Legal position last verified: 3 August 2026.
An organisation’s first material AI risk may not arrive through a strategic technology procurement. It may begin when an employee pastes a customer file, workforce assessment, contract draft or source code into a publicly accessible generative AI service. Türkiye’s Personal Data Protection Authority describes this uncontrolled use as “Shadow AI”: the use of generative AI outside organisational knowledge, approval and control, leaving the business unable to see which tools, data and outputs are shaping its operations.1
The commercially sensible response is not a blanket ban. The Authority’s February 2026 workplace guidance warns that prohibition-only models may push use outside corporate visibility and instead recommends clear direction, balanced controls, awareness and managed access.2 Recent organisational research reaches a similar conclusion: Shadow AI is a socio-technical governance failure, and practical countermeasures include AI tool registers, role-specific training, internal audit and escalation routes that enable controlled use rather than merely suppressing it.3
The board should not ask only, “Do we use AI?” It should ask, “Which business decision uses which data, through which supplier, under whose accountability, with what testing and with what genuine human control?”
This playbook gives companies a practical route from initial idea and procurement to contracting, pilot approval, production deployment and incident response. Its red–amber–green screening model is an internal prioritisation tool; it is not a substitute for formal legal classification under the EU Artificial Intelligence Act (“AI Act”).
The 2026 position: which obligations matter now?
| Issue | Position as at 3 August 2026 | Business consequence |
|---|---|---|
| Türkiye’s KVKK | Law No 6698 applies in a technology-neutral manner whenever an AI use involves personal data.4 | Inputs, outputs, logs, accounts, feedback and integrations must be assessed separately. |
| GDPR | Where the organisation or processing falls within the GDPR, it applies alongside the AI Act; AI Act compliance does not displace data protection law.5 | A supplier’s “AI Act compliant” claim does not establish a lawful basis, transparency or transfer compliance. |
| AI Act territorial scope | The Act may apply to providers placing systems on the EU market, deployers located in the EU and certain third-country providers or deployers where system output is used in the EU.6 | Establishment in Türkiye does not, by itself, guarantee that the organisation is outside scope. |
| AI literacy | Following amendments in force from 27 July 2026, providers and deployers must take measures supporting the AI literacy of staff and other people operating or using systems on their behalf, having regard to context and affected persons.7 | A generic annual presentation is not a complete programme; evidence should be role- and use-case-specific. |
| Article 50 transparency | Interaction notices, machine-readable marking and defined disclosure duties for synthetic content apply from 2 August 2026.8 | Chatbots, synthetic media and corporate publishing workflows need immediate review. |
| Standalone high-risk systems | The requirements linked to Article 6(2) and Annex III apply from 2 December 2027 following Regulation (EU) 2026/1744.9 | Hiring, credit and other consequential systems should not be ignored; inventory and evidence should be built before the deadline. |
| Product-safety high-risk systems | Requirements linked to Article 6(1) and Annex I apply from 2 August 2028.10 | Businesses with long product cycles should start technical-file and supply-chain work now. |
| Limited legacy transition | Systems placed on the market before 2 August 2026 receive a transition only for the Article 50(2) machine-readable marking duty, until 2 December 2026.11 | It is not a general grace period for chatbot notices or deployer disclosures. |
1. The first 48 hours: create visibility before drafting a long policy
Even a company with no AI committee, inventory or formal policy can establish six interim controls within two business days:
- Appoint one accountable executive owner. The matter needs a named person who can join legal, information security, procurement and business decisions.
- Issue a temporary input rule. Customer files, workforce data, trade secrets, source code, financial data, health data and live legal matters must not be entered into unapproved public tools.
- Publish an approved-tool list. Replace “everything is prohibited” with a clear statement of which tool is permitted for which task.
- Open a simple use-case declaration form. An employee should be able to disclose the tool, purpose, data and effect on a decision in five minutes.
- Create a no-blame rapid reporting route. People should report accidental uploads or harmful outputs promptly rather than conceal them.
- Pause consequential automated outcomes. AI should not produce a final hiring, dismissal, promotion, remuneration, credit, insurance, health or legal outcome before review.
These are not a permanent governance programme. They are containment controls that reveal what the organisation is already doing and reduce unmanaged data flows.
2. Build an AI use-case register—not merely a product list
The question “Which AI products have we purchased?” is too narrow. The same service may be low impact when editing public copy and highly consequential when ranking applicants. The inventory unit should therefore be the use case, not the tool.
