“The volume of AI-generated content is growing faster than editorial teams' capacity to review it.”

Adoption of artificial intelligence (AI) by authors has accelerated production of academic articles and increased submission rates to journals, thereby straining review capacity and hurting journal outcome metrics, such as turnaround time and decision accuracy. … As with any operations system, peer review faces substantial stress when the supply of submissions overcomes the capacity to process them. Such a scenario is occurring now as artificial intelligence (AI) tools, especially generative AI and large language models (LLMs), have accelerated authors’ ability to generate research articles. … “Peer review … is under unprecedented pressure. Too many submissions are funnelled through a limited pool of reviewers, creating unsustainable workloads and threatening the effectiveness, speed and fairness of peer review” (Cambridge University Press 2025, p. 11).Abstract Adoption of artificial intelligence (AI) by authors has accelerated production of academic articles and increased submission rates to journals, thereby straining review capacity and hurting journal outcome metrics, such as turnaround time and decision accuracy. Academic journals face an imperative to improve review quality and productivity by incorporating generative AI tools in the review workflow. As a concrete first step toward this idea, we outline a particular workflow that deploys large language models as a structured, trained, frontline reviewer. … ## 1. Introduction and Motivation Peer review is essential to academic research and the profession. It requires multiple expert reviewers with specialized knowledge to devote substantial time in assessing the quality of submitted articles. As with any operations system, peer review faces substantial stress when the supply of submissions overcomes the capacity to process them. Such a scenario is occurring now as artificial intelligence (AI) tools, especially generative AI and large language models (LLMs), have accelerated authors’ ability to generate research articles. For instance, Google’s AI Co-Scientist has successfully generated novel, literature-grounded hypotheses for biology laboratories, some of which aligned with hypotheses developed independently by the same laboratories (Gottweis et al. 2025). … A recent editorial commentary in Information Systems Research underscores that generative AI is already reshaping research practices and disciplinary expectations while warning that the community must develop new norms to manage these shifts responsibly (Gopal et al. 2025). From a recent report by a top publisher: “Peer review … is under unprecedented pressure. Too many submissions are funnelled through a limited pool of reviewers, creating unsustainable workloads and threatening the effectiveness, speed and fairness of peer review” (Cambridge University Press 2025, p. 11). This article articulates the operations crisis confronting academic peer review, develops a formal model to capture the essence of the process, proposes an AI-infused peer review workflow, and applies an extended model to examine under what conditions and to what extent this new approach can help address the crisis (Cambridge …

Journal operations are increasingly strained by growth in low-quality AI-augmented submissions, affecting volunteer editors and reviewers alike. While most of these submissions are rejected, many during the initial deputy editor screen, the burden they impose is substantial. … Our evidence implies that the sharp rise in submission rates at Organization Science since late 2022 was driven by AI. If this trajectory continues, the volunteer-based review system will be unable to keep pace.… Our goal was not to cast judgment on what defines appropriate AI use, but rather to provide evidence on how growing AI usage might be reshaping how we evaluate and promote scientific research. A clear picture is emerging: AI language models, combined with strong publish-or-perish incentives, are pushing our field to produce more rather than better research. Journal operations are increasingly strained by growth in low-quality AI-augmented submissions, affecting volunteer editors and reviewers alike. While most of these submissions are rejected, many during the initial deputy editor screen, the burden they impose is substantial. We believe that a better equilibrium is possible, one in which we work with AI to produce research once thought impossible. Reaching this equilibrium will require more than just better tools, however; it will require better institutions. In this report, we analyze how submission and review activity at Organization Science have changed following the release of ChatGPT, the first of many modern commercial large language models (LLMs), in November 2022. … #### 5.4.1. Making the System Sustainable and the Research Stronger. The most immediate operational challenge is the rise in volume. Our evidence implies that the sharp rise in submission rates at Organization Science since late 2022 was driven by AI. If this trajectory continues, the volunteer-based review system will be unable to keep pace. This poses an important organizational design challenge (Hasan et al. 2025). Journals need tools and policies that maintain volume at sustainable levels or facilitate low-cost rejections of low-quality AI-driven research. We emphasize that the goal is not to screen out research enabled by AI, since our objective is to refine and promote great research on organizations. …

