Abstract
Background: Safety and efficacy decisions for food supplements, foods for special medical purposes and novel ingredients increasingly depend on large, heterogeneous bodies of evidence: randomised trials, observational cohorts, animal toxicology, in vitro mechanistic work and regulatory opinions. Machine learning and large language models (LLMs) promise to shorten the slowest stages of evidence synthesis, but published evaluations show that their reliability varies sharply between tasks.
Objective: To describe a human-in-the-loop framework in which AI tools accelerate retrieval, de-duplication, prioritisation and draft data extraction, while every inclusion decision, risk-of-bias judgement, certainty rating and conclusion remains with a named, qualified human reviewer.
Methods: Narrative methodological review of reporting and appraisal standards (PRISMA 2020, PRISMA-S, PRISMA-P, PECO, RoB 2, ROBINS-I, AMSTAR 2, GRADE, the OHAT approach and the EFSA guidance on systematic review methodology for food and feed safety), of published evaluations of machine learning and LLMs in systematic review tasks, and of the obligations of Regulation (EU) 2024/1689 (the AI Act) relevant to AI-assisted scientific work.
Conclusions: The evidence supports AI as an accelerator of search, screening prioritisation and draft extraction, not as an autonomous reviewer. Unsupervised LLM output can contain fabricated references and unreliable screening decisions. A defensible workflow pre-registers the protocol, logs every AI-assisted step, verifies every citation against a bibliographic registry, checks every number against the source text and keeps the final scientific judgement with accountable experts.
1. Introduction
A product dossier for a food supplement or a medical nutrition product is only as strong as the evidence review behind it. Regulators, clinicians and brand owners need to know not just what a study found but how the studies were found, why some were excluded and how confident the reviewer is in the overall body of evidence. Systematic review methodology answers these questions through a transparent, reproducible sequence: a pre-specified question and protocol, a documented search, independent screening, structured data extraction, risk-of-bias appraisal and a graded statement of certainty.
For food and feed safety, the European Food Safety Authority published guidance on applying this methodology to risk assessment questions as early as 2010 [10]. In environmental health and toxicology, the National Toxicology Program's OHAT approach extended systematic review to the integration of human, animal and mechanistic evidence for hazard identification [6]. These frameworks are demanding: a single review can involve thousands of records and weeks of expert time.
Machine learning tools for search, screening and extraction have matured over the past decade [15, 16], and general-purpose LLMs are now widely available. The question for any organisation producing scientific dossiers is therefore no longer whether to use AI, but where AI output can be trusted, where it must be verified and where human judgement cannot be delegated.
2. Reporting and appraisal standards that do not change with AI
2.1 Question, protocol and search
A review starts with a structured question. For interventions this is typically PICO (population, intervention, comparator, outcome); for safety and exposure questions the PECO framework (population, exposure, comparator, outcome) is better suited, because it accommodates exposures that are not chosen by the participant and comparisons across exposure levels [9]. The protocol should be written before screening begins and reported according to PRISMA-P, a 17-item checklist [3].
The search must be reproducible. PRISMA-S adds 16 reporting items specific to literature searches, including the databases and platforms used, the full search strings, limits, date of last search and de-duplication method [2]. These items apply regardless of whether a query was drafted by a person or suggested by an AI tool: the final search string, as executed, is what must be reported.
2.2 Selection, extraction and reporting
PRISMA 2020 provides a 27-item checklist and a flow diagram that accounts for every record from identification to inclusion [1]. Conventional practice is independent double screening by two reviewers. A methodological systematic review comparing single with double screening found a median of 5% of relevant studies missed by single screening (range 0–58%), rising to a median of 13% for less experienced reviewers [11]. Agreement between reviewers is commonly quantified with Cohen's kappa [17] and interpreted on the Landis and Koch scale, on which values of 0.61–0.80 indicate substantial and 0.81–1.00 almost perfect agreement [18].
2.3 Risk of bias and certainty
Each included study is appraised with a design-appropriate tool: RoB 2 for randomised trials, structured in five bias domains [4], and ROBINS-I for non-randomised studies of interventions, structured in seven domains [5]. Existing systematic reviews used as evidence can be appraised with AMSTAR 2 [8].
The certainty of the body of evidence for each outcome is then rated with GRADE, which considers risk of bias, inconsistency, indirectness, imprecision and publication bias, and allows rating up for large effects, dose–response gradients and residual confounding that would reduce the observed effect [7]. The OHAT approach adapts this logic to environmental and toxicological evidence and translates confidence ratings into levels of evidence for a health effect, integrating human and animal streams [6].
None of these steps is a clerical task. Deciding whether a study population is close enough to the target population, whether a dose is comparable, or whether a confounder is serious enough to downgrade certainty requires domain expertise and accountability.
3. What the evidence says about AI in systematic reviews
3.1 Where machine learning helps
Marshall and Wallace reviewed machine learning methods proposed for search, screening and data extraction and concluded that several are ready for practical use, provided the review team understands their strengths and failure modes [15]. Active learning tools, which re-rank unscreened records as reviewers label them, let teams find relevant studies earlier in the screening process; open-source, transparent frameworks of this kind are available [16]. Used this way, the human still makes every inclusion decision; the machine only changes the order in which records are seen.
