§ Master’s thesis · Faculty of Management, University of Warsaw · DIMA 2025–2027
Konsylium
Documenting, not detecting.
Universities try to detect AI-written text — and are losing that race. Konsylium asks a different question: not “was AI used?” but “how was AI used, and was it documented?”. It is an advisory project for the University of Warsaw and a working web app I designed and built.
Project status
- Thesisworking draft v13 · 225 pp.
- Field researchOct–Dec 2026
- DefenceMarch 2027
A race that can’t be won
AI-text detectors fail both ways: paraphrased text slips through and honest authors get falsely accused. An institution that bases sanctions on them takes a double risk — ethical and legal.
AI detectors tested proved reliable
None of the 14 tools — Turnitin included — reached 80% accuracy.
Weber-Wulff et al., 2023of essays by non-native writers wrongly flagged as AI-generated
61.3% on average across 7 detectors — while native writers’ essays were classified almost flawlessly.
Liang et al., 2023Three layers of one problem
- Layer 01
Operational
No tool, no procedure. Documentation duties are scattered, and widespread AI use goes undocumented.
- Layer 02
Legal
A double risk: doing nothing clashes with the spirit of the AI Act; over-monitoring breaches the GDPR.
- Layer 03
Cultural
An entrenched cat-and-mouse game between student and institution — to be replaced by openness.
We change the question.
Documenting, not detecting
The name refers to a medical consilium, where specialists decide together and in the open. In Konsylium the role of AI is disclosed, not hunted down.
Three arguments behind the shift
- Technical
Detection mechanisms fail.
- Legal
A register of AI use sits closer to AI Act and GDPR compliance than a grading tool.
- Pedagogical
Openness teaches the responsible use of AI.
The duty to document AI use already follows from University of Warsaw resolution URK no. 98/2023 and the DIMA guide — Konsylium doesn’t create it, it puts it into practice.
Six modules, one workflow
Each module tackles a different slice of the problem, so the faculty can roll it out step by step. The statuses are honest: some parts run in the prototype, others are mock-ups for later phases.
- Live
- Partial / early
- Mock-up
- 01
Onboarding & thesis editor
Guides the student through the process; an editor with a version register.
Status: Liveversion register: mock-up · F2Grounded inKim et al., 2025Where in the appProject dashboard; milestone schedule - 02
Document templates & AI declarations
Generates the appendix that documents AI use in the thesis.
Status: Partialdocument templates: F2Grounded inPérez-Pérez et al., 2026Chaudhry et al., 2022Where in the appDocumenting AI use (appendix export) - 03
Formal requirements validator
Automatically checks compliance with an 86-page set of requirements.
Status: Live19 rules from the catalogueGrounded inHalaweh, 2023Where in the appRequirements check - 04
Evidence repository
A versioned record of how the thesis came to be.
Status: Mock-upfull view: F2–F3Grounded inChan, 2023Where in the appA tile on the project dashboard - 05
AI use log
A declarative register kept by the student and exported to the appendix — instead of detection.
Status: Early versiondeclarative version: F2Grounded inLiang et al., 2023Chan, 2023Where in the appDocumenting AI use · in the prototype, the “AI audit” view - 06
Ethics & EBM
Decision support grounded in evidence (Evidence-Based Management).
Status: Assistant liveliterature review: F3Grounded inBond et al., 2024Alduais et al., 2025Where in the appLiterature review; built-in assistant
Statuses as listed in the thesis’ consistency matrix. Eight of the 41 reviewed publications feed the module functions directly — the rest form the argumentative backdrop of the whole concept.
- 26app views in 8 groups
- 19formal rules in the validator
- 10milestones with the supervisor
- 86pages of the DIMA guide turned into software
Five scenes from the app
The app’s UI is in Polish — translated here.
An assistant that helps but never ghost-writes
The built-in chat answers questions about methodology, requirements and the schedule. Asked to write a chapter, it refuses, cites Appendix B of the DIMA guide, records the attempt in the AI Audit module (prototype v1 — see scene 2) and offers ideation, theory or proofreading instead.
- Gemini 2.5 Flash
- allowed / needs documentation / forbidden
- in prototype v1 every prompt goes into the log
Click a suggestion — see how the prompt gets classified.
From audit to log
The prototype labels entries as “violations”. A self-review against Nielsen’s heuristics showed that this reproduces the detection logic the thesis rejects (severity 4). The target version (F2) is to be declarative: the student describes the use, the system stores no full prompts, and judging compliance is left to the supervisor.
Switch between the two versions — arrow keys work too.
