Something I have realised lately is that it’s no longer a question of whether AI-assisted screening tools will find their way into manuscript submission and peer review systems. It’s more of a question of how they can be implemented responsibly and where they add the greatest value. Across the industry, platform after platform is moving in the same direction: wiring automated quality, language, and integrity checks directly into the editorial workflow so they run the moment a manuscript is submitted, rather than as a separate step bolted on afterwards.
When that integration happens for a publisher, the instinct is often to treat it as a turning point– the moment editorial work starts being replaced rather than supported. We don’t see it that way. The distinction between replacing editorial judgement and supporting it matters because it shapes how publishers prepare for what’s next.
Integration Was Always Coming
Screening tools that flag formatting issues, statistical irregularities, and potential integrity concerns are not new. What’s changing now is the plumbing: these checks are increasingly built into submission platforms themselves rather than run as a separate, optional layer, which is a sensible and predictable evolution. Once a capability exists and proves useful, building it into the core workflow is simply how software develops.
At PA EDitorial, we made a deliberate choice some time ago not to compete in that space by building our own automated screening tool to sit alongside those already on the market. Our choice wasn’t one of caution but rather reflected a clear view that the highly standardised administrative layer of manuscript handling was always going to move towards automation and, in many cases, offshore, regardless of which vendor got there first. Instead, we’ve focused our experience on developing services around this reality, knowing that the shift is coming, so we can support all the publishers and editors who rely on us to know where the ground is actually moving.
So, when integration happens, for any publisher on any platform, it doesn’t ask us to rethink our position; it confirms it.
What ‘Humans–in-the–Loop’ Actually Means
The phrase ‘humans-in-the-loop’ gets used a lot in conversations about AI, often as a reassurance rather than a description of anything specific. In an editorial context, it should mean something precise because the vaguer version of the phrase contributes very little to these conversations.
This human isn’t someone who rubber-stamps an AI-generated flag because checking it properly would take too long. It’s someone with the authority, context, and accountability to overrule the system when it’s wrong, escalate when it’s right but the situation calls for more than a standard response, and decide, case by case, what happens next; an example of this is our PA Triage Editor. If that person doesn’t have real authority to change the outcome, they’re not exercising editorial judgement; they’re simply confirming a decision that’s already been made.
Publishers need to get this distinction right as integration becomes the norm rather than the exception. It’s the difference between a workflow where a flagged manuscript sits in a queue for someone to glance at and approve, and one where a flagged manuscript reaches someone with the standing to ask follow-up questions, request clarification from an author, escalate to an editor-in-chief, or simply decide the flag doesn’t apply and move on. The first version satisfies a checkbox, and the second protects the quality and integrity of what gets published.
What We’ve Learned from Early Adoption
We’ve watched AI-assisted screening tools operate in live editorial environments across the industry, and the pattern has been fairly consistent regardless of the specific platform or tool involved. Manuscripts get returned to authors over genuinely minor issues, flagged with a precision that technically isn’t wrong but contextually misses the point. The author-facing communication often feels mechanical: efficient on paper, frustrating in practice, and occasionally tone-deaf in exactly the moments where tone matters most to someone whose career may hinge on the outcome.
None of the above is a criticism of the technology in isolation. Screening tools do what they’re built to do: they screen. The trouble starts when screening is mistaken for judgement, and a flagged inconsistency is treated as a settled verdict rather than the opening question it actually is.
The Paradox Nobody Wants to Hear
There’s a comfortable assumption running through a lot of commentary on AI in publishing: that automation reduces editorial workload. We think the opposite is closer to the truth, certainly in the medium term and probably for longer than current forecasts assume. AI-assisted workflows will screen manuscripts, flag potential issues, and generate recommendations at a volume and speed no editorial office could match manually. That’s genuinely useful. But every flag, every recommendation, and every automated output still has to be interpreted by someone who understands what it means in context: for this journal, this discipline, this author, this specific situation.
Someone has to decide whether a flagged issue is a real problem or a false positive shaped by a model that doesn’t know the field’s conventions. Someone has to determine what action follows from that decision. Someone then has to communicate it to an author in a way that doesn’t read as though it was generated by the very system that flagged the problem.
If integration increases the sheer number of things that get flagged, the volume of decisions requiring genuine human judgement doesn’t shrink. It grows. That’s the part of the conversation that tends to get skipped over in favour of a simpler story about automation freeing up everyone’s time.
What AI Can Do and Where Do People Still Have to Step In?
It’s worth being specific about where the line actually falls because vague reassurance helps no one, and overclaiming helps no one either.
AI can screen. People still need to interpret.
A screening tool can scan a manuscript against a rule set faster and more consistently than any human team, identifying potential issues such as scope mismatches, reporting omissions, ethical concerns or signals that merit closer investigation. However, these outputs are indicators, not conclusions.
A manuscript that appears to fall outside the scope could represent an emerging area of research; a reporting omission may be entirely appropriate for the study design; an image duplication or unusual similarity score may have a legitimate explanation that requires investigation rather than automatic rejection. Equally, a manuscript may satisfy every screening criterion yet still lack the novelty, relevance or scientific contribution expected by the journal.