Minimum AI use-case record
| Field | Evidence to record |
|---|---|
| Business owner | Executive and function accountable for the outcome |
| Purpose | Measurable business need; “using AI” is not a purpose |
| Affected people | Employees, applicants, customers, suppliers, children or other groups |
| Inputs | Personal data, special-category data, trade secrets, source code or public material |
| Outputs | Draft, score, recommendation, classification, decision, media or code |
| Decision effect | Information support, recommendation, strong influence or de facto final outcome |
| System/version | Product, foundation model, version, integration and hosting mode |
| Supply chain | System provider, model provider, subprocessors and support access |
| Legal roles | KVKK/GDPR controller–processor roles and AI Act provider–deployer–importer roles |
| Locations | Hosting, support, remote-access and onward-transfer countries |
| Retention/training | Retention of prompts, outputs and logs; use for training or improvement |
| Human oversight | Reviewer, authority, review criteria and ability to override |
| Testing | Accuracy, discrimination, security, privacy, hallucination and misuse tests |
| Legal records | Notice, contract, DPIA, TIA, risk decision and relevant approvals |
| Status | Idea, review, pilot, approved, suspended or retired |
| Reassessment | Trigger date for changes to version, purpose, data, country or subprocessor |
The register should not become a static spreadsheet. A model update, new subprocessor, changed data-use term, broader integration or increased decision influence should reopen the record. Scholarship conceptualising generative AI as a complex adaptive socio-technical system likewise supports continuing feedback and re-evaluation rather than one-off documentation.12
3. An internal red–amber–green screening model
This model helps legal and risk teams allocate effort. It is not the AI Act’s formal legal classification. A separate assessment must determine whether a practice is prohibited, high-risk or subject to other duties.
Red: do not start before specialist approval
- applicant ranking, recruitment, promotion, performance, pay, scheduling, dismissal or discipline;
- creditworthiness, essential-service access, insurance or economic-opportunity scoring;
- personalised health, legal or safety determinations;
- biometric identification or sensitive biometric categorisation;
- solely automated decisions producing legal or similarly significant effects;
- systems intended to infer employees’ emotions;
- use involving children, health information, criminal-conviction data or intensive workforce monitoring;
- entry of personal data, trade secrets or customer files into public generative AI services.
The use of AI to infer emotions in workplaces is prohibited under Article 5 of the AI Act, subject to the narrow medical or safety exception.13 A supplier label such as “engagement analytics” or “interview sentiment” does not change the functional assessment.
Amber: proceed only through a controlled pilot
- customer-service chatbots;
- meeting recording, transcription and summarisation;
- summarisation of customer or employee documents in a controlled enterprise environment;
- personalised marketing or proposal recommendations;
- coding assistants used in production development;
- fraud, security-event or contract-risk alerts;
- corporate text, image, audio or video generation;
- supplier or customer prioritisation.
The pilot should use limited data, avoid final adverse outcomes, include validation and preserve a rollback route.
Green: simplified controls may be proportionate
- summarising public material with no personal or confidential data;
- language and style editing;
- generating anonymous training examples;
- general brainstorming;
- synthetic test data disconnected from real systems;
- internal drafts containing no personal data or trade secrets.
“Green” does not mean zero risk. Copyright, factual accuracy, reputation and Article 50 transparency may still matter.
4. Twenty-five questions to ask before signing
A security certification or an “enterprise” label is not a full due-diligence response. The procurement file should contain written and verifiable answers to the following questions.
A. System, purpose and roles
- What is the precisely stated intended purpose, and which uses are prohibited?
- Which legal entity provides the service, foundation model, hosting and support?
- Is the customer a deployer under the AI Act, and could customisation make it a provider or another operator?
- How has the supplier classified the system under the AI Act, and what evidence supports that conclusion?
- Will the supplier give advance notice of a model, feature or material architecture change?
B. Data and privacy
- For which separate purposes are prompts, files, outputs, feedback, telemetry and account data processed?
- Are any of those data used for training, fine-tuning, quality review or human evaluation?
- Is the no-training commitment contractually binding and extended to foundation-model providers and subprocessors?
- What is the retention period for each data category, and when are backups deleted?
- What are the parties’ controller/processor roles and the lawful basis for each purpose?
- In which countries are data stored or remotely accessed, and is there an up-to-date subprocessor list and objection mechanism?