AI is reshaping nearly every newsroom beat, but this transformation has been particularly acute for fact-checkers, tasked with debunking a growing volume of AI-generated falsehoods, and facing a technology that is both a disrupting force and a powerful new instrument. … On one hand, AI has accelerated the production and spread of misleading content. On the other, it has enabled small fact-checking teams to operate at a scale previously unimaginable. … Jiménez Cruz, CEO and co-founder of Spanish fact-checker Maldita and Chair of the European Fact-checking Standards Network, noted that AI has made the job more demanding, often forcing her team to respond to obvious AI-generated viral content.3. How AI is transforming fact-checking AI is reshaping nearly every newsroom beat, but this transformation has been particularly acute for fact-checkers, tasked with debunking a growing volume of AI-generated falsehoods, and facing a technology that is both a disrupting force and a powerful new instrument. In a panel moderated by our own Eduardo Suárez, Clara Jiménez Cruz from Maldita, Tai Nalon from Aos Fatos, and Chris Morris from Full Fact examined how fact-checking is evolving amid rapid advances in generative AI and the proliferation of misinformation. All three described the rise of generative AI as a watershed moment comparable to the early days of the internet: disruptive, democratising, and destabilising. On one hand, AI has accelerated the production and spread of misleading content. On the other, it has enabled small fact-checking teams to operate at a scale previously unimaginable. “We are in danger of getting to a place where no one believes anything they'd read or see or hear anywhere,” said Morris, who is the CEO of the British fact-checker Full Fact. … “Additionally, there were 2.1 million interactions, including likes and shares, on Facebook and Instagram related to AI-powered disinformation. These overall figures help us understand the scale of what is happening and the magnitude of what we are dealing with.” Jiménez Cruz, CEO and co-founder of Spanish fact-checker Maldita and Chair of the European Fact-checking Standards Network, noted that AI has made the job more demanding, often forcing her team to respond to obvious AI-generated viral content. But she also said it has enabled the development of new tools. Maldita and Full Fact have developed systems that use large language models to detect and classify claims across millions of sentences. A newer iteration integrates a generative AI component built around a ‘harms’ model to prioritise the most damaging narratives.

By some estimates, the majority of content being produced on the internet is already created by artificial intelligence. 21 From automated news articles, to hyper realistic images, catchy tunes that sound like your favourite artists, or fake videos, AI-generated content is transforming the web. … The fear is that human-generated, verified content will soon be drowned out by machine-made ‘AI slop’.6. AI slop, deep-fakes, and misinformation By some estimates, the majority of content being produced on the internet is already created by artificial intelligence. 21 From automated news articles, to hyper realistic images, catchy tunes that sound like your favourite artists, or fake videos, AI-generated content is transforming the web. Social media feeds have been particularly afflicted, with a recent Guardian investigation suggesting that nearly one in ten of the fastest-growing YouTube channels globally only show AI-generated video. 22 TikTok has revealed that its platform already hosts more than one billion AI videos, including one of rabbits bouncing on a trampoline (right) that has racked up 25 million views. The fear is that human-generated, verified content will soon be drowned out by machine-made ‘AI slop’. Much of this is not trying to deceive and is in keeping with a long internet tradition of playfulness, surrealism, and satire, but others worry about the wider implications of powerful new AI tools that can generate realistic images and video, especially when it comes to news.

AI has supercharged the march towards the dead internet by dramatically accelerating the pace and scale by which automated tools can develop content. … This scenario is already beginning to materialize, and we are still in the early days of the AI revolution. As the technology advances, it’s not hard to imagine a world in which most online activity occurs without human prompting, oversight, or review.The concept, which was born on message boards over a decade ago, suggests that the internet will eventually consist primarily or entirely of inanimate algorithms speaking to each other, replacing human interaction with “bot” activity. AI has supercharged the march towards the dead internet by dramatically accelerating the pace and scale by which automated tools can develop content. Today, it is not uncommon for an AI agent to autonomously initiate a query and complete that action using materials that were themselves AI-generated. For example, an AI agent could initiate the drafting of a competitive analysis report without human prompting, which it develops by leveraging online reviews, press releases, and social media content, all of which could be generated by other automated tools. The competitive analysis itself may then be fed into an organization’s internal database and leveraged by other internal AI tools to develop more reports, all without human oversight. This scenario is already beginning to materialize, and we are still in the early days of the AI revolution. As the technology advances, it’s not hard to imagine a world in which most online activity occurs without human prompting, oversight, or review. In fact, that world may have already arrived. The Dead Internet Tracker

With AI-generated content making it increasingly difficult to tell reality from fiction, newsrooms are changing the way they handle photos and videos. However, as AI becomes more advanced, fact-checkers say it’s getting hard to keep up. … Today, it’s flooded with AI-generated and fake content, which is disrupting the information ecosystem.Photo by Drew Yang/ Review of Journalism Seeing isn’t always believing. With AI-generated content making it increasingly difficult to tell reality from fiction, newsrooms are changing the way they handle photos and videos. However, as AI becomes more advanced, fact-checkers say it’s getting hard to keep up. “It’s astonishing how fast the technology is developing. There are things that, even a year ago, we relied on to reveal something as fake that we can’t rely on anymore,” says David Michael Lamb of CBC’s visual verification team. “It means that the work of verification is getting harder and harder, almost by the day.” … Lamb started working with the team in the summer of 2024, joining at the start of its development. He says CBC developed it to combat the threat of AI, recognizing that the problem was steadily getting worse. In 2024, about 24 percent of Canadians used social media as their primary source of information, according to Statistics Canada. Newsrooms have also been taking advantage of social media to quickly gain real-time information on developing situations. Today, it’s flooded with AI-generated and fake content, which is disrupting the information ecosystem. “Before, if we saw a video and we wanted to verify its authenticity…it was, ‘Is the video being misrepresented, or has it been edited in some way?’” says Melissa Goldin, an AP news verification reporter and editor who has focused on mis- and disinformation since 2018. “Now, the question has to be, ‘Is it even a real video to begin with?’”