Automated tools are also limited in their ability to recognise more nuanced patterns, such as unusual reviewer suggestions, inconsistencies across multiple submissions or behaviours that simply warrant a closer look from an experienced editor. Screening tools will inevitably produce false positives, and someone still needs to determine whether a flag represents a genuine concern or merely an artefact of the screening process. Determining what any of these signals actually mean requires editorial judgement, informed by the journal’s policies, the discipline’s expectations, and the broader context of the research field.
AI can flag. People still need to investigate.
Flagging is pattern-matching, but investigation is judgement: following a flagged concern back to its source, establishing whether it reflects a genuine research integrity issue, an honest error, or simply an edge case the model wasn’t built to recognise.
There’s another consideration, too. If we become overly reliant on automated screening alone, those intent on undermining research integrity will inevitably adapt. As AI detection tools become more sophisticated, so will the techniques used to evade them. It then becomes an ongoing arms race between systems designed to identify problems and those designed to avoid detection. Human editorial judgement provides the critical layer that can recognise context, question unexpected results, and identify patterns that even the most advanced automated tools may miss.
These differences matter most in the cases where the stakes are highest. Research integrity questions rarely come as clean, binary signals. A duplicated figure might be a genuine concern or a formatting artefact carried over from a previous draft. A statistical anomaly might point to fabrication, or it might point to a legitimate methodology the model simply hasn’t encountered often enough to recognise as normal. Treating every flag as equally serious – or equally dismissible – is its own kind of failure, and it’s a failure no screening tool is currently equipped to avoid on its own.
AI can recommend. People still need to decide.
A tool can surface a suggested course of action based on patterns it has identified or criteria it has been trained to recognise. But a recommendation is not a decision. Determining whether an apparent concern reflects a genuine issue, an innocent anomaly or a false alarm generated by the underlying algorithm requires scientific expertise, editorial experience, and an understanding of the wider research context. It also requires consideration of factors that are difficult to quantify: the significance and potential impact of the research, the consequences of accepting flawed work or incorrectly rejecting valuable contributions, and the human reality of getting the decision wrong in either direction. These are judgements that cannot be reduced to an algorithm alone
AI can process. People still need to communicate.
Processing volume is exactly what these tools are good at. Explaining an automated decision to an author or editor in a way that feels considered rather than formulaic is something else entirely. That remains a distinctly human skill.
AI can accelerate. People still need to take responsibility.
Speed is the genuine, uncomplicated benefit of deeper integration. What speed doesn’t do is absorb accountability. When a decision affects a researcher’s career, a journal’s reputation, or the integrity of the published record, responsibility sits with a person who can be asked to explain it, not with a model.
Why Authors Won’t Accept a Robot Gatekeeper – and Shouldn’t Have To
There’s a further dimension that’s easy to overlook from within a platform roadmap: how all this affects the people on the receiving end. Will academics be comfortable with a world where a manuscript can be effectively rejected before a human being has looked at it? This isn’t an emotional objection by the author; it’s a reasonable expectation that a system this consequential to someone’s career should have a person accountable for the outcome. As an industry, we should expect and welcome some pushback wherever that line gets blurred, because it reinforces the distinction that matters: automation can screen, but it shouldn’t get the final word.
What This Means for Publishers, Practically
For editorial offices weighing up how to prepare, the practical question isn’t whether to adopt AI-assisted screening at some point – that decision is rapidly being made for them by the platforms they already use. The more useful question at this stage is how to structure the human layer around it: who reviews which categories of flag, what authority they have to override a recommendation, and how decisions are communicated so authors experience a considered editorial process rather than an automated checkpoint with a person’s name attached after the fact.
That’s the design work that matters most as integration becomes standard, and it’s work that becomes more relevant, not less, as the underlying tools improve. A fully integrated screening layer doesn’t remove the need for that work; it simply changes where the work starts.
Holding the Line on the Message
Peer review management has always depended on someone applying context to outputs, whether those outputs come from a junior editor’s first pass or an AI model’s automated triage. This hasn’t changed, and we don’t expect it to.
There is now broad acceptance across the industry that AI will absorb many of the routine checks editorial teams currently perform by hand. What it still struggles with – and what we don’t expect it to resolve any time soon – is context: policy, discipline-specific convention, research integrity questions that don’t fit a template, and situations that require someone to pick up the phone rather than send an automated notice. Confidence in the published record depends on someone being able to stand behind a decision, explain it, and adjust it when the situation calls for nuance a model can’t supply.
The role of the editorial office will change as integration becomes the norm. The balance of work within it will shift, probably faster than many publishers currently expect. But the need for experienced people to interpret AI outputs, manage exceptions, and provide recommendations that publishers can trust isn’t disappearing; it’s becoming more visible, precisely because the routine work around it is being automated away.
Confidence is, in the end, what publishers want and need when they invest in editorial support: confidence that a flagged manuscript was examined properly, confidence that a difficult call was made by someone qualified to make it, and confidence that the published record holds up to scrutiny long after the screening report has been filed away. AI can generate a great many things, but it can’t generate that kind of confidence on its own. That kind of confidence has always required people, and we don’t see deeper integration changing that.
Our role is to help publishers use these technologies well, applying editorial judgement where it matters most and ensuring automated screening supports, rather than replaces, informed decision-making.
About PA EDitorial
At PA EDitorial, we are committed to providing a customer experience that prioritises professionalism. We welcome inquiries from individuals and businesses, so feel free to contact us to start working together.
You can learn more about PA EDitorial services here.