- What technical support is available for access, correction, erasure, objection and automated-decision requests?
- How is customer data separated from other tenants?
C. Security and administrative control
- Does the product support SSO, MFA, role-based access, SCIM, DLP, IP restrictions and administrator logs?
- How is data encrypted in transit and at rest, and who controls the keys?
- Can prompts, outputs and audit events be exported for investigation?
- What vulnerability-management, penetration-testing and independent assurance evidence exists?
- Within how many hours will a personal data breach or serious AI incident be notified, and what will the notice contain?
- How are plugins, agents, APIs and third-party connectors permissioned and restricted?
D. Quality, oversight and rights
- What accuracy, error, hallucination and discrimination results exist for the customer’s intended use?
- Where are known limitations, unsuitable contexts and performance deterioration documented?
- To what extent can the system explain sources, main factors or the basis of an output?
- Are logs sufficient to support human review, override, challenge and correction?
- Are Article 50 interaction and machine-readable marking functions available?
- On exit, how will data, configurations, logs and evaluation evidence be exported, and how will deletion be verified?
A supplier’s blanket reliance on “trade secrets” is a particular warning sign for systems that score or decide about people. In Dun & Bradstreet Austria, the Court of Justice held that “meaningful information about the logic involved” under GDPR Article 15 must enable a person to understand and challenge the automated procedure; trade-secret interests do not automatically remove the information from access and may require controlled disclosure to an authority or court.14
5. The minimum AI contract schedule
A standard data processing addendum does not necessarily address model and output risk. Read together, the order form, main agreement, data processing terms and security schedule should cover at least:
| Clause | Minimum contractual result |
|---|---|
| Purpose and permitted use | Defined use case, users, data and prohibited activity |
| Data use | No training/improvement use of prompts, files and outputs, or a precisely bounded alternative |
| Confidentiality/trade secrets | Binding protection across supplier personnel, model providers and subprocessors |
| Roles and instructions | Separation of controller/processor purposes and prohibition of off-instruction use |
| Subprocessors/countries | Current list, advance notice, objection, onward-transfer mechanism |
| Security | Access, encryption, logging, tenant separation, vulnerability and secure-development controls |
| Incident notice | Time, content and cooperation for breaches, harmful model behaviour and serious incidents |
| Model/system change | Notice for model, version, hosting, data-use or material performance change |
| Testing/performance | Use-case acceptance criteria, disclosed limitations and error notification |
| Human oversight | Technical support for explanation, records, override, challenge and correction |
| AI Act cooperation | Role/classification evidence, Article 50 features and technical-file assistance where relevant |
| Intellectual property | Allocation for inputs, outputs, training material, licences and third-party claims |
| Audit | Documentary assurance, targeted review and remediation rights |
| Suspension | Right to stop the integration where legality, security or fundamental-rights risk arises |
| Exit/deletion | Portability, configuration/log export and verifiable deletion including backups |
Research on AI outsourcing governance supports a combined model: measurable formal controls should sit alongside information sharing, collaboration and change management rather than relying exclusively on either a long contract or informal relationship management.15
6. The data protection gateway: AI Act analysis is not enough
The AI Act expressly preserves EU data protection law.16 A conclusion that a system is “not high-risk” does not establish that its personal data processing is lawful.
Before production, the business should answer:
- Which personal data enter or emerge from the system? Map prompts, outputs, logs, feedback, conversation history, device and account data.
- What is the purpose and processing condition? Assess KVKK Articles 5 and 6 and, where applicable, GDPR Articles 6 and 9.
- Is minimisation real? “The model works better” is not, by itself, a reason to upload an entire customer or employee file.
- Is transparency intelligible? Affected people should understand the AI-supported processing and its material consequences.
- What role does the supplier play? Security, product development or model-improvement purposes may create separate controllership.
- Is there an international transfer? Consider hosting, support access, model APIs, subprocessors and logging platforms.
- Is an impact assessment required? GDPR Article 35 requires a DPIA for processing likely to produce high risk, particularly with new technology; documented impact assessment is also sound accountability practice under the KVKK.17
- Is the model genuinely anonymous? EDPB Opinion 28/2024 requires a case-specific assessment of whether personal data can be extracted from the model or obtained through queries.18
- Can individual rights be actioned? There must be a technical route for access, correction, erasure, objection and automated-assessment issues.
- Where does deletion occur? Cover the primary store, chat history, security logs, backups, evaluation sets and supplier support copies.
