic data and the danger of synthetic authority

Synthetic data appears to offer an elegant solution: generate artificial records that resemble real data, then share or analyse them with less privacy risk. In principle, synthetic data can support model training, rare-event simulation, teaching, software testing, privacy-preserving research and fairness analysis. In practice, it can also create a dangerous illusion of safety.

The central risk is synthetic authority. Synthetic authority arises when generated data acquires the visual, statistical or rhetorical appearance of empirical fact. A clean synthetic dataset may look more complete than the real world. A simulated population may appear more balanced than actual sampling would allow. A model output may look precise while resting on assumptions that cannot carry the policy weight placed upon them.

This risk connects directly to my own research practice. I use AI to assist with structure, language and source discipline, but I do not treat generated text, generated examples or simulated patterns as empirical evidence. In security, policy and academic analysis, that boundary is essential. Synthetic material may help illustrate a concept, test a scenario or expose a methodological risk; it must not be allowed to masquerade as observed reality.

Synthetic data must therefore be labelled, documented and validated. Researchers should assess statistical fidelity, task-specific utility, privacy leakage, representativeness, bias reproduction and fairness effects. They should avoid presenting synthetic data as direct evidence unless the research design justifies that use. Synthetic data may illustrate, test or simulate; it does not automatically observe.

The privacy issue is especially serious. Synthetic does not necessarily mean anonymous. Synthetic datasets can still leak information about the underlying training data, especially where models reproduce rare patterns, preserve local structure, or permit membership or attribute inference. Any claim that synthetic data is privacy-preserving should be treated as a hypothesis requiring evaluation, not as a default assumption.

7. Disclosure should be proportionate, not punitive

The proper disclosure rule is materiality. Minor spelling correction does not require the same disclosure as AI-assisted thematic coding, synthetic data generation, statistical interpretation or substantive drafting. A reference manager is not the same as an AI system that proposes an argument. A grammar tool is not the same as a chatbot that writes a literature review.

Where AI materially affects research design, analysis, interpretation, writing, data generation or presentation, disclosure should be expected. Disclosure should not be framed as confession. It should be framed as provenance. The reader has a legitimate interest in knowing how the work was produced, especially where AI influenced method, evidence selection or argument structure.

For disabled scholars, disclosure must be handled carefully. A requirement to disclose every assistive action may become intrusive and discriminatory. The standard should not demand disclosure of disability status. It should require disclosure of material AI involvement in the work. A scholar may state that AI tools were used for language support, structuring, source checking or accessibility assistance without being forced to disclose medical or diagnostic details.

A practical institutional formula would be a standard protected disclosure rather than a personal explanation. For example: ‘AI-supported tools were used for accessibility, language, formatting, organisation, or research-process support. All substantive claims, evidence selection, analysis and conclusions remain the author’s responsibility and were reviewed by the author.’ Such wording normalises responsible assistive use without requiring a scholar to reveal dyslexia, ADHD, disability status or other protected personal information. It also keeps the ethical focus where it belongs: on accountability for the scholarly product, not on the medical or neurological status of the researcher.

8. A governance standard for accountable cognition

Accountable cognition can be translated into a practical governance standard. The researcher should retain control over the research question, the evidence hierarchy, the interpretation and the final judgement. The AI system may support the process, but it must not become the hidden authority behind the work.

In practical terms, the responsible researcher defines the role of AI before using it, classifies the data being handled, protects confidential or personal information, verifies factual claims and references, separates observed evidence from generated illustration, records material AI involvement where it affects the research process, tests the argument against counter-evidence, and ensures that the final output can be personally defended. These are not bureaucratic hurdles; they are the operating conditions of trustworthy AI-supported scholarship.

This standard is stricter than both blanket prohibition and careless adoption. Blanket prohibition ignores accessibility, real practice and the impossibility of returning to a pre-AI research environment. Careless adoption ignores hallucination, bias, privacy, dependency and institutional trust. Accountable cognition accepts that AI is now part of the scholarly environment, but insists that human responsibility must become more explicit, not less.

9. What this means for academia

Universities and journals should avoid two failures. The first is punitive purity: treating polished AI-assisted work as inherently suspect, especially when it comes from disabled, neurodivergent or non-native-language scholars. The second is permissive vagueness: allowing AI use without clear standards, training, disclosure norms or data-protection rules.

