General-purpose assistants like ChatGPT are genuinely useful for drafting, brainstorming, and explaining concepts. But research teams keep asking a sharper question: is it safe to paste confidential or unpublished research into one? The honest answer is that it depends entirely on what the tool does with your text, what account and settings you are using, and how sensitive the work is. This post walks through the actual risks, separates myth from reality, and lays out safer patterns for teams that handle pre-publication results, proprietary methods, or regulated data.
What actually happens to the data you send
When you submit a prompt to a hosted assistant, your text leaves your machine and is processed on the provider's servers. From there, three things matter: whether the content is retained, whether it can be used to train future models, and who inside or outside the provider could access it. Consumer-tier products have historically defaulted to using conversations to improve models unless you opt out, while business and enterprise tiers usually promise no training on your data by contract. Those distinctions are easy to miss, and the default a researcher lands on is often the least protective one.
Even with training turned off, your content typically still transits and is stored on third-party infrastructure, may be retained for a period for abuse monitoring, and is subject to that provider's security posture and legal jurisdiction. None of that is automatically disqualifying. It simply means the safety question is not yes or no - it is whether the data handling matches the sensitivity of what you are pasting.
The unpublished-IP and peer-review risk
For research teams, the highest-stakes concern is rarely a dramatic breach. It is the slow erosion of control over unpublished intellectual property. Novel methods, unfiled results, grant ideas, and draft manuscripts derive much of their value from being first and being yours. Pasting them into a general assistant means a copy now lives outside your environment, governed by terms you did not write.
Peer review raises the bar further. Many journals and funders treat manuscripts under review as confidential, and uploading a reviewer's manuscript - or your own under-review draft - into a third-party tool can violate confidentiality obligations. The same applies to data covered by IRB protocols, NDAs with industry partners, or export-control rules. The tool being technically secure does not resolve a contractual or ethical commitment that the content stay in approved hands.
Hallucinated citations are a separate hazard
Confidentiality is only half the problem. General assistants also fabricate citations - they generate references that look real, with plausible authors, journals, and years, but that do not exist or do not say what the answer claims. This happens because the model is predicting fluent text, not retrieving and quoting verified sources. For research work, a fabricated citation that slips into a literature review or a grant proposal is a serious credibility risk.
We cover the mechanism and the fix in more depth in our companion post on why AI fabricates citations and how grounded retrieval addresses it. The short version: an assistant that answers from its training weights cannot show you the passage it is standing on, so you cannot verify it.
Safer patterns for sensitive research
If the work is non-sensitive - public papers, general background, or material you would happily post publicly - a mainstream assistant on a no-training tier is often a reasonable choice. The calculus changes for confidential or unpublished work. Here the goal is to keep data inside a boundary you control and to make answers verifiable. A few patterns help:
- Use a private knowledge base scoped to your team, so answers come from your own uploaded documents rather than a shared public model, and each tenant's data stays isolated.
- Bring your own provider key, so any managed AI calls run under your own provider account and contractual terms rather than a shared consumer account.
- Run fully local or self-hosted for the most sensitive corpora, so documents and queries never leave your environment at all.
- Prefer a system that cites every answer to a specific passage and declines when the evidence is missing, instead of one that always produces a confident guess.
How Researcher Center approaches this
Researcher Center was built around exactly this gap. It is a private AI knowledge base for research teams: answers are grounded in your own documents and cited to the passage they came from, and the system declines when your sources do not support an answer rather than guessing. You choose how the AI runs - platform-managed, your own provider key, or fully local and self-hosted - and data stays isolated per tenant and workspace, with hard spend caps so costs cannot run away. You can read more about the deployment choices in our post comparing self-hosted, managed, and bring-your-own-key AI.
If your team is weighing whether a general assistant is safe for confidential or unpublished research, the most useful next step is to try a controlled setup on a small, non-confidential corpus first. Early Access is an invite-only free trial - you can request access on our pilot page and start with documents you are comfortable testing on.