As a neuroscientist and published author with over 10 years of science communication experience, I love the potential of AI to make literature reviews and science writing more accessible and easier to understand.

While AI tools for research can speed up scientific searches, they don’t work like a scientist does. Most research AI tools simply extract and summarise information that matches your prompt. It’s essential to understand how these tools work (and their limitations) before using them for anything critical. Often, you’ll need to conduct manual research alongside using an AI tool to avoid ending up with flawed, incomplete, or biased data.

In this guide, I’ll walk you through:

• how AI literature reviews compare with real researchers
• which AI tools for academic research are worth exploring, their benefits and limitations
• what AI can (and can’t) tell you about a scientific study
• how to use AI for science and report writing

Using AI for Research: What You Need to Know

While AI tools for research aren’t capable of replacing a rigorous, manual literature review yet, they can help speed up the process. But if you’re not careful, they can give you a flawed or misleading perspective. That’s why you need to know how to use them effectively. 

In a previous blog, I explored how Gen AI is impacting scientists and businesses, and what it means for health tech. I recommend reading that post for a deeper understanding of how scientists are currently using Gen AI research tools in their work (and what they’re concerned about), and what the risks are when using them in a business.

AI Literature Review Tools vs. Real Researchers

If you’ve ever done a real literature review, you’ll know it’s a layered, iterative process.

In academia, it takes a team of experts who independently screen hundreds (or thousands) of studies, assess bias, and document the rationale for inclusion or exclusion. Only then will they begin analysing the combined data. This can take between 6 months to 2 years.

In industry, the review process is faster and more targeted. For concept validation, you might combine PubMed with Google Scholar, look at trial registries, and scan grey literature like industry reports. Relevance is judged by the methods, journal quality, citation patterns, sample generalisability, and whether the findings apply to the product context.

AI tools do not conduct research the same way a scientist would. They typically just search for studies and extract data that semantically matches your query. That may be helpful as a starting point, but it also misses nuance, novelty, and interdisciplinary relevance. 

We also need to be cautious of how much trust we place in these tools, and consider whether outsourcing tasks like scientific research will make us less capable of them in the long-term. Indeed, a study by Microsoft found that having higher confidence in AI led to less critical thinking, and people engaged less in their work over time while using AI. The authors said:

While GenAI can improve worker efficiency, it can inhibit critical engagement with work and can potentially lead to long-term over-reliance on the tool and diminished skill for independent problem-solving.

For more context regarding the deeper issues raised by AI in science, read how AI can create a monoculture of knowing.

Best AI Tools for Literature Reviews and Research Writing (2025)

I personally tested several AI tools for academic research while working on client projects. I compared each tool against the manual scientific search I carried out myself. This post isn’t sponsored, and I’m not an advocate of any particular one. This is simply a non-exhaustive review of tools and my personal opinion of what worked, what didn’t, and what you need to watch out for.

Deep Research (OpenAI)

ChatGPT's Deep Research model was the most intuitive tool I tried due to its excellent user experience. You can specify whether you want peer-reviewed research or grey literature, and it links citations directly to the source text (within a paper or article) rather than just the source itself.

That said, the relevance of studies varied. Sometimes the tool pulled outdated or off-topic papers. It also often cited the wrong study — quoting findings from other papers mentioned in the introduction or discussion section, rather than the actual results. It can only pull data from open-access papers, which skews the output and limits depth.

• Best for: user experience and flexible query outputs.
• Limitations: confirmation bias is a real issue here. Results can shift depending on who you say the report is for. Not recommended unless you’re already a subject matter expert.

Elicit

I wanted to love Elicit. It explains its logic well, helps you reframe your question, and walks you through its paper selection process. It works by screening articles and extracting data from a subset of these papers (like a scientist would), but the output was very random and outdated. It missed a lot of highly relevant papers that I found in my independent search, despite most of them being open-access. It also has strict limits on paper extraction unless you upgrade, and for 50 credits per month (recommended for literature reviews), it only included three searches, two of which returned completely useless results. 

• Best for: simple questions in well-researched fields (if you extract from >500 papers using the Pro version).
• Limitations: recent research and relevant papers were often missed. Not ideal for anything novel, contested, or interdisciplinary. Also not beginner-friendly, setting the right parameters takes time, and bad inputs = bad results.

Stanford’s Co-Storm

Disappointing. It pulled data from reputable websites but very little actual academic research. In my experience, it returned helpful results only about 20% of the time. Might work better for well-established areas of science. Not recommended for literature reviews.

• Best for: media scanning and finding press releases from reputable websites.
• Limitations: does not seem to use data from peer-reviewed sources.

Notebook LM

Not necessarily a lit review tool, but it can be useful for summarising papers you collect as part of the review process. It has a low risk of hallucination because it limits its answers to the sources you provide. That means its great for asking questions across several papers, especially if you’re trying to summarise your own research base.

• Best for: summarising papers you already have and turns complex results into digestible podcasts that you can listen to.
• Limitations: it doesn’t always answer clearly, and sometimes it doesn’t interpret the findings correctly. Also, it’s hard to track which comment relates to which study, so label your sources clearly.

Undermind & Sci Reports

Neither worked as intended. I tested the pro versions and couldn’t get meaningful results. Avoid.

Dexa

I'm just including Dexa as an honourable mention, but it's not for literature reviews. It's great for finding podcast content based on your interests. The search function works much better than Apple or Spotify. Remember to treat podcast science (even Huberman ones) with caution. Opinions ≠ evidence.