Generative artificial intelligence has changed this constraint by reducing the time and cost required to generate academic style text and enabling large scale submission growth. This introduces a new infrastructure level risk where the critical threat is not only low quality writing but sustained overload of editorial and reviewer capacity. … This collapse occurs when baseline peer review utilization is already high, such that a moderate proportional increase in submissions is sufficient to push reviewer demand beyond available service capacity.Description Scientific peer review functions as the primary quality control mechanism in academic publishing. For decades the system remained manageable partly because producing a full manuscript required significant human effort which naturally limited submission volume. Generative artificial intelligence has changed this constraint by reducing the time and cost required to generate academic style text and enabling large scale submission growth. This introduces a new infrastructure level risk where the critical threat is not only low quality writing but sustained overload of editorial and reviewer capacity. This study frames AI-enabled submission flooding as a system stability problem and models the peer review workflow as a two stage queueing pipeline consisting of editorial screening and peer review. A stress testing approach is applied by increasing manuscript arrival rates above baseline levels and evaluating system utilization and backlog formation and expected review delay. The results show a tipping point behavior where moderate increases in submission volume can push the system beyond its stability boundary causing persistent backlog growth and rapid inflation of review timelines from weeks toward months and years. This collapse occurs when baseline peer review utilization is already high, such that a moderate proportional increase in submissions is sufficient to push reviewer demand beyond available service capacity. The findings indicate that content-based AI detection tools alone cannot prevent collapse under high volume conditions because congestion emerges when arrivals exceed service capacity regardless of manuscript origin. Therefore long term resilience requires defenses that reduce adversarial scaling and restore the balance between attacker cost and defender capacity through stronger identity verification and submission throttling and proof-of-personhood style controls.

Overall, our audit highlights the immediate need for greater transparency and updated editorial standards regarding the use of AI in journalism to maintain public trust.… Despite this prevalence, we find that AI use is rarely disclosed: a manual audit of 100 AI-flagged articles found only five disclosures of AI use. A factuality analysis shows AI-generated articles are 8.2 times more likely to contain hallucinated claims than human-written news. Overall, our audit highlights the immediate need for greater transparency and updated editorial standards regarding the use of AI in journalism to maintain public trust.

When you add a 33% submission increase on top of a reviewer base that is already turning down half its invitations, the math becomes unfavorable for authors waiting on decisions. Editors cannot simply expand the reviewer pool fast enough to absorb that volume. The response, which editors are sometimes reluctant to say out loud, is to screen harder before papers ever reach reviewers. … The editors who scan your title page and abstract for those first ten seconds are doing it with more manuscripts in queue than ever, under more time pressure, with reviewer capacity that has not kept up.… But the rate at which those reviewers accept invitations has fallen sharply. Editors now need an average of 4.5 invitations to secure a single completed review, nearly double the rate from 2018. Per 100 invitations sent, editors wait a combined 407 days for responses from reviewers who ultimately decline or do not reply. That is more than a year of accumulated waiting, burned on people who never reviewed the paper. When you add a 33% submission increase on top of a reviewer base that is already turning down half its invitations, the math becomes unfavorable for authors waiting on decisions. Editors cannot simply expand the reviewer pool fast enough to absorb that volume. The response, which editors are sometimes reluctant to say out loud, is to screen harder before papers ever reach reviewers. Desk rejection rates at many journals are rising not because editorial standards changed but because the editorial filter is now doing more work at an earlier stage. What Journals Are Doing About It … ## What the Flood Means If Your Paper Is Real If you are preparing a genuine clinical study, a well-designed retrospective analysis, or a systematic review that took two years to complete, the submission surge is unfair to you in a specific way. Your paper is competing for editorial attention alongside a much larger volume of thin submissions. The editors who scan your title page and abstract for those first ten seconds are doing it with more manuscripts in queue than ever, under more time pressure, with reviewer capacity that has not kept up. The bar for what catches attention has raised, not because your science got worse but because the noise floor around it got louder. Several things are worth doing before you submit, and some of them are different from what they were three years ago.