7. The production gate: fifteen pieces of evidence
The following evidence should exist before an AI system is allowed to affect a live business process:
- approved use case and accountable owner;
- data-flow diagram;
- operator and privacy-role matrix;
- lawful-basis and transparency analysis;
- international-transfer and subprocessor review;
- DPIA, TIA or equivalent assessment where required;
- supplier responses and executed schedules;
- provenance and lawful use of test data;
- accuracy, error and hallucination results;
- discrimination and group-performance analysis;
- evidence of security, access and logging configuration;
- human oversight procedure;
- user training and policy acknowledgement;
- incident and kill-switch route;
- reassessment date and change triggers.
A language-editing service can use a lighter gate; an applicant-scoring system requires greater evidence. Risk-based governance means increasing control with impact, not forcing every use case through an identical file.19
8. Genuine human oversight is more than clicking “approve”
A meaningful reviewer:
- understands the intended purpose and known limitations;
- can access the context and inputs needed to make the decision;
- has sufficient time and authority to assess independently;
- can disagree with and override the output;
- is not incentivised to rubber-stamp recommendations;
- records the reason for intervention or acceptance;
- can reconsider the result following a challenge.
The Turkish Authority’s workplace guidance similarly recommends treating generative AI output as decision support rather than the final basis of a decision, with human review for accuracy, context and discriminatory effects.20 Empirical literature on automation bias shows why this matters: people can become complacent and accept automated recommendations without adequate scrutiny.21
In SCHUFA, the Court of Justice held that creating a credit score can itself constitute an automated decision under GDPR Article 22 where a third party attributes a determining role to that score.22 A company should therefore look at operational reality, not merely say that “a human signs the final form”. Under Türkiye’s KVKK Article 11(1)(g), a data subject also has the right to object where analysis exclusively through automated systems produces an adverse result.23
9. The one-page employee rule
A detailed policy may sit behind the programme, but the rule shown to employees should be concise and operational:
- Use only approved corporate accounts and tools.
- Do not enter customer, employee or third-party data into a public AI service.
- Do not use health, financial, legal, criminal-conviction or identity data without specific approval.
- Do not upload trade secrets, source code, product roadmaps or confidential contracts.
- Anonymise, generalise and remove unnecessary detail wherever possible.
- Never assume AI output is true; verify sources and context.
- Do not use an AI output alone to score a person or produce an adverse decision.
- Check the disclosure rule before publishing AI-generated text, image, audio or video.
- Report suspicious output, discrimination, leakage or accidental upload immediately.
- Do not install a new AI tool, plugin, agent or integration without approval.
The Authority recommends that corporate guidance define approved tools, purposes, permitted inputs, output-use rules, security and training, and it cautions against sharing personal or sensitive corporate material with public services.24
10. AI literacy: train for the role, not for attendance
The current Article 4 requires measures supporting AI literacy with regard to the users’ knowledge, experience, context and people affected.25 A single annual video is therefore not a mature response.
| Role | Minimum learning outcome | Evidence |
|---|---|---|
| Board/executive team | risk appetite, accountability, critical uses, dashboard | decision and periodic review |
| Procurement/legal | role analysis, supplier questions, contracts, transfers, AI Act scope | completed checklist and sample file |
| IT/security | access, agent/API permissions, logs, DLP, model change, response | technical test and configuration evidence |
| HR | hiring/workforce decisions, discrimination, automation and challenge | scenario-based assessment |
| Marketing/comms | accuracy, copyright, brand, synthetic content and Article 50 | publication control |
| End user | approved tools, input rules, hallucination, verification, reporting | short test and acknowledgement |
Do not measure success only through attendance. Track unsafe-input events, unapproved use, reporting time, validation failures and adherence to the production gate.
11. Article 50: a practical customer and publishing checklist
From 2 August 2026, businesses need to distinguish four issues.
11.1 Direct interaction with an AI system
A provider must design a system so that a person is informed when directly interacting with AI unless this is obvious to a reasonably well-informed, observant and circumspect person.26 A company building or materially rebranding a chatbot should assess whether it has become a provider.
Operational step: Give a clear notice at the beginning of the interaction and provide a route to human assistance for material matters.