AI use log
Planned after a heuristic self-reviewA validator for 19 formal requirements
Times New Roman 12 pt, 25 mm margins, 1.5 line spacing, length, bibliography, the individual workload form, the AI report… Green means: ready for the JSA scan (Poland’s national anti-plagiarism system).
- 19 formal rules
- DIMA guide · 86 pages
- ready for JSA
Run the scan again — the rules resolve one by one.
Thesis validator
Comments pinned to the content
Pins anchor to elements of the page, threads can be resolved, and notifications reach the in-app bell and arrive via Web Push.
- pins on page elements
- resolvable threads
- Web Push (VAPID)
Click the pin to open or close the thread.
10 milestones with the supervisor
The schedule follows Table 3 of the guide: every meeting has an input, an expected output and what the supervisor expects. Material reaches the supervisor at least 3 days ahead.
- Table 3 of the guide
- Feb 2026 → Feb 2027
- material ≥ 3 days ahead
Pick a meeting — arrow keys work too.
10 milestones with the supervisor
04 · Literature review
- Input
- ≥ 30 sources (incl. foreign-language ones: AI in education, EU AI Act)
- Output
- theoretical framework + final research problem
- The supervisor expects
- a theory that fits the problem
A real app, not a mock-up
Screenshots of the working prototype in the version prepared for the master’s thesis — from the project hub, through the team dashboard and the validator, to the presentation for the Dean.
Scroll to leaf through the stack. Click a screen to enlarge it.
§ 01The project hubThe way into the project: 12 tiles in three groups — product, strategy documents, research — and an assistant that knows the whole project. 
§ 02The team dashboardThesis progress, supervisor meetings, bibliography and repository in one row, with the next milestone and what the supervisor expects from it underneath. 
§ 03The Konsylium assistantA chat grounded in the DIMA guide. Its greeting already says it won’t write any part of the thesis, and it suggests questions for the current screen. 
§ 04The presentation for the Dean30 slides for the authorities of the Faculty of Management: the concept, a demo, the value for stakeholders and the research plan. 
§ 05Bibliography and simulators41 publications with bibliometric metrics, each with an interactive simulator — here, a model of how assessment design affects AI misuse. 
§ 06The design systemInk and paper, UW navy and a terracotta “§” — the same palette this page is set in. 
§ 07The validator on a phoneThe 19 formal requirements in the mobile layout, with a bottom bar for the key modules.
Three users, one source of truth
The team writes, the supervisor reviews, the partner observes. One source of truth instead of email attachments and five versions of a file.
- § 1Defining the research problemthe client, diagnosis, knowledge gap, research questions — in focus in this view
- § 2Literature reviewdetection, transparency, the EU AI Act and GDPR — in focus in this view
- § 3The research hypothesesH1–H5: unreliable detection, stakeholder attitudes — in focus in this view
- § 4Study design and methodologytriangulation, desk research, four field studies — in focus in this view
- § 5Data collection (fieldwork)Oct–Dec 2026 schedule, pilot, quality control — in focus in this view
- § 6Analysis of resultsH1–H2 in the literature, the data analysis plan — in focus in this view
- § 7Testing the hypothesesdecision rules, the status of H1–H5 — in focus in this view
- § 8The Konsylium app prototypeUX/UI design, usability research — in focus in this view
- § 9The legal and formal rollout pathlegal analysis, stakeholders, risk, economic model — in focus in this view
- § 10Conclusionssynthesis, limitations, recommendations for the university — in focus in this view
“As consultants, you stop being merely authors of text and become providers of decisions.”
Evidence first, code second
The diagnosis rests on a structured review of 41 publications and an analysis of the thesis process at the University of Warsaw’s Faculty of Management. The hypotheses are falsifiable — and for those the field studies will test, the decision rules were fixed before any data was collected.
The literature review
A structured review, with the entries verified against CrossRef and PubMed.
- Ethics & integrity1332%
- Assessment & teaching1332%
- Adoption & policy1229%
- AI literacy37%
A roughly even split — which keeps the diagnosis from being one-sided.