As a result, journalists use AI frequently without oversight. They work with the organization’s information using personal AI tools, sometimes free versions that offer no meaningful data protection and could accidentally make sensitive company information public. Then the ‘human-in-the-loop’ is retained as a concept rather than as a practical safety measure. … “The recent crisis at El Espectador in Colombia, where AI-generated misinformation went unnoticed for months, underscores the risks of this oversight gap.Working group member Claudia Báez shares her perspective: “In my experience working with AI in Latin American newsrooms, there is a clear gap between having an AI policy ‘on paper’ and making it widely accessible for everyone or democratizing it. While large legacy and digital organizations often create formal frameworks or transparency statements, these documents are rarely integrated into the newsroom’s daily workflow. As a result, journalists use AI frequently without oversight. They work with the organization’s information using personal AI tools, sometimes free versions that offer no meaningful data protection and could accidentally make sensitive company information public. Then the ‘human-in-the-loop’ is retained as a concept rather than as a practical safety measure. “The recent crisis at El Espectador in Colombia, where AI-generated misinformation went unnoticed for months, underscores the risks of this oversight gap. These examples speak to the importance of ensuring that policy development does not only live on paper but includes active connections to and evaluation of practices. One promising example comes from La Silla Rota, a Mexican legacy media organization, which has created an internal AI policy tool for its team. This simple custom GPT is shared with the newsroom to answer journalists’ questions about when to use AI and when not to. practitioners, AI technologists and development policy actors.”

From a journal's perspective, Stage 1 GenAI usage undermines writing quality as a screening signal. Editors encounter submissions that appear competent at first glance yet remain unready for publication. Hence, submissions focused on superficial writing quality increase screening loads (i.e., more time needed to screen), while at the same time eventually driving up desk rejections (i.e., poor quality submissions), and eventually strain reviewer capacity. … Reviewer capacity. Reviewer time becomes the binding constraint when submission numbers increase faster than the reviewer pool. … Congestion occurs when demand (here: submissions) surpasses capacity (here: review capacity). Journals encounter longer delays and heavier screening loads in the short term, and they risk declining certification quality in the longer term.… 2025) . There is a risk that these errors compound across studies, systematically degrading research quality (Gopal et al. 2025; Ji et al. 2023 ). To address these risks, authors must verify claims and sources, validate code and outputs, and take responsibility for errors. From a journal's perspective, Stage 1 GenAI usage undermines writing quality as a screening signal. Editors encounter submissions that appear competent at first glance yet remain unready for publication. Hence, submissions focused on superficial writing quality increase screening loads (i.e., more time needed to screen), while at the same time eventually driving up desk rejections (i.e., poor quality submissions), and eventually strain reviewer capacity. Journals should clarify disclosure expectations and enforce accountability through strict desk-reject policies for violations such as hallucinated sources, e.g. … 2024) , while Organization Science reports a decline in writing quality over the same period (Gartenberg et al. 2026) . At BISE, we observe submissions that look competently packaged but vary widely in substantive contribution. With absolute volume of submissions still rising, the screening effort on our side is growing. Reviewer capacity. Reviewer time becomes the binding constraint when submission numbers increase faster than the reviewer pool. At BISE, reviewer acceptance of invitations dropped from 66.0% in 2019 to 57.5% in 2025. Time from submission to reviewer agreement rose from 31 to 43 days. On-time review completion settled at 55.7% in 2025, near the historical low. While submissions doubled between 2020 and 2025, the reviewer pool contracted from 288 to 270. … ## Addressing the Congestion Problem Congestion occurs when demand (here: submissions) surpasses capacity (here: review capacity). Journals encounter longer delays and heavier screening loads in the short term, and they risk declining certification quality in the longer term. Simply adopting better tools may not be enough to reach a better equilibrium (Gartenberg et al. 2026) . During our last BISE editorial board meeting, we already brainstormed potential policy bundles that could address the congestion problem. …

A faster draft is not a productivity win if verification and repair cost more than the manual baseline.Measure correction cost Track false support, wrong attribution, omitted evidence, theme instability, reviewer minutes, and time to an approved deliverable. A faster draft is not a productivity win if verification and repair cost more than the manual baseline. Preserve the accepted output, corrections, prompt, model, source manifest, and approval record. Sources and verification

As research output in fields like artificial intelligence continues to accelerate, the volume of submissions to major conferences has reached levels that strain traditional human-centric review processes. For example, the AAAI Conference on Artificial Intelligence experienced a massive influx of papers, nearly doubling from 15,000 in 2025 to over 30,000 in 2026. This growth places an immense burden on human reviewers, leading to potential issues with consistency, depth, and timely feedback.Introduction The academic peer review system is facing an unprecedented scaling challenge. As research output in fields like artificial intelligence continues to accelerate, the volume of submissions to major conferences has reached levels that strain traditional human-centric review processes. For example, the AAAI Conference on Artificial Intelligence experienced a massive influx of papers, nearly doubling from 15,000 in 2025 to over 30,000 in 2026. This growth places an immense burden on human reviewers, leading to potential issues with consistency, depth, and timely feedback. Figure 1: The AAAI-26 AI Review Pilot integrated a custom AI pipeline into the official conference workflow, providing supplementary feedback for every full-review submission.

