Policy for using genAI tools at Centrum Cyfrowe

#The social aspects of technology

Since 2022, generative AI changed and keeps changing how we write, create, research, learn and process information.

At the Centrum Cyfrowe Foundation, we use various genAI tools depending on the tasks we perform. We use them for interview transcription, text analysis, support in generating a wider range of variants, translation, and at various stages of research.

We use these tools when creating emails, social media posts, and writing applications. It is worth clearly distinguishing here what data can be processed in corporate office tools (for which the Foundation has purchased licenses), vs. privately used or open tools – the latter are acceptable, but only in situations that do not require the use of confidential, sensitive data or data covered by organizational non-disclosure agreement. We are also aware of wide-spread intellectual property infringement – materials used without the author’s/creator’s permission for the purposes of model training, which is why we wish to choose our tools consciously and responsibly.

The overriding principle that guides us is that we, people using these tools, take full responsibility for the results of the published output. We also do not delegate the entirety of tasks to generative artificial intelligence, only selected parts – particularly related to organizing the content we have already prepared, reviewing it for its logical flow and potential gaps in our reasoning. Below are the detailed rules we apply when working with genAI tools at the Centrum Cyfrowe Foundation.

What principles guide us?

Transparency – we disclose when and why (for what purpose) we used genAI, unless these are routine activities (e.g. translating a text which is later quality-checked by a person – like in the case of this very sentence).
Limited trust – we do not upload data that could be considered sensitive, personal, classified etc., into genAI tools.
Doubting by default – we assume that genAI may always provide erroneous information¹, e.g., referring to non-existent or misattributed publications and/or wrongly linking conclusions from different sources. We remain vigilant regarding biases that may result from the models’ training data, lack of transparency, or the general attitude and business goals of the producers of these tools.
Limited use – we do not shun genAI tools away, but we use them when it makes sense or is necessary — keeping in mind the carbon footprint, as well as our intellectual work’s hygiene and discipline.
Human first, machine second – genAI tools serve as a supplement to people’s work. This is why we first come up with an idea for research, a text, or research conclusions – and only then use genAI tools to search for gaps in reasoning or discover previously unseen connections between data.
We avoid both the ”humanisation” (anthropomorphization) of tools based on artificial intelligence and their ”demonisation”. We believe that digital tools can improve daily work according to the users’ individual needs or habits, and above all, support (and not displace) human knowledge, creativity, sensitivity, intuition, and reasoning. The key to a sustainable use of genAI is awareness of the mechanisms, possibilities, and risks associated with it.

Most common examples of our use of genAI:

Transcriptions of interviews in qualitative research.
Text translations – both for working purposes (e.g., translating a scientific article in a foreign language) and as support in publications (translations published on the Centrum Cyfrowe website always undergo substantive verification and editing process by specialists).
Building a base of concept alternatives (ideation).
Critique of written text.
Proposals for titles or project names.
Deepening and broadening argumentation.
Support in preparing for speeches.
Searching through large data sets.

Use of large language models in the writing process:

Support in creating communication texts and their variants tailored to different channels and audiences (e.g., updates on the website, social media content, press releases).
Drafts of emails, especially summarizing meetings based on rough, handwritten notes.
Summaries of meetings, debates, conversations.
Developing points in a discussion prepared by a person/people, fragments of articles, and/or prepared applications.
Support in linguistic and stylistic editing and proofreading.

Use of genAI in the research and design process:

Transcribing interviews, discovery workshops, etc.
Desk research analysis, especially for drawing conclusions from a large number of documents – always verifying the context within the source, on which the model based its proposal.
Qualitative data analysis — as a supplement to manual coding (triangulation).
Creating project documentation based on various types of information (e.g., meeting notes, emails, GitHub threads), with appropriate protection of sensitive or classified data.

Use of genAI and other AI-based tools in other cases:

Use of automation in improving the creation of video and audio materials (e.g., support in editing, cutting out pauses, etc.).
Support in retouching or editing photos (e.g., expanding the background to improve composition) – without interfering with the appearance of the people depicted.
Searching for information, e.g., about the author, source, or authenticity of visual materials (photos, graphics) through AI-based reverse image search tools.
Generating alternative texts (alt texts) for visual materials (enabling the ”reading” of a digital image by screen readers for the visually impaired, as well as improving content recognition by platform algorithms or search engines).
Furthermore, in our daily work, we use Google Workspace tools, which have functionalities powered by artificial intelligence models.

When and for what purposes do we not use genAI tools?

To create entire texts and/or presentations, speeches, communications.
To generate visual materials from scratch, e.g., graphics or deepfakes to illustrate our statements or any other publications. We use the wealth of resources from the public domain and Creative Commons. Exceptions may be materials used (or created) during training workshops – where we simultaneously discuss issues surrounding AI-generated visual materials.
To come up with ideas for a text or content: first, we write about our idea (in a more or less structured way), and only then do we use genAI to point out any weak points, significant gaps, potential directions for expanding the topic, etc.

¹ We deliberately avoid using the term “hallucinations”, which ascribes agency and – at times – human characteristics to genAI tools. To hallucinate implies the possession of cognitive functions, which GenAI does not yet possess; see: https://news.northeastern.edu/2023/11/10/ai-chatbot-hallucinations/.

This text was originally written and published in Polish, without the use of AI. After an initial AI translation via Gemini, it was proofread and edited by our team.