11.2 A provider generates synthetic content
Article 50(2) requires providers of systems generating synthetic audio, image, video or text to mark outputs in a machine-readable format and make them detectable. The limited transition until 2 December 2026 applies only to systems placed on the market before 2 August 2026 and only to this marking duty.27
Operational step: Ask the supplier which provenance standard is used, which formats retain it and whether platform conversion strips the marker.
11.3 A deployer uses deepfakes or emotion/biometric systems
A deployer using deepfake image, audio or video must disclose that the content was artificially generated or manipulated. People exposed to emotion-recognition or biometric-categorisation systems must also receive information, while the separate workplace emotion-inference prohibition must be checked first.28
11.4 A deployer publishes AI text on a matter of public interest
A deployer publishing AI-generated or manipulated text to inform the public on a matter of public interest must disclose that fact. The specific exception may apply where the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for the publication.29
Operational step: “Someone looked at it” is not a defensible workflow. Record the editor, verification step and entity accepting publication responsibility.
12. The first 24 hours of an AI incident
An AI incident is not limited to a personal data breach. It may include discriminatory applicant ranking, an incorrect credit refusal, confidential code sent outside the business, fabricated legal sources, misleading customer communication or an agent taking an unauthorised action.
Response sequence
- Stop and contain: Disable the integration, API key, automated publication or decision flow.
- Preserve evidence: Retain prompts, outputs, model version, user, timestamps, logs, data set and supplier notices.
- Scope impact: Identify affected people, data, countries, decisions and customer outputs.
- Engage the supplier: Request deletion, logs, model-behaviour and subprovider information.
- Assess legal triggers: Consider KVKK/GDPR breach notification, contractual notice, AI Act serious-incident or sectoral duties.
- Correct outcomes: Re-review decisions, correct publications and provide human reconsideration.
- Find the root cause: Determine whether policy, training, access, contract, model or oversight failed.
- Prevent recurrence: Update the rule, technical control, training and register entry.
A punishment-led reporting culture hides Shadow AI. Controlled enablement treats prompt reporting as a defence layer.
13. A 30–60–90-day implementation plan
Days 1–30: visibility and containment
- appoint an executive sponsor and cross-functional AI working group;
- publish the interim rule and approved-tool list;
- collect existing tools and use cases;
- pause red uses or return them to human decision;
- add supplier questions to procurement;
- open the incident route;
- check chatbots and content workflows for Article 50.
Deliverable: first inventory, red-risk list, interim policy and responsibility matrix.
Days 31–60: contracts, data and testing
- review priority suppliers and contract terms;
- map data, roles, countries and retention;
- complete notices, processing conditions, DPIAs/TIAs and data schedules where needed;
- run accuracy, discrimination, security and misuse tests;
- begin role-based training;
- document human oversight and challenge.
Deliverable: completed review files, contract remediation and test reports.
Days 61–90: permanent governance and measurement
- make the production gate mandatory;
- connect the register to procurement, security and project management;
- monitor model and subprocessor changes;
- launch the board dashboard;
- run an incident exercise and sample data-subject request;
- decide which uses to retire, remediate or scale.
Deliverable: an operating governance cycle, management report and update calendar.
14. Eight board-level metrics
The board needs more than the number of AI licences purchased. Report:
- registered use cases and detected unregistered uses;
- red–amber–green distribution;
- systems reaching production before legal/security review;
- scenario-test performance as well as training completion;
- incidents involving personal or confidential data in unapproved tools;
- average detection and containment time;
- unresolved critical supplier-contract gaps;
- AI recommendations changed, rejected or challenged by humans.
These metrics distinguish “using more AI” from producing measurable value under control.
15. Three practical scenarios
Scenario 1 — Sales-call summarisation
Calls may contain personal data and commercial information. A controlled enterprise environment, contractual no-training commitment, restricted retention, access controls and appropriate transparency should be assessed. If output is written automatically into the CRM, human review and correction are needed.
Scenario 2 — CV ranking
The use can affect access to employment. Data provenance, discrimination testing, applicant transparency, automated-decision limits, genuine reviewer authority and a challenge route are central. Calling the output a “recommendation” does not settle the issue if it is operationally decisive.
Scenario 3 — AI-assisted public report
In addition to source validation, copyright and brand review, Article 50(4) may apply. The editorial-control exception should be supported by actual review, an identified editor and accepted publication responsibility; otherwise the AI use should be disclosed.
Frequently asked questions
Is a complete ban on ChatGPT and similar services the safest corporate approach?
