Generative AI can lift the administrative weight of literature reviews and manuscript drafting, but only when it stays grounded in evidence and firmly under human control.
The problem we know too well
Every clinician who also publishes lives with a quiet collision. On one side sits the expectation to produce scholarly work; on the other, a clinical workload that rarely relents. The old imperative to “publish or perish” has not softened. If anything, it now competes with more patients, more documentation, and less time.
The traditional machinery of research makes this harder. A systematic review or an original manuscript can take months of painstaking data extraction and synthesis. Much of that work is logistical rather than intellectual, and it is exactly the labour that leaves researchers depleted.
There is also an equity problem that is easy to miss. English is the first language of less than 10% of the world’s population, yet it dominates nearly 75% of academic publishing. Researchers who think and treat patients in another language must spend considerably more time and resources to reach high-impact journals. That imbalance does more than slow people down. It can shape study design, reporting clarity, and even peer-review outcomes, quietly reinforcing who gets seen and who advances.
This is the burden that generative AI is now being asked to ease. In our recent review in the Journal of ISAKOS, we set out a structured way for orthopaedic researchers to use these tools well. The principles, we believe, travel across the life sciences.
Generative AI: from black box to grounded assistance
The instinctive worry about generative AI in research is that it invents things. That worry is justified. Large language models, or LLMs, are the systems behind tools such as ChatGPT and Claude. They are probabilistic engines trained to predict plausible text. They are genuinely good at outlining, summarising, and polishing language, but they have no built-in sense of whether a source is real or an argument is sound. Left ungrounded, they hallucinate, fabricating facts, data, and even citations with complete confidence.
What makes AI usable for serious research is the move away from this opaque “black box” generation toward something verifiable. The key technique is retrieval-augmented generation, or RAG. In plain terms, RAG ties the model to a specific set of documents you supply, so its answers come from those sources rather than from the open sea of its training data. It does not eliminate hallucination, and it never removes the need to check, but it turns a confident improviser into a constrained, traceable assistant.
How it works in practice
Our framework breaks the research workflow into three phases.
The first is discovery. Semantic search tools read the meaning of a query rather than merely matching keywords. They surface related papers, map citation networks, and let a researcher follow the lineage of an idea, so that seminal and emerging studies are both caught early. Established databases such as PubMed remain the backbone; these tools are the high-yield complement.
The second phase is extraction. Once a curated library of sources exists, RAG-based systems pull structured data from those documents and anchor every answer to its place in the original file. This stage has always been the bottleneck, and it is where grounding matters most.
The third phase is drafting. LLMs help structure and refine prose, ideally section by section. The discipline we stress is simple: the researcher writes the substance first and uses the tool to critique and sharpen, not to author. The machine is a drafter and editor, never the intellectual source.
The opportunities
Used this way, the gains are real. Literature synthesis and data extraction become faster. Because every claim is tied to a source and every step is logged, the work becomes more reproducible and easier to audit. Semantic discovery surfaces connections a keyword search would miss. And for non-native English speakers, careful language refinement helps level the field, widening access to high-level publishing rather than narrowing it.
The risks, and the guardrails
None of this comes for free. Hallucination remains the central danger. Early AI-generated reviews, though fast, have shown reference inaccuracy reaching up to 70% in workflows that lacked grounding. The fluency of the text makes errors harder to spot, not easier.
Bias is the second risk. Models can reproduce and amplify whatever skew sits in their training data, subtly shaping how prevalence, risk, or outcomes are framed. A third and more insidious risk is homogenisation. When everyone leans on the same models and prompts, scientific writing drifts toward a uniform, postgraduate-level sameness that erodes individual voice and critical reasoning. In orthopaedics, that convergence has already been documented.
Then there is privacy. Patient-level data and unpublished manuscripts must never be entered into ungoverned platforms, where content may be retained or absorbed into future training. Regulations such as HIPAA in the United States and GDPR in Europe are not optional here. Finally, accountability. AI cannot be an author. It cannot take responsibility for a study’s design, interpretation, or final approval, and that responsibility stays human, completely.
The guardrails follow directly. Verify everything against the original sources. Keep an audit trail that records the tool used, the prompts, and the human check at each step. Use only secure, compliant systems for anything sensitive. Disclose AI involvement openly, as the major publishers now require. These are the practices that stop efficiency from curdling into what the field has begun to call “scientific pollution”. Its starkest form is the paper mill: an estimated 3% of medical publications in 2022 showed features consistent with such mass-produced, fraudulent work.
What comes next
The near future points toward agentic AI: systems that chain several steps together, retrieving, reasoning, and acting across connected tools with less prompting. The promise is more automation; the condition is unchanged. These systems still need human oversight, robust governance, and clear lines of responsibility. The field also needs shared standards for evaluation and disclosure, along with curated registries of compliant tools, the kind of infrastructure professional societies are well placed to maintain.
The constant
For all the capability on offer, our conclusion is deliberately modest. AI works best as a cognitive exoskeleton. It can carry the administrative and logistical weight so that clinicians and researchers can spend their attention where it belongs, on interpretation, rigour, and relevance. It augments expertise; it does not replace it. The technology will keep changing. The need for human judgment will not.
Author Bios
Ricardo Bastos
Prof. Dr. Ricardo Bastos is an orthopaedic surgeon with over 20 years of experience, specialising in knee surgery and sports medicine. Internationally trained and academically accomplished, he blends clinical excellence with research. He also completed an Executive MBA at Porto Business School and Kozminski University, strengthening his focus on leadership and innovation in healthcare.
Américo Dias
Américo Dias is an analog IC design engineer with over 20 years of experience in electronics, including 14 years in the semiconductor industry. He completed a dual Executive MBA at Porto Business School and Kozminski University in 2026, and his current interests include AI assisted technical work, engineering decision making, and the interface between technology and business.
Sthefan Berwanger
Prof. Sthefan Berwanger is a Researcher at INESC TEC with over 25 years of experience in business consulting and more than 20 years in executive and higher education. He is an award winning lecturer at international business schools and a guest lecturer at Porto Business School.