New work is created in devising effective prompts and in editing outputs. Perceived efficiency gains are variable and additional research is needed to evaluate any real performance gains across a range of common tasks. … There is an unmet opportunity to design new interfaces to support journalistic work with generative AI, in particular to enable the human oversight needed for the efficient and confident checking and verification of outputs. … The last prominent concern deals with the lack of quality control (8.1%, 16 of 196), as respondents worry that outputs from generative AI will not be verified sufficiently. A respondent states: “I worry that we do not hav[e] ‘standards staff’ in place to fact check AI. News organizations could be viewed as more trustworthy if we can show that real people enforce the news standards.”… • Changing Work Structure and Organization. There are a host of new roles emerging to grapple with the changes introduced by generative AI including for leadership, editorial, product, legal, and engineering positions. Almost half of respondents indicated that tasks or workflows have already changed because of generative AI. New work is created in devising effective prompts and in editing outputs. Perceived efficiency gains are variable and additional research is needed to evaluate any real performance gains across a range of common tasks. Overall, these findings underscore the need for training initiatives and for more fine-grained evaluations to measure actual shifts in productivity. • Work Redesign. There is an unmet opportunity to design new interfaces to support journalistic work with generative AI, in particular to enable the human oversight needed for the efficient and confident checking and verification of outputs. Journalists will need well-designed editing interfaces in order to effectively use generative AI for various tasks. … time for an investment in training. Another challenge concerns not having regulation and guidelines in place (11.1%, 22 of 196) or as one respondent put it: “We should have basic guidelines on what kind of things we check when taking on a tool.” The last prominent concern deals with the lack of quality control (8.1%, 16 of 196), as respondents worry that outputs from generative AI will not be verified sufficiently. A respondent states: “I worry that we do not hav[e] ‘standards staff’ in place to fact check AI. News organizations could be viewed as more trustworthy if we can show that real people enforce the news standards.” A potential gateway into dealing responsibly with these concerns and challenges, respondents mention, is by deciding what uses should be banned (15.5%, 34 of 196), which we elaborate further in the next section. Banned Uses

Scientific peer review faces mounting strain as submission volumes surge, making it increasingly difficult to sustain review quality, consistency, and timeliness. … While this established review process has long endured, the rising scale of submissions means we face increasingly overburdened reviewers with more papers assigned, the need to recruit an ever-wider pool of potentially less experienced reviewers, and increasingly compressed timelines.Abstract Scientific peer review faces mounting strain as submission volumes surge, making it increasingly difficult to sustain review quality, consistency, and timeliness. Recent advances in AI have led the community to consider its use in peer review, yet a key unresolved question is whether AI can generate technically sound reviews at real-world conference scale. Here we report the first large-scale field deployment of AI-assisted peer review: every main-track submission at AAAI-26 received one clearly identified AI review from a state-of-the-art system. … Unfortunately, despite this rapid growth, the peer review process has remained largely static, with a large cohort of human reviewers providing detailed reviews and ratings for papers, and a smaller group of senior researchers comparing those reviews and ratings to make final paper acceptance recommendations. While this established review process has long endured, the rising scale of submissions means we face increasingly overburdened reviewers with more papers assigned, the need to recruit an ever-wider pool of potentially less experienced reviewers, and increasingly compressed timelines. Maintaining the quality, consistency, and timeliness of peer review is thus increasingly challenging. For example, the scale of AAAI-26 submissions required the recruitment and oversight of over 28,000 Program Committee members, Senior Program Committee members, and Area Chairs, nearly three times the size of the committee in AAAI-25 [Association for the Advancement of Artificial Intelligence 2025].

Yet, as AI-generated content becomes more sophisticated, it challenges the foundations of journalistic expertise by enabling non-professionals to create news-like materials at unprecedented speed and scale. … One of the recurring themes in our interviews regarding AI experimentation in the newsrooms is how surprised journalists were with the speed AI brings to the table, allowing journalists to perform tasks such as summarizing large amounts of information quickly.… Previous research on digital technologies has already argued that journalistic authority is no longer solely grounded in expertise or institutional backing but must now be continuously renegotiated in the face of external challenges, such as technological advancements and public skepticism (Anderson,, 2008; Robinson,, 2007). As journalism evolves, so too must the strategies by which authority is asserted and maintained. Yet, as AI-generated content becomes more sophisticated, it challenges the foundations of journalistic expertise by enabling non-professionals to create news-like materials at unprecedented speed and scale. Journalists must not only defend their authority against other actors in the digital ecosystem but also against the encroachment of the technologies they are increasingly expected to integrate into their work (Dodds et al.,, 2025; Simon,, 2022) … As this quote reflects, experimentation inside newsrooms often takes place collaboratively. The same editor also explained: “We are doing another test tonight, where we are going to write prompts together. […] We have small working groups where we work together with colleagues” (P3). One of the recurring themes in our interviews regarding AI experimentation in the newsrooms is how surprised journalists were with the speed AI brings to the table, allowing journalists to perform tasks such as summarizing large amounts of information quickly. Regarding the first tests for using AI tools for transcription, a journalist argued, ”It just works super-fast. And it’s more accurate than doing it manually, I think” (P1). AI for generating summaries and quick recaps has seemingly become standard practice, mainly when dealing with complex and time-sensitive topics like the ongoing conflict in Ukraine. …