Generally not. Türkiye’s 2026 workplace guidance warns that blanket prohibitions may push use outside organisational visibility. A stronger model combines approved tools, permitted purposes, prohibited data, access control, training and rapid reporting.
Can a company established in Türkiye fall within the EU AI Act?
Yes. The Act may apply to providers placing a system on the EU market, deployers located in the EU and certain third-country providers or deployers where output is used in the EU. The company’s role, market and output location require case-specific analysis.
Is a supplier statement that customer data is not used for training sufficient?
No. The business should verify which prompts, files, outputs, feedback and logs the promise covers; whether it binds model providers and subprocessors; whether retention or human review continues; and whether it is contractual.
Does a human final approval remove automated-decision risk?
Not always. Where a reviewer routinely confirms the system or the output is de facto determinative, operational reality matters. The reviewer needs knowledge, time, authority, independent judgement and a genuine ability to override.
Is a DPIA mandatory for every AI project?
No. GDPR Article 35 turns on whether the processing, particularly through new technology, is likely to result in high risk to people’s rights and freedoms. A documented initial screening for every use case is nevertheless sound evidence of how the DPIA decision was reached.
Must every AI-assisted corporate text be labelled as AI-generated?
No. Article 50 distinguishes roles and content. Deepfakes, certain public-interest publications and system-level machine-readable marking require separate analysis. The human review/editorial control and editorial-responsibility exception for defined public-interest text should be applied to the actual workflow.
Conclusion: the strongest policy is an evidence system, not a ban
Enterprise AI is no longer only a technology-purchasing decision. Customer relations, employment, privacy, trade secrets, intellectual property, cyber security and decision quality can converge in a single use case.
Effective governance does three things at once:
- it does not bury beneficial low-impact uses in unnecessary bureaucracy;
- it applies a strong gate where people’s rights or critical corporate assets are affected;
- it can later demonstrate which data, system, supplier, test and human judgement produced the outcome.
The starting point is not a hundred-page policy. It is a use-case register, supplier questions, minimum contract schedule, production gate and rapid reporting route. A company that makes those elements operational within 90 days creates both a clearer route to innovation and a defensible accountability record under the KVKK, GDPR and AI Act.
Legal information notice
This article provides general information only. It is not a legal opinion, data protection impact assessment, fundamental-rights impact assessment or technical security assessment for any specific system or organisation. Legal treatment depends on the company’s role, intended and actual use, data, affected people, jurisdictions, sector rules and current official law.
Bibliography
Primary and official materials
- Law No 6698 on the Protection of Personal Data.
- Turkish Personal Data Protection Authority, Use of Generative Artificial Intelligence Tools in Workplaces (February 2026; Turkish original).
- Regulation (EU) 2016/679 (GDPR).
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated version of 27 July 2026.
- Regulation (EU) 2026/1744.
- European Commission, Guidelines on Transparency of AI-Generated Content (20 July 2026).
- European Data Protection Board, Opinion 28/2024.
- OQ v Land Hessen (SCHUFA Holding AG) (C‑634/21) EU:C:2023:957.
- CK v Magistrat der Stadt Wien (Dun & Bradstreet Austria) (C‑203/22) EU:C:2025:117.
Academic literature
- Bartels M, ‘A Balancing Act: Data Protection Compliance of Artificial Intelligence’ (2024) 73(6) GRUR International 526.
- Beulen E, Plugge A and van Hillegersberg J, ‘Formal and Relational Governance of Artificial Intelligence Outsourcing’ (2022) 20 Information Systems and e-Business Management 719.
- Ebers M, ‘Truly Risk-based Regulation of Artificial Intelligence: How to Implement the EU’s AI Act’ (2025) 16(2) European Journal of Risk Regulation 684.
- Janssen M, ‘Responsible Governance of Generative AI: Conceptualizing GenAI as Complex Adaptive Systems’ (2025) 44(1) Policy and Society 38.
- Parasuraman R and Manzey DH, ‘Complacency and Bias in Human Use of Automation: An Attentional Integration’ (2010) 52(3) Human Factors 381.
- Silic M, Silic D and Kind-Trüller K, ‘From Shadow IT to Shadow AI—Threats, Risks and Opportunities for Organizations’ (2025) Strategic Change 1.
