AI Data Policy

This page explains how content you produce on ResearchArk interacts with the AI features in the platform — what we do with it, what we don't, and where it is processed. It complements our Privacy Policy and GDPR Compliance pages.

Plain-language summary

• We do not use your content to train AI models. Not ours, not anyone else's.
• Your drafts and uploads stay in your account and are accessible only to you and the people you explicitly share them with.
• AI features process your content briefly to answer your request — for example, drafting a section, summarising an opportunity, or matching a profile. The processing is done either on our infrastructure or via contracted model providers under data-processing agreements.
• Generative AI features are opt-in by use. You can use ResearchArk without engaging with any of them. ArkAssist, ArkAssist Chat, and AI-assisted matching are all optional — if you don't open or invoke them, no request is sent to a third-party model. Search runs on ResearchArk's own local embedding model regardless (see below) — it isn't a third-party AI feature and isn't something you opt into — so it behaves the same whether or not you use the generative features.

What "we don't train on your content" means

When you use ArkAssist or any other AI feature, the request is sent to a language model. The model returns an answer. Once the answer is returned, the request is not added to any training dataset. We do not aggregate user prompts, drafts, or uploads into a corpus that is then used to fine-tune any model — neither models we operate ourselves nor models operated by third-party providers.

We hold contractual commitments from each model provider that user content sent through their API is not used by them for training either. Those commitments are part of the data-processing agreements in place with each provider and are reviewed periodically.

Model training and your ideas

• Your prompts and the ideas in them are not used to train AI models. The requests you send to ArkAssist, ArkAssist Chat, the ResearchArk Agent, and AI-assisted matching are used only to produce your answer. They are not added to any training dataset — neither models we operate nor models operated by our contracted providers — and each provider is bound by a data-processing agreement that prohibits training on your content.
• The ideas you develop here remain yours. You keep full intellectual-property ownership of everything you submit or generate on the platform. By using the Service you grant us only a limited licence to process and store that content so we can operate the platform for you; we do not use it for marketing, promotion, or redistribution, and we do not share it outside the ResearchArk ecosystem without your explicit consent.
• The ResearchArk Agent's inference hop is configured for zero retention. When the ResearchArk Agent generates text, it does so through an inference endpoint that holds prompts and outputs only in volatile memory for the duration of the request and does not store them afterward. Other providers, used by ArkAssist and ArkAssist Chat, may briefly retain request data for abuse monitoring under their own data-processing agreements — see the retention window for each provider below. ResearchArk itself separately retains what you explicitly save (see below).

Retention. Operational metadata for each AI request (provider, model, status, latency, request identifier — never the prompt body or the model output) is retained only as long as necessary for service operation, debugging, and abuse prevention. The drafts, conversations, and Studio content you save are retained by ResearchArk under its Data Retention Schedule and can be deleted at any time from the platform or on request to privacy@researchark.eu.

Where your content is processed

Active model providers

ArkAssist may route a request to one of the following providers depending on the model selected and provider availability:

• Anthropic (Claude family)
• OpenAI (GPT family)
• Google (Gemini family)
• Fireworks AI (curated open-weight models: DeepSeek, GLM, Kimi, MiniMax, and Qwen)

All four providers are accessed exclusively through paid, enterprise-grade API agreements — never a free or consumer-facing product. This distinction matters most for Google Gemini, whose free tier permits training on submitted content under different terms that never apply here. Each provider is bound by a data-processing agreement that prohibits training on customer content. We may add or remove providers; this page will reflect the current set.

What is logged, and for how long

We log operational metadata for each AI request — provider, model, response status, latency, request identifier — so we can debug failures and detect abuse. Logged metadata does not include the prompt body or the model output. Operational logs are retained only as long as necessary for service operation and abuse detection.

We also retain the drafts and outputs you explicitly save in your account. You can delete them at any time from the relevant section of the platform.

Your controls

• Avoid generative AI features entirely: ArkAssist, ArkAssist Chat, and AI-assisted matching are all opt-in by use. If you don't open or interact with those features, no request is sent to a third-party model. The rest of the platform — profile editing, organizations, programmes, funders, ArkSphere — works the same way regardless. Search always runs on ResearchArk's own local embedding model (see "Where your content is processed" above), independent of these opt-in features.
• Delete saved drafts: ArkAssist → Drafts → select the draft → Delete.
• Export your content: Account → Settings → Data Export → request a copy of your account data (including drafts and saved outputs). We deliver the export within 30 days, in line with GDPR Article 20.
• Right of erasure: as described in the Privacy Policy and GDPR Compliance pages.

Profiles and discoverability

If your profile visibility is set to "public" or "connections only", your profile text is embedded so other users can discover you through search and matching. Embeddings are computed on our infrastructure (EmbeddingGemma-300M, run locally) and stored on ResearchArk infrastructure. Profiles with visibility set to "private" are never embedded and are excluded from search.

When your visibility is "connections only", your profile only appears in another researcher's search results if you are already connected to that researcher.

Changes to this policy

We will update this page when our practices change — for example if we add a new model provider or change a data-processing region. Material changes are announced via the platform notification system. The current version and last-updated date are shown at the bottom of this page.

Contact

Questions about this policy can be sent to privacy@researchark.eu.