I’d be curious to know which tools you’re using and what your workflow is.
Generic commercial chat portals from frontier labs are not (primarily) optimized for epistemology. They’re optimized for likeability, perceived usefulness, engagement, etc. They are not inherently truth-seeking. They do not stake out and commit to factual positions. You can nudge them in a particular direction and push them pretty reliably into whatever space confirms your biases.
Literature reviews are useful, if you demand links to sources and double-check them. Grammar/spelling checks are great. But using a public chatbot and expecting a rigorous lab partner is not a good idea.
The big labs have all recently announced/released science tools:
OpenAI — GPT-Rosalind (announced April 16–17, 2026, named after Rosalind Franklin)
Anthropic — Claude for Life Sciences (launched last October as a suite of connectors and skills)
Google DeepMind — AI Co-Clinician (announced April 30, 2026)
None of these are open. GPT-Rosalind is gated trusted-access; Co-Clinician is a research project, not deployable; Claude for Life Sciences is enterprise-targeted via API credits.
LLMs don’t necessarily lead to what you’re calling “epistemic immunodepression”. They can and will if used naively. The same way AI coding tools can easily lead to mountains of technical-debt slop without some basic rigor. The tools need to be optimized for that particular role and they need to be used responsibly. Right now, neither is happening at scale.
My workflow is modest. For daily work, I use the public versions of Claude (cowork), GPT, and Gemini. No GPT-Rosalind, no Claude for Life Sciences, no Co-Clinician. I do not have institutional access to any gated research tool. I work as an independent clinical researcher with no funding, so what I can use is what is on the open web at a paying-user tier. For research, when I need to evaluate the output of LLMs, I use some local models through Ollama and Google API key through Google AI studio.
What I try to do is detect the model disagreement rather than agreement. I prompt for the strongest objections to my own claim before I ask for support. I ask three models the same question and compare answers. I make the model name the assumption it is hiding. None of this is sophisticated. It is the minimum hygiene that a non-native English clinician with a full-time hospital job can manage.
And yes, I agree with your core point. A generic chatbot used naively will produce confirmation, not correction. That is exactly the failure mode I am trying to diagnose. My paper is not arguing that LLMs are bad. It is arguing that without structured frameworks for use, the default drift is toward what I call “epistemic immunodepression”. The tools you listed (Rosalind, Co-Clinician, Life Sciences) are promising because they impose structure. The problem is the gap while we wait for them.
I will share more about the workflow and the empirical study behind this paper in the next post of the sequence. I am building this in public, mostly because I have no other option, and partly because I think the failure modes are easier to see from outside the institutions.
I’d be curious to know which tools you’re using and what your workflow is.
Generic commercial chat portals from frontier labs are not (primarily) optimized for epistemology. They’re optimized for likeability, perceived usefulness, engagement, etc. They are not inherently truth-seeking. They do not stake out and commit to factual positions. You can nudge them in a particular direction and push them pretty reliably into whatever space confirms your biases.
Literature reviews are useful, if you demand links to sources and double-check them. Grammar/spelling checks are great. But using a public chatbot and expecting a rigorous lab partner is not a good idea.
The big labs have all recently announced/released science tools:
OpenAI — GPT-Rosalind (announced April 16–17, 2026, named after Rosalind Franklin)
Anthropic — Claude for Life Sciences (launched last October as a suite of connectors and skills)
Google DeepMind — AI Co-Clinician (announced April 30, 2026)
None of these are open. GPT-Rosalind is gated trusted-access; Co-Clinician is a research project, not deployable; Claude for Life Sciences is enterprise-targeted via API credits.
LLMs don’t necessarily lead to what you’re calling “epistemic immunodepression”. They can and will if used naively. The same way AI coding tools can easily lead to mountains of technical-debt slop without some basic rigor. The tools need to be optimized for that particular role and they need to be used responsibly. Right now, neither is happening at scale.
Thanks for the question.
My workflow is modest. For daily work, I use the public versions of Claude (cowork), GPT, and Gemini. No GPT-Rosalind, no Claude for Life Sciences, no Co-Clinician. I do not have institutional access to any gated research tool. I work as an independent clinical researcher with no funding, so what I can use is what is on the open web at a paying-user tier. For research, when I need to evaluate the output of LLMs, I use some local models through Ollama and Google API key through Google AI studio.
What I try to do is detect the model disagreement rather than agreement. I prompt for the strongest objections to my own claim before I ask for support. I ask three models the same question and compare answers. I make the model name the assumption it is hiding. None of this is sophisticated. It is the minimum hygiene that a non-native English clinician with a full-time hospital job can manage.
And yes, I agree with your core point. A generic chatbot used naively will produce confirmation, not correction. That is exactly the failure mode I am trying to diagnose. My paper is not arguing that LLMs are bad. It is arguing that without structured frameworks for use, the default drift is toward what I call “epistemic immunodepression”. The tools you listed (Rosalind, Co-Clinician, Life Sciences) are promising because they impose structure. The problem is the gap while we wait for them.
I will share more about the workflow and the empirical study behind this paper in the next post of the sequence. I am building this in public, mostly because I have no other option, and partly because I think the failure modes are easier to see from outside the institutions.