Footnotes
Turkish Personal Data Protection Authority, Use of Generative Artificial Intelligence Tools in Workplaces (February 2026) 5–10 https://www.kvkk.gov.tr/SharedFolderServer/CMSFiles/MTY5YTdkNjdjNzJlMjM.pdf accessed 3 August 2026 (Turkish original).↩︎
ibid 10–15.↩︎
Mario Silic, Dario Silic and Kathrin Kind-Trüller, ‘From Shadow IT to Shadow AI—Threats, Risks and Opportunities for Organizations’ (2025) Strategic Change 1 https://doi.org/10.1002/jsc.2682.↩︎
Law No 6698 on the Protection of Personal Data https://www.mevzuat.gov.tr/mevzuatmetin/1.5.6698.pdf; Turkish Personal Data Protection Authority (n 1) 9–10.↩︎
Regulation (EU) 2016/679 (General Data Protection Regulation) [2016] OJ L119/1 https://eur-lex.europa.eu/eli/reg/2016/679/oj; Regulation (EU) 2024/1689 (Artificial Intelligence Act), art 2(7) and recital 10, consolidated version of 27 July 2026 https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727 accessed 3 August 2026.↩︎
Artificial Intelligence Act, art 2(1).↩︎
Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 simplifying the implementation of harmonised rules on artificial intelligence [2026] OJ L, art 1; Artificial Intelligence Act, art 4, consolidated version (n 5) https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R1744.↩︎
Artificial Intelligence Act, art 50; European Commission, ‘Guidelines on Transparency of AI-Generated Content’ (20 July 2026) https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content accessed 3 August 2026.↩︎
Regulation (EU) 2026/1744 (n 7), amending Artificial Intelligence Act art 113.↩︎
ibid.↩︎
ibid; European Commission, ‘Transparency Obligations under Article 50 of the AI Act’ (July 2026) https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act accessed 3 August 2026.↩︎
Marijn Janssen, ‘Responsible Governance of Generative AI: Conceptualizing GenAI as Complex Adaptive Systems’ (2025) 44(1) Policy and Society 38 https://doi.org/10.1093/polsoc/puae040.↩︎
Artificial Intelligence Act, art 5(1)(f).↩︎
CK v Magistrat der Stadt Wien (Dun & Bradstreet Austria) (C‑203/22) EU:C:2025:117, paras 60–71 and operative part https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62022CJ0203.↩︎
Erik Beulen, Albert Plugge and Jos van Hillegersberg, ‘Formal and Relational Governance of Artificial Intelligence Outsourcing’ (2022) 20 Information Systems and e-Business Management 719 https://doi.org/10.1007/s10257-022-00562-7.↩︎
Artificial Intelligence Act, art 2(7) and recital 10.↩︎
GDPR art 35; Marvin Bartels, ‘A Balancing Act: Data Protection Compliance of Artificial Intelligence’ (2024) 73(6) GRUR International 526 https://doi.org/10.1093/grurint/ikae060.↩︎
European Data Protection Board, ‘Opinion 28/2024 on Certain Data Protection Aspects Related to the Processing of Personal Data in the Context of AI Models’ (18 December 2024) paras 29–50 https://www.edpb.europa.eu/documents/opinion-of-the-board-art-64/opinion-282024-on-certain-data-protection-aspects-related-to_en accessed 3 August 2026.↩︎
Martin Ebers, ‘Truly Risk-based Regulation of Artificial Intelligence: How to Implement the EU’s AI Act’ (2025) 16(2) European Journal of Risk Regulation 684 https://doi.org/10.1017/err.2024.78.↩︎
Turkish Personal Data Protection Authority (n 1) 13–14.↩︎
Raja Parasuraman and Dietrich H Manzey, ‘Complacency and Bias in Human Use of Automation: An Attentional Integration’ (2010) 52(3) Human Factors 381 https://doi.org/10.1177/0018720810376055.↩︎
OQ v Land Hessen (SCHUFA Holding AG) (C‑634/21) EU:C:2023:957, paras 43–73 https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62021CJ0634.↩︎
Law No 6698, art 11(1)(g).↩︎
Turkish Personal Data Protection Authority (n 1) 11–15.↩︎
Artificial Intelligence Act, art 4, consolidated version (n 5).↩︎
Artificial Intelligence Act, art 50(1).↩︎
Artificial Intelligence Act, art 50(2); Regulation (EU) 2026/1744 (n 7).↩︎
Artificial Intelligence Act, arts 50(3)–(4) and 5(1)(f).↩︎
Artificial Intelligence Act, art 50(4).↩︎