A better approach would include discipline-specific guidance, accessible AI-literacy training, clear examples of permitted and prohibited use, proportional disclosure rules, strong confidentiality warnings, and assessment models that require oral defence, process logs, annotated sources or reflective methodology statements where appropriate. The aim should not be to catch every use of AI. The aim should be to make responsible use visible and irresponsible use harder to defend.

Academic integrity in the AI era should be rebuilt around ownership, not nostalgia. The old ideal of unaided cognition was always partly fictional. Scholars have always worked through tools, institutions, editors, assistants, instruments and infrastructures. The future standard should ask whether the work is honest, accountable, verifiable and intellectually owned.

Conclusion: the case study is the argument

The day I became a case study was the day I realised that my use of AI was not peripheral to the ethics debate. It was central to it. My workflow shows why AI cannot be treated only as a threat to academic integrity. It can also be an accessibility technology, a cognitive scaffold and a route into fuller scholarly participation.

But that argument only holds if responsibility is preserved. AI assistance is legitimate when it strengthens access, clarity, verification and accountable judgement. It becomes ethically dangerous when it conceals evidence, fabricates authority, substitutes for understanding, exposes sensitive data, or allows the researcher to hide behind the machine.

The responsible academic future is not AI-free and it is not AI-surrendered. It is accountable cognition: human judgement supported by tools, disciplined by evidence, transparent about material assistance, and answerable for consequence. That is the standard by which AI use in research should be judged.

Author workflow disclosure

This article was produced through an AI-assisted but human-directed workflow. AI support was used for accessibility assistance, article structuring, language refinement, source-discovery prompts, revision planning, and conversion of editorial comments into specific amendments. The author retained responsibility for the argument, accepted or rejected suggested changes, checked the logic of the claims, and remained accountable for the final text. AI-generated material was not treated as empirical evidence, and synthetic or illustrative examples were not presented as observed data.

Selected references and source notes

Almgren Bäck, G., Lindeblad, E., Elmqvist, C. and Svensson, I. (2024) ‘Dyslexic students’ experiences in using assistive technology to support written language skills: a five-year follow-up’, Disability and Rehabilitation: Assistive Technology, 19(4), pp. 1217-1227. Published online 15 January 2023. DOI: 10.1080/17483107.2022.2161647.

Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P. and Roberts, K. (2024) Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. National Institute of Standards and Technology. DOI: 10.6028/NIST.AI.600-1.

European Commission, Directorate-General for Research and Innovation (2026) Living Guidelines on the Responsible Use of Generative AI in Research. Updated May 2026. Available at: https://research-and-innovation.ec.europa.eu/news/all-research-and-innovation-news/updated-era-living-guidelines-responsible-use-generative-ai-research-2026-05-08_en.

Kaabachi, B., et al. (2025) ‘A scoping review of privacy and utility metrics in medical synthetic data’, npj Digital Medicine, 8, Article 7. DOI: 10.1038/s41746-024-01359-3.

Perelmutter, B., McGregor, K.K. and Gordon, K.R. (2017) ‘Assistive technology interventions for adolescents and adults with learning disabilities: an evidence-based systematic review and meta-analysis’, Computers & Education, 114, pp. 139-163. DOI: 10.1016/j.compedu.2017.06.005.

UK Research and Innovation (2024) Generative Artificial Intelligence in Application and Assessment Policy. Published 20 September 2024; last updated 3 December 2024. Available at: https://www.ukri.org/publications/generative-artificial-intelligence-in-application-and-assessment-policy/.

UNESCO (2023, updated 2026) Guidance for Generative AI in Education and Research. Paris: UNESCO.

Zhao, S. (2025) ‘The use of generative AI by students with disabilities in higher education’, The Internet and Higher Education, 66, Article 101014. DOI: 10.1016/j.iheduc.2025.101014.

CONTACT

© 2026 Dr Danie Adendorff. All rights reserved. Rumbls.com is an independent analytical blog.

TERMS AND CONDITIONS

DISCLAIMER

COOKIE POLICY

PRIVACY POLICY