Benchmarking Report of AI Tools

For a more comprehensive report of which AI tools are best for literature reviews, I would highly recommend checking out Nuance AI Lab's report, which benchmarked 20+ tools and evaluated eight of those on things like prompting quality, citation quality, writing style, speed, and costs.

Interestingly, they found Elicit performed best when comparing how these tools stack up relative to one another (rather than comparing the results to a human scientist conducting a review). It's likely that performance can vary substantially depending on the scientific field and research question, so keep this in mind.

Final thoughts

If you already know how to conduct a literature review, tools like Deep Research or Notebook LM can complement existing processes,  especially for scanning and organising information. But they don’t replace the fundamentals: a clear search strategy, critical appraisal, bias checking, or the ability to interpret relevance across disciplines.

They also don’t give you access to the most recent or relevant studies, which is often where the most valuable insights are. For most use cases, that’s a critical issue. Whether you're developing writing a scientific report, validating ideas, or tracking emerging research, you need more than an outdated summary.

You need to understand what matters, what’s missing, and why.

Currently, AI tools aren’t built to do that.

The Hidden Flaws in AI Research Tools

So why do these issues occur when trying to use AI tools for academic research and writing? It’s because Gen AI tools were never designed for scientific reasoning in the first place. The way these models are programmed actually contradicts the principles and moral code that scientists follow.

Gen AI doesn’t understand science, it mimics it

LLMs generate responses by predicting the most likely next word based on probability, not comprehension. They don’t “know” whether a study was well-designed, peer-reviewed, or contradicted by more recent work. Ask it to summarise a paper, and it might sound confident, but it won’t tell you if the findings were retracted, misrepresented in the media, or have since been disproven. This is especially risky in health, where subtle details in a study’s design or sample population can completely change how applicable the findings are.

You only get a slice of the literature

Most LLMs are trained on open-access papers, which make up a relatively small (~30% of research), omitting critical paywalled studies. High-quality, peer-reviewed studies in major journals are often locked behind paywalls, meaning AI simply can’t “see” them. Even if you ask it to provide the latest research, you might be getting an outdated or skewed picture.

You could be basing scientific claims or investment slide decks on a skewed or outdated subset of the research without even realising it.

LLMs sound confident, even when they’re wrong

A 2024 Nature Machine Intelligence study found that tools like GPT-4 often express the same level of confidence regardless of whether they’re 40% or 90% confident in their answer. This creates a dangerous illusion of reliability.

Even more concerning is that people in the study were more likely to trust longer, more detailed answers, even when those answers didn’t improve accuracy. In other words, AI is trained to sound persuasive, not accurate. And that’s a problem when you’re relying on it to make claims that affect real people’s health.

Built to please

AI models tend to reinforce the assumptions in your question. If you ask, “Does X supplement reduce anxiety?” you’ll likely get an answer that starts with: “Yes, several studies suggest…” — even if the overall body of evidence is weak or contradictory. A recent study showed that as LLMs become more instructable, their reliability decreases and they are more likely to give wrong answers instead of declining to answer.

I’ve experienced this myself when summarising the results from studies. I will often get a different answer depending on how I frame the question.

No warnings for bias

Scientific studies are often shaped by who funds them, how they’re designed, and who’s included or excluded from the participant pool. Some journals are rigorous. Others will publish just about anything if you pay. A good scientist will look at these variables and adjust their interpretation accordingly.

AI can’t gauge the reputational weight of a journal or author. It doesn’t factor in whether the study you’re quoting came from The Lancet or a predatory journal with no peer review. Unless you’re checking every study, you could end up referencing something unreliable without knowing it.

It can’t weigh evidence

In science, one study doesn’t prove anything. Good researchers look for systematic reviews, meta-analyses, replication studies, and debates within the field. Gen AI doesn’t do this. It cherry-picks what’s most statistically likely to answer your prompt. It won’t tell you that one meta-analysis contradicts another, or that the study you’re quoting was based on 12 participants in a single clinic. These details shape how generalisable the results are, especially if you’re building for diverse real-world users.

While AI may give you “an answer,” that answer may not reflect the scientific consensus, or even the majority view.

AI Report Writing and Science Communication: What’s Safe?

If you already know what you want to say, Gen AI can help polish your draft, improve structure, reduce jargon, or tighten the tone. But it can’t generate real insight because it doesn’t understand what it’s saying. As the Science Editor-in-Chief, Holden Thorp, put it:

‍“ChatGPT is fun, but not an author.”

Why? Because AI has no ideas, no opinions, and no narrative agenda. It just stitches together familiar phrases from its training data. As Nature editors wrote, what you get is often “clichéd nothingness” — language that sounds good but says very little.

Use AI in writing for:

• Adjusting the structure, formatting, and improving flow
• Making technical ideas clearer (if you understand the concept)
• Editing your own ideas into something more readable

Don’t use AI for:

• Creating original scientific content
• Writing about health topics without expert review
• Generating or trusting citations
• Replacing your own thinking

If your goal is clarity, AI can help.

If your goal is credibility, you still need a human.

Need Help Navigating AI in Research?

If you're using AI for research, developing health products, or writing reports, our team of experts can help you do it responsibly (without compromising on accuracy or credibility).

At Sci-translate, we offer:

• Research & Data Analysis
• Science Communication
• Medical writing
• Product Testing

Book a call with me with me (Dr Anna McLaughlin, founder of Sci-translate) to get started.

Let’s make sure your science stays sound, even in the age of AI.