As research volumes grow and AI becomes embedded across writing, discovery and evaluation, a central question emerges: how can science scale without losing credibility while meeting the challenges of our time? … At Springer Nature, we are using AI to augment human judgement, not replace it. Editorial responsibility remains with people, while technology operates within defined and auditable boundaries.Chief Product Officer Each year, more than three million scholarly articles are added to the global scientific record, shaping decisions across medicine, public policy, climate action, economics and technology. As research volumes grow and AI becomes embedded across writing, discovery and evaluation, a central question emerges: how can science scale without losing credibility while meeting the challenges of our time? For science, scale without trust is not progress. At Springer Nature, trust is not assumed — it is the outcome of deliberate technical and editorial choices about where automation belongs and where responsibility must remain human. … Scientific publishing is fundamentally a validation process. From millions of annual submissions, only a proportion becomes part of the permanent scholarly record. In 2025, we received 3.1 million submissions and published 539,000 primary research articles, showing increased selectivity compared with the previous year and a deliberate focus on rigour and quality, supported by improved systems – not automation. At Springer Nature, we are using AI to augment human judgement, not replace it. Editorial responsibility remains with people, while technology operates within defined and auditable boundaries. This is because we believe, used well, AI improves consistency and sharpens focus. It allows editors and reviewers to direct their expertise where it matters most, without compromising quality, relevance or ethics. Publishing, at its best, makes responsibility explicit through shared standards, transparent decisions and consistent expectations across the whole research cycle. This approach depends on coherence. …

On the editorial side, growth is real (+15% since 2022, reaching 37,000), but far below the explosion in submissions. On the reviewer side, Elsevier mobilised 1.9 million experts in 2025, up from 1.4 million in 2022, a +35% increase. The effort is real, but it is not enough. … The volume of submissions is growing far faster than the number of available reviewers, and nothing suggests this gap will close on its own.3. Editors and reviewers are struggling to keep up On the editorial side, growth is real (+15% since 2022, reaching 37,000), but far below the explosion in submissions. On the reviewer side, Elsevier mobilised 1.9 million experts in 2025, up from 1.4 million in 2022, a +35% increase. The effort is real, but it is not enough. The number of submissions per reviewer rose from 1.93 in 2022 to 2.21 in 2025. In absolute terms, the gap may seem modest. What it measures is a growing pressure on each evaluator: more manuscripts to assess, within deadlines that are not extending, with a time budget that remains, itself, constant. … The data describe a dynamic that shows no sign of stabilising. Two questions follow, and they deserve to be asked seriously. How much pressure can the peer review system absorb? The volume of submissions is growing far faster than the number of available reviewers, and nothing suggests this gap will close on its own. Recruiting more reviewers will not be enough. The answer will likely come, in part, from AI itself: growing assistance in the evaluation process could release capacity, but at the cost of biases and risks that the literature is beginning to document, and which deserve serious collective discussion before the practice becomes entrenched. What data are we still missing? These figures describe volumes. …

42% more content was published per month by AI-using companies in the article-level analysis. … 97% had some form of editing or review process for AI content. … Nearly all surveyed companies edit or review AI content, supporting a human-led workflow rather than autonomous publication.… - 13% reported not using AI to create content. - 89% of respondents at marketing teams with more than 100 people used AI. - 44% used ChatGPT for content creation. - 15% used Gemini for content creation. - 10% used Claude for content creation. - 94 tools were named across responses. - 42% more content was published per month by AI-using companies in the article-level analysis. - 17 articles was the median monthly publishing frequency among AI users. - 12 articles was the median monthly publishing frequency among non-users. - 87% used AI for blog posts. Adoption and tools statistics … ## Writing workflow statistics - 76% used AI for brainstorming topic ideas. (Ahrefs, 2025) - 73% used AI to create content outlines. (Ahrefs, 2025) - 67% used AI to improve or update content. (Ahrefs, 2025) - 97% had some form of editing or review process for AI content. (Ahrefs, 2025) - 80% manually reviewed AI content for accuracy. (Ahrefs, 2025) - 35% worked with subject-matter experts to review AI content. (Ahrefs, 2025) - 30% worked with an editor. (Ahrefs, 2025) - 14% asked AI to review AI-created content for accuracy. (Ahrefs, 2025) - 4% published primarily pure AI-generated content. … ## What the current evidence shows - The available evidence through September 10, 2026 shows high adoption, especially for ideation, outlining, updating and blog production. - Faster output is the clearest reported benefit; comparatively few respondents say AI directly improves content quality. - Nearly all surveyed companies edit or review AI content, supporting a human-led workflow rather than autonomous publication. - Ranking studies use probabilistic content detection and observational SERPs; they cannot prove that AI authorship alone caused ranking differences. - For current planning, accuracy review, source verification, expert input and editorial judgment remain the critical controls. In summary