- 40 + 1peer-reviewed papers and one preprint
- 38 of 41from 2023–2026 — a current cohort
- 20 of 41in Top Tier or High Prestige journals
- 41interactive simulators in the hub — one per publication
Five falsifiable hypotheses
RQ1: Why detection doesn’t fit the problem
desk researchHypothesis 1
AI detectors falsely accuse non-native writers more often.
grounded in literatureH1 · MethodLiterature synthesis
CriterionConsistent findings from several independent studies, with no contrary results of comparable strength.EvidenceLiang et al., 2023 · Weber-Wulff et al., 2023not tested by a study of our ownHypothesis 2
Paraphrasing significantly lowers a detector’s recall.
grounded in literatureH2 · MethodLiterature synthesis
CriterionThe literature consistently reports lower recall after paraphrasing or “humanising” the text.EvidenceDalalah & Dalalah, 2023 · Weber-Wulff et al., 2023 · Eke, 2023not tested by a study of our own
RQ2: Stakeholders’ needs and attitudes
field studies · Oct–Dec 2026Hypothesis 3
Over 50% of DIMA students use GenAI for their thesis.
awaiting data · Oct–Dec 2026H3 · MethodCAWI + confidence interval
Decision ruleAccepted if the lower bound of the 95% confidence interval for the share of GenAI users exceeds 50%.InstrumentRES-CAWI-STD-01 · App. Cfixed before data collectionHypothesis 4
Students prefer documentation to detection and sanctions.
awaiting data · Oct–Dec 2026H4 · MethodCAWI + one-proportion test
Decision ruleAccepted if over 50% of respondents choose the transparency model and the lower bound of the 95% CI exceeds 0.5 (α = 0.05).InstrumentRES-CAWI-STD-01 · App. Cfixed before data collectionHypothesis 5
Supervisors see the lack of tools for managing AI as a practical problem.
awaiting data · Oct–Dec 2026H5 · MethodIDI + thematic analysis
Decision ruleAccepted if the code “no tools for managing AI” reaches thematic saturation, with inter-coder agreement of κ ≥ 0.70.InstrumentRES-IDI-PRO-01 · App. Bfixed before data collection
Rules before data
For H3–H5 the decision rules were set in advance, so the conclusions can’t be bent to fit expectations. H1–H2 are judged by the consistency of the literature: an open measure of evidence strength, not an a priori test.
The “Method” button turns a card over.
Four field studies
October–December 2026 · with statistical power analysisThe protocols are ready: each has a study plan — sampling, instrument, analysis plan, ethics, risks — and a full script. Fieldwork is timed so the material reaches decision gate B#1 in March 2027.
- pilot
- fieldwork
- buffer
- RES-CAWI-STD-01App. C
CAWI survey
DIMA students
Students’ experience, their use of AI and their response to the Konsylium concept.
- Sample
- n ≥ 100
- Duration
- 10–15 min
- Tests
- H3H4
- pilot 5–11 Oct
- fieldwork 19 Oct – 8 Nov
- buffer to 15 Nov
- RES-IDI-PRO-01App. B
In-depth interviews (IDI)
Supervisors
Supervisors’ workload, friction points in supervision, attitudes to AI and response to the prototype.
- Sample
- n = 6–10
- Duration
- 45–60 min · 1:1
- Tests
- H5
- pilot 5–11 Oct
- fieldwork 19 Oct – 8 Nov
- buffer to 22 Nov
- RES-EXP-B2B-01App. A
Expert interviews
B2B partners
Partners’ needs, barriers to cooperation, data quality and confidentiality (NDA).
- Sample
- n = 5–8
- Duration
- ~45 min · 1:1
- pilot 26 Oct – 1 Nov
- fieldwork 2–22 Nov
- buffer to 29 Nov
- RES-FGI-UCZ-01App. D
Focus group (FGI)
University representatives
Formal requirements and system scope as seen by the dean’s office, quality assurance, the DPO and IT.
- Sample
- 5–8 people
- Duration
- 60–90 min
- pilot 23–29 Nov
- fieldwork 7–13 Dec
- buffer to 20 Dec
Analytical toolkit and legal frame
- BMC
- PESTEL
- Porter’s Five Forces
- VRIO
- digital maturity
- GE matrix
- benchmarking
- EU AI Act
- GDPR
- Law on Higher Education and Science
- UW resolution URK no. 98/2023
- DIMA thesis guide
A drafted grant application: scientific rationale, aims and research plan.
application draftedSix phases, five decision gates
A stage-gate design: before money goes into the next phase, the previous one has to confirm its assumptions. After every gate the project can be deliberately stopped.
- to date
- planned
- decision gate (B#, Polish “bramka”)
- Jun 2026 – Mar 2027 We are here
Phase 1: Research & concept validation
The master’s thesis: field research (Oct–Dec 2026) → a requirements specification.
Gate B#1 · Mar 2027Problem–Solution Fitmeasures: indicators from the field research - Apr–Sep 2027
Phase 2: UX/UI & architecture, mock-ups
Mock-ups, a design system, usability tests with supervisors.
