Appendix B – Data Security and Privacy in the World of Artificial Intelligence

Purpose and context

This document is an appendix and a detailed elaboration of the principles set out in the "Guiding Policy on the Responsible and Ethical Use of Artificial Intelligence at the University of Akureyri" (hereinafter referred to as the "Principal Policy").

The main policy allows the use of AI tools if strict conditions are met on privacy, secure data handling, copyright, and academic integrity.

The purpose of this document is to provide the in-depth, technical and legal analysis necessary to interpret and enforce these conditions in an ever-changing technological environment. The document is intended for the entire academic community, students, teachers, researchers and administrative staff, as a fundamental guide to the safe and responsible use of this revolutionary technology. The document will be updated as needed, but is intended to reflect and ensure that users are aware and informed about the requirements for the use of AI and how different services handle the user's data.

An ever-changing risk landscape

An analysis of the top AI solutions demonstrates the fundamental risks associated with the use of large language models (LLMs). That risk lies in the fact that users' data is sent for processing on service providers' servers, which is a fundamental difference from traditional, local software. The pace of development in this field is unprecedented, and new models regularly appear on the scene, in addition to which the University of Akureyri (UNAK) has made certain tools available to the entire academic community. These developments call for a deeper and more comprehensive analysis that shifts the debate from general principles to a specific assessment supported by evidence at individual service providers and service levels.

Two cornerstones of true data security

The analysis in this document reveals a crucial conclusion: true data security in the world of generative artificial intelligence is not determined by branding, model performance, or whether the service is paid for. It is based on two inalienable pillars:

1. Legally binding contract terms: Enterprise solutions that ensure, in a clear and legally binding manner, that customer data is never used to train or improve the service provider's models.
2. Jurisdiction and data sovereignty: In which country and under which legislation the service provider operates and where users' data is stored. For a European university subject to the EU General Data Protection Regulation (GDPR), this is essential.

These two factors will serve as a common thread throughout the analysis and form the basis for the risk categories and practical recommendations presented.

Free consumer versions:

In this model, the user is not the customer, but rather is part of the product. The data that a user enters – questions, snippets, ideas – is the currency that is paid for the service. Service providers such as OpenAI (for free ChatGPT) and Google (for Gemini) reserve the right to use this data to train, refine, and develop their models. The default setting is almost without exception that data collection for training is enabled. The user must show initiative, find the right settings and opt out (e. opt-out) of this use.

Paid Individual Subscriptions (Plus/Pro):

This often creates serious misunderstandings and a false sense of security. When a user pays for a subscription like ChatGPT Plus or Gemini Advanced, they are buying performance and access to newer and more powerful models. He is not buying increased privacy. In the vast majority of cases, the same privacy terms apply to paid individual subscriptions as they do to free versions. The user still needs to opt out of data usage for training manually.

Enterprise Solutions (Enterprise/Business/API):

At this stage, the business relationship changes dramatically. Here, the customer is a company or organization that pays for services where data security is a fundamental prerequisite. The customer's data is no longer a currency but a valuable asset that should be protected. All major service providers offer legally binding contracts that ensure that a customer's data is never used to train their public models. This is not a setting that needs to be opted out; This is the core of the service.

Contractual commitment to no training:

A clear and legally binding guarantee that a customer's data (both input and output) will not be used to train their generic models.

Data Isolation and Security Certifications:

The services are designed to isolate customer data and meet stringent global security standards, such as SOC 2 and ISO/IEC 27001.

Regulatory compliance:

The service providers commit to comply with the requirements of major data protection regulations, including the GDPR in Europe.

Powerful management tools:

Enterprise customers get access to dashboards that allow them to control employee access and monitor usage.

Verification of technical foundation

The university's borg.unak.is chatbots are based on OpenAI's APIs, specifically the services covered by OpenAI's General Business and Corporate Terms and Conditions (formerly known as Teams). This means that all interactions with these chatbots are subject to OpenAI's strictest data protection terms.

Analysis of API protection measures

Using OpenAI's business API, the university ensures the following basic protections:

• No use for training: OpenAI guarantees by legally binding means that data sent through APIs is not used to train or improve their public models.
• Data ownership: The University of Akureyri retains full ownership of the data submitted to the system and the results obtained from it.
• Strict security requirements: All data transmission is encrypted and the service is certified according to international security standards such as SOC 2 Type 2 and ISO/IEC 27001.
• Data storage and deletion: For API services, OpenAI stores data for a maximum of 30 days for the sole purpose of monitoring abuse, after which it is deleted.

Bottom Line

The University's internal chatbots on borg.unak.is are formally classified as a secure enterprise solution. They meet all the basic requirements for data protection and security. However, caution should always be exercised and data that is subject to special statutory restrictions (e.g. sensitive personal data) should be placed in any cloud system unless absolutely necessary and following a special risk assessment.