For us now it is very important to evaluate what we have done so far and rethink what we can realistically invest in in terms of human resources, financial resources and technology. AI powered technologies are evolving faster than the capacities of small newsrooms and organisations.… translation were also mentioned as areas of interest. These tools would aid in monitoring social media platforms, curating relevant content, verifying information, and translating content across different languages. The goal is to improve news production, content quality, and audience engagement. Others, especially smaller newsrooms are assessing their use of AI and working on aligning their future strategy with the resources available to them: For us now it is very important to evaluate what we have done so far and rethink what we can realistically invest in in terms of human resources, financial resources and technology. AI powered technologies are evolving faster than the capacities of small newsrooms and organisations. We are currently conducting an internal discussion to strategise our next steps in terms of AI related activities both in our newsrooms and training and support programmes for other small independent media in the region. 4.1 The Need for Education and Training

The problem is that output volume scaled instantly while review capacity didn't. One writer can now generate ten times the drafts, but a human editor still takes the same amount of time per piece. … The AI content review backlog is structurally different because generation scales instantly while review capacity stays fixed. Traditional editorial backlogs grew linearly with team output. AI tools let one writer produce 10x the drafts overnight, creating 3-5x backlogs within months.Why AI Content Generators Created a 5x Review Backlog AI tools like Jasper, Copy.ai, and ChatGPT let small teams produce content at a scale that would have required entire departments five years ago. The problem is that output volume scaled instantly while review capacity didn't. One writer can now generate ten times the drafts, but a human editor still takes the same amount of time per piece. Studies show content teams using AI generators report review backlogs growing 3x to 5x within the first six months of adoption. The Fundamental Asymmetry Between Creation and Verification … ### AI content review backlog vs traditional editorial backlog? The AI content review backlog is structurally different because generation scales instantly while review capacity stays fixed. Traditional editorial backlogs grew linearly with team output. AI tools let one writer produce 10x the drafts overnight, creating 3-5x backlogs within months. The gap isn't process, it's that human review has cognitive limits AI generation doesn't. What happens when review queues grow faster than your team can approve?

The arrival of AI is accelerating dynamics that the system was already struggling to manage, and generating debates that, like the crisis narrative itself, often move faster than the evidence beneath them. … Breaches of information ethics or publishing standards / policies throughout the research lifecycle are growing, along with a steadily rising number of new manuscripts. And the advent of intelligent automation makes fraudulent research that much easier to produce. … We do not yet have reliable ways to measure AI usage across the peer review lifecycle, for authors, reviewers, or editors, which means policy is currently running ahead of evidence.… New research communities are emerging across Asia, Africa, and Latin America, bringing capacity and energy that the system has not yet learned to use well. But the rate at which scholars accept review invitations fell by nearly half in the same period. It now takes an average of 4.5 invitations to secure a single review. The arrival of AI is accelerating dynamics that the system was already struggling to manage, and generating debates that, like the crisis narrative itself, often move faster than the evidence beneath them. The pipeline has not broken. But it has slowed generally, if unevenly, and in ways that are impossible to ignore. (cont.) … The fundamental purpose of peer review is quality control. However, the system is strained by increasing cases of falsified research data, misrepresentation of authors’ identities, and other forms of research dishonesty — compounding the long timelines to editorial decisions and declining rates of reviewer engagement. Breaches of information ethics or publishing standards / policies throughout the research lifecycle are growing, along with a steadily rising number of new manuscripts. And the advent of intelligent automation makes fraudulent research that much easier to produce. Most research integrity interventions rely on checking manuscripts against various criteria and flagging failed checks for editors’ final decision-making. … Conclusion As with any analysis, this report sparks an open set of questions, which will be important to keep asking as peer review continues to evolve. We do not yet have reliable ways to measure AI usage across the peer review lifecycle, for authors, reviewers, or editors, which means policy is currently running ahead of evidence. The data on reviewer pool fatigue surfaces patterns — geographic, demographic, gendered — but it is less clear what interventions those patterns should inform, or who has the standing to make them. That work cannot happen without the people it affects. …

In newsrooms specifically, the friction has several sources. AI-drafted content requires human review, which takes time. Factual accuracy concerns have led to cautious editorial processes.Editors at Bohiney Magazine note that the gap between AI promise and AI practice in newsrooms follows a pattern familiar from other technology adoption cycles. Initial enthusiasm identifies theoretical productivity gains; actual implementation reveals friction, error rates, training costs, and quality-control requirements that substantially reduce the net benefit. In newsrooms specifically, the friction has several sources. AI-drafted content requires human review, which takes time. Factual accuracy concerns have led to cautious editorial processes. Legal liability concerns have produced conservative deployment patterns. Union negotiations have slowed adoption in newsrooms where contracts require bargaining over AI deployment. The New-Role Phenomenon