For draft data extraction, a proof-of-concept study found that an LLM extracted 160 predefined data elements from 10 open-access randomised trials with an overall accuracy of 96.3%; most of its six errors were missed items [12]. The authors framed this as potential for efficiency, based on a small convenience sample, not as a replacement for human extraction.
3.2 Where unsupervised AI fails
Screening performance is less reassuring. A pre-registered evaluation of a general-purpose LLM working "human-out-of-the-loop" found that apparent accuracy was inflated by chance agreement and dataset imbalance. After adjustment, screening performance ranged from none in balanced datasets to moderate in imbalanced ones, and data extraction performance was moderate. The authors concluded that substantial caution is needed [13].
Reference generation is the most serious failure mode. When LLMs were asked to reproduce the reference lists of 11 published systematic reviews, hallucination rates (references with at least two of title, first author or year wrong) were 39.6%, 28.6% and 91.4% for the three models tested, and precision was at most 13.4% [14]. A dossier that cites a non-existent paper loses its credibility with regulators and clinicians, regardless of the quality of the rest of the work.
3.3 Implications
The pattern is consistent. AI is useful when it proposes and a human disposes: candidate records, a screening order, a draft extraction table. It is unreliable when its output is accepted without verification. The design question is therefore how to make every AI contribution verifiable, logged and owned by a person.
4. A human-in-the-loop workflow
The workflow below sets out the human-in-the-loop design principles that Olympia Biosciences uses when planning evidence reviews for supplement, nutraceutical and medical nutrition development. When a review is conducted and reported as systematic, each AI-assisted stage is expected to produce the listed artefact, which a human checks and signs off before the next stage begins.
- Question and protocol
AI-assisted contribution: Suggests synonyms, outcome lists and candidate PECO/PICO elements
Human responsibility: Fixes the question, eligibility criteria and analysis plan before screening
Verification artefact: Protocol reported to PRISMA-P [3]
- Search
AI-assisted contribution: Proposes search-string variants and candidate records from semantic search
Human responsibility: Approves the final strings; runs and documents them in bibliographic databases
Verification artefact: PRISMA-S search record [2]
- De-duplication and prioritisation
AI-assisted contribution: Groups duplicates; ranks records by predicted relevance [16]
Human responsibility: Confirms merges; screens every record that determines inclusion
Verification artefact: PRISMA 2020 flow diagram [1]
- Screening
AI-assisted contribution: Flags likely includes and excludes as a second opinion
Human responsibility: Makes and records every include/exclude decision with reasons; resolves conflicts
Verification artefact: Decision log; inter-rater agreement (kappa) [17, 18]
- Data extraction
AI-assisted contribution: Drafts extraction tables from full texts [12]
Human responsibility: Checks every extracted number against the source text
Verification artefact: Extraction table with page/figure anchors
- Risk of bias and certainty
AI-assisted contribution: None delegated
Human responsibility: Applies RoB 2, ROBINS-I, AMSTAR 2, GRADE or OHAT [4–8]
Verification artefact: Signed appraisal tables
- Citation integrity
AI-assisted contribution: Collects identifiers
Human responsibility: Resolves every DOI in a bibliographic registry and confirms title, authors and year
Verification artefact: Citation verification register
- Conclusions
AI-assisted contribution: Drafts plain-language summaries
Human responsibility: Writes and signs the conclusion, including explicit evidence gaps
Verification artefact: Named, accountable author
Three rules make the workflow defensible:
- No fabricated evidence. Every reference is resolved against a bibliographic registry before publication. A reference that cannot be resolved is removed, and the gap is stated as "no direct evidence identified" rather than filled.
- Every number is traceable. Doses, effect sizes, sample sizes and limits are checked against the abstract or full text, not against an AI summary.
- AI outputs are candidates, not findings. Records, extractions and summaries proposed by AI are inputs to human review. Inclusion, appraisal, certainty and conclusions are human decisions, recorded with the reviewer's name.
5. Regulatory context: the EU AI Act
Regulation (EU) 2024/1689 lays down harmonised rules on artificial intelligence in the European Union [19]. Two provisions are directly relevant to AI-assisted scientific work. Article 4 requires providers and deployers of AI systems to ensure a sufficient level of AI literacy among the staff who operate them. Article 14 requires high-risk AI systems to be designed for effective human oversight, so that a person can understand the system's capacities and limitations, remain aware of automation bias, correctly interpret its output and decide not to use or to override it.
Literature-review support tools are not, as such, listed as high-risk systems. Applying the Article 14 oversight principles voluntarily is nevertheless good scientific practice: it is the same principle that systematic review methodology has always required, namely that the reviewer, not the tool, is accountable for every judgement.
6. Limitations
This is a narrative methodological review, not a systematic review of AI tools. Published evaluations of LLMs in evidence synthesis are recent, often small and model-specific, and model performance changes between versions. The figures cited above describe the models and datasets tested in those studies and should not be generalised to other tools. Any organisation adopting AI-assisted review should validate its own workflow against human reference standards and report that validation.
7. Conclusion
AI can make rigorous evidence synthesis faster. It cannot make it more trustworthy on its own. For food supplements, medical nutrition and novel ingredients, where a dossier must withstand regulatory and clinical scrutiny, the defensible model is human-in-the-loop: established standards (PRISMA 2020, PRISMA-S, RoB 2, ROBINS-I, GRADE, OHAT), AI used only where its output can be verified, every citation and number checked against the source, and named experts accountable for every conclusion.