Dual functionality, unified security:

Scite.ai offers two main features:

1. Citation Analysis (Smart Citations): This is the core functionality of Scite.ai. The system analyzes how published scientific journals cite other published publications. Use of this feature includes searching and analyzing publicly available, published data. The risks associated with this use are therefore negligible.
2. Scite Assistant: This is the generative AI component of the service where users can submit questions or upload their own text for analysis. Here, the user submits his/her own, possibly unpublished, and sensitive data for processing.

Technical implementation analysis and data security

Scite Assistant is not a self-contained language model. The service is an intermediary that connects securely to enterprise solutions (Enterprise API) from authorized service providers such as OpenAI and Anthropic. This technical implementation is key. This means that all data that a user enters into Scite Assistant is subject to the strictest data protection terms of these backend services. The main point is that the data is not used to train OpenAI 's or Anthropic's public models, in full compliance with their terms for APIs and enterprise solutions. The data protection is therefore comparable to that guaranteed in the University's internal chatbots.

Conclusion and recommendations for the use of Scite.ai

Based on this analysis, both parts are Scite.ai considered safe for use within the academic community.

• Smart Citations: The use of this section Scite.ai is safe and strongly encouraged.
• Scite Assistant: The use of this section is safe for working with sensitive research data. As the service uses secure corporate connections, it is suitable for working with, analysing and compiling unpublished research, grant applications and other confidential data.

Since the University of Akureyri provides formal access to Scite.ai, the tool as a whole is classified as a secure and accepted solution for the academic community.

Activity analysis

The tool uses collections of large-scale language models to analyze lesson plans and automatically create draft e-reading lists. Teachers upload a document (e.g., Word, PDF) and the AI identifies references to course materials, extracts information, and creates an organized reading list. The system then tries to link each source to the library's collections.

Data processing and verification of security

Data security is key. Ex Libris (which is part of Clarivate) has made a clear statement about the handling of data in this process. In the frequently asked questions (FAQs) about the tool, the following is stated:

"Are you using my data to train your AI?" "No. Your data is not used to directly or indirectly train LLMs. This feature utilizes pre-trained large language models (LLM) to process your content and create a list of resources. Your input is not stored by the large language model or used for any other purpose than to build the list."

This means that HA's data is not used to train AI models. Lesson plans uploaded are sent for processing only to create the reading list and are not saved by the language models. This policy is fully compliant with the University's security requirements.

Risk factors and limitations

Despite secure data processing, there are certain risk factors that concern functionality:

• Limited accuracy: Ex Libris verifies that the tool is not 100% accurate. All lists require manual review.
• Icelandic content: The AI has primarily been tested in English. Activity with Icelandic lesson plans is the single biggest risk factor and requires thorough testing.
• Unexplained errors: The vendor acknowledges that some plans don't work without a known explanation.

Conclusion and recommendations

The Leganto AI Syllabus Assistant is a secure tool from a data protection standpoint and falls under Category 1. However, its implementation requires caution. It is recommended that:

1. Define a clear procedure for all auto-generated lists to be manually reviewed.
2. Perform a thorough testing phase with Icelandic material before the tool goes into general use.
3. Prepare educational material that explains the advantages and limitations. 

For students:

• Use the university's tools: For all learning activities, it is best to use the tools provided by the university (Category 1), such as Copilot and Scite.ai. They're safe for you and your data.
• Protect your personal information: Remember that the use of Category 3 and 4 tools is your responsibility. Be careful never to put personally identifiable or sensitive information on such services.
• Protect your study work: Consider unfinished assignments and essays as confidential. Use of Category 3 or 4 tools for such data is your sole responsibility and is strongly discouraged.

For researchers and teachers:

1. Protect intellectual property: All unpublished research papers, manuscripts, and grant applications are valuable intellectual property. The use of tools in Categories 3 and 4 to work with such data is the sole responsibility of the user and in violation of the University's policy on the protection of intellectual property.
2. Choose the right category: Use Category 1 tools (e.g. borg.unak.is, Copilot, Scite.ai) for daily work. If more powerful tools are needed, ensure that you are using a Tier 2 solution (e.g., ChatGPT Team with an enterprise agreement).
3. Mentor students: Teachers are responsible for educating their students about these principles and directing them to safe tools (Category 1).

For administrative staff:

1. Only approved systems for sensitive data: All processing of student or staff personal data, financial information or other administrative data shall only take place in systems in Category 1.
2. Responsibility for the handling of administrative data: The use of tools in Categories 3 and 4 to work with administrative data is the sole responsibility of the user. Such use constitutes a serious deviation from the University's privacy and security policy and may constitute a violation of it.