When generic content becomes easier to produce in an AI era, newsroom advantage shifts toward what is harder to replicate: original reporting, trusted relationships and tighter audience–journalist communities. … 43% of newsrooms agree that advances in AI will reduce the number of people employed in their workplaces over the next three years even as 39% of newsrooms expect their overall editorial outputs to increase.◈ Trust is moving from institutional authority to relational signals. Reporters spend 38% of their time on production, but only 11% on post-publication work such as community engagement. This limits the ability to build visible expertise and direct audience relationships. When generic content becomes easier to produce in an AI era, newsroom advantage shifts toward what is harder to replicate: original reporting, trusted relationships and tighter audience–journalist communities. Our Future Newsrooms Study, which draws on 448 survey responses from newsrooms across 86 countries, shows that newsrooms broadly understand the direction of travel — engagement, clearer trust signals, AI, audience focus and new formats. Still, they face four interconnected gaps that prevent them from turning understanding into consistent execution: a strategy gap, an audience trust gap, a capability gap and a skills gap. … During our interviews, multiple newsroom leaders told us that a more robust conceptualisation of AI success must consider how to tie efficiencies to an overarching objective that can drive audience or commercial value. In the current state, however, newsrooms appear to be defaulting toward time savings as the main indicator of AI success, and this risks orienting newsrooms toward doing the same work, just faster and with fewer people. 43% of newsrooms agree that advances in AI will reduce the number of people employed in their workplaces over the next three years even as 39% of newsrooms expect their overall editorial outputs to increase. As newsrooms look to double down on audience engagement, it will be important to move toward a more mature approach to AI that measures its ability to enable journalism that wasn’t previously achievable, such as pursuing investigations that weren’t possible in the past or identifying a way to interact with audiences that couldn’t be served effectively in the past. Slightly or strongly agree

Risk of an increasing number of publications which can be quickly produced but are of LITTLE or NO RELEVANCE. E.g. systematic reviews with searches, data extraction and synthesis and even discussion done by AI which do not address relevant health research questions or combine data in a meaningless way.… Research "Familiarity with the system and how to enhance use of AI to complement research." Thematic Analysis of Medical Journal Editors Regarding Chief Journal/Publisher Policy on AIC use No AI in Authorship or Peer Review | This theme refers to policies prohibiting the use of AI in authorship or peer review. "Do not use AI for editorial purposes." This theme refers to policies allowing AI use for improving grammar or language clarity. Permitted For Language/Grammar Use Risk of an increasing number of publications which can be quickly produced but are of LITTLE or NO RELEVANCE. E.g. systematic reviews with searches, data extraction and synthesis and even discussion done by AI which do not address relevant health research questions or combine data in a meaningless way." This term refers to the Medical Editors' perceptions of Integration, ethical and confidentiality implications with the use of AICs in the scholarly publishing process. "If we input unpublished papers into AI, isn't that problematic because then AI can use the content in AI searches?" References

Clarkesworld magazine (Feb 2023) suspended new short-fiction submissions after a flood of AI-generated submissions; editor Neil Clarke reported a ratio of roughly 700 legitimate to 500 AI submissions per day at peak [Historical] … Clarkesworld 2025 submission volume — the magazine reopened with manual screening rather than automated AI detection. Neil Clarke (Aug 2025): "the US is now the primary producer of slop submissions."500+ banned in one day Clarkesworld magazine (Feb 2023) suspended new short-fiction submissions after a flood of AI-generated submissions; editor Neil Clarke reported a ratio of roughly 700 legitimate to 500 AI submissions per day at peak [Historical] 14,805 legitimate submissions in 2025 (~1,233/month) Clarkesworld 2025 submission volume — the magazine reopened with manual screening rather than automated AI detection. Neil Clarke (Aug 2025): "the US is now the primary producer of slop submissions." Top 3 reader objections to AI covers

Content teams spent years looking for ways to produce more. More articles, more posts, more formats, more [add your option]. The bottleneck was always capacity. AI removed it. … Something shifted in the market around 2022. Content volume went up everywhere, and fast. This is the kind of change that happens when a constraint disappears overnight. The constraint was production capacity, which was removed by AI. Suddenly a team of three could publish what a full editorial department used to. So they did. And so did everyone else. … Another unsolved problem is brand alignment. Every team knows they need AI to sound like them, but almost none have a system for it. Some teams rely on detailed prompts or brand guidelines pasted into every session. Some teams have a single editor who "knows the voice." The irony: the teams spending the most on content production are the ones with the least infrastructure to protect what makes their content theirs.// Why this report exists Content teams spent years looking for ways to produce more. More articles, more posts, more formats, more [add your option]. The bottleneck was always capacity. AI removed it. This report covers what actually happened when B2B tech companies got access to unlimited content production. We published this report for a simple reason: the market needs more signal from people doing the work. And we need content not for the sake of content. … We had a strong feeling — Everyone published more. Nobody asked why. Something shifted in the market around 2022. Content volume went up everywhere, and fast. This is the kind of change that happens when a constraint disappears overnight. The constraint was production capacity, which was removed by AI. Suddenly a team of three could