Gate B#2 · Sep 2027Mock-ups + PhDmeasures: how usable the mock-ups are - Oct 2027 – Jun 2028
Phase 3: Development & sandbox
An R&D grant: code, AI guardrails, a sandbox pilot with one DIMA cohort.
Gate B#3 · Jun 2028Proof of Conceptmeasures: adoption and compliance in the test environment - Jul 2028 – Dec 2029
Phase 4: Integration & scaling across UW
SSO/USOS, APD, JSA, EU AI Act certification, more faculties.
Gate B#4 · Dec 2029Market readinessmeasures: impact across the faculty - 2030–2031
Phase 5: Commercialisation (SaaS)
A spin-off company, SaaS for Polish universities (B2G/B2B).
Gate B#5 · 2031Commercial scalabilitymeasures: a validated, profitable commercial model - 2031+
Phase 6: European expansion
Multilingual support, Bologna frameworks, eduGAIN sign-in.
No gate — an execution stage
The goal: a University of Warsaw spin-off and a SaaS model for universities in Poland and the EU.
Phase 1 is under way — everything beyond it is a plan that every gate can still put to the test.
More than a prototype
Konsylium isn’t a single screen but a whole ecosystem in one repository: the thesis, the app, the presentation for the Dean, the implementation documents and a literature base.
The project in numbers
- 225pages of the thesis — working draft v13
- 42klines of TypeScript/TSX in the repository
- 30slides in the presentation for the Dean
- 16legal & implementation documents
- 16technologies and services in the stack
- 12hub tiles in 3 groups
Counted in the project repository and the working draft of the thesis.
Product
- 01Prototype — thesis
- 02Prototype — demo
- 03Design system
Strategy documents
- 04Presentation for the Dean
- 05Business problem
- 06Rollout path
Research
- 07Bibliography
- 08Literature review
- 09PRELUDIUM grant
- 10Study protocols
- 11DIMA guide
- 12Master’s thesis
Architecture
- Hub12 tiles and a RAG assistant up front
- The app26 views in 8 groups
- APIAI assistant, RAG, comments, Web Push, PDF export
Tech stack
- Next.js 15 (App Router)
- React 19
- TypeScript
- Tailwind CSS v4
- shadcn/ui
- Neon Postgres
- Drizzle ORM
- pgvector
- Vercel AI SDK
- Gemini 2.5 Flash
- gemini-embedding-001 (768-d)
- RAG
- JWT (jose)
- Web Push (VAPID)
- PDFKit
- Vercel
From concept to code
My role
Jakub Nowak- Information architecture: the project hub and the app’s 26 views.
- UX/UI design and a “manuscript” design system — one language for the app, the documents and every project material.
- Full-stack build: Next.js, Postgres with pgvector, a RAG assistant, comments with Web Push, PDF export.
- Co-authorship of the thesis (50/50 share), leading the legal part and the prototype.
- Design and build of the deck for the faculty’s leadership (30 slides), prepared together with Maksymilian Cała.
Authorship
Authors
- Supervisor
- Dr Edyta Barmentloo
- Faculty
- Faculty of Management, University of Warsaw · Digital Marketing (DIMA) · 2025–2027
- Thesis title
- Advisory project for the University of Warsaw: modern teaching tools as a response to the need for transparency in the use of artificial intelligence in higher education
- 1 · Research problemJakub Nowak 60%, Maksymilian Cała 40%
- 2 · Literature reviewJakub Nowak 45%, Maksymilian Cała 55%
- 3 · HypothesesJakub Nowak 55%, Maksymilian Cała 45%
- 4 · MethodologyJakub Nowak 40%, Maksymilian Cała 60%
- 5 · Data collectionJakub Nowak 45%, Maksymilian Cała 55%
- 6 · AnalysisJakub Nowak 40%, Maksymilian Cała 60%
- 7 · Hypothesis testingJakub Nowak 50%, Maksymilian Cała 50%
- 8 · The Konsylium prototypeJakub Nowak 60%, Maksymilian Cała 40%
- 9 · Legal & formal pathJakub Nowak 65%, Maksymilian Cała 35%
- 10 · ConclusionsJakub Nowak 50%, Maksymilian Cała 50%
- OverallJakub Nowak 50%, Maksymilian Cała 50%
source: the workload form, working draft v13
The thesis follows its own rule: Appendix E openly states what we used GenAI tools for — among other things, organising the literature, language editing and analysis code. The concept, hypotheses, conclusions and recommendations are entirely the authors’ own.
Evidence, not opinion.
I’d be glad to demo the app live and talk through the design decisions.
§ Konsylium · DIMA · Faculty of Management, University of Warsaw · 2025 / 2027
Back to top
§ 2.3 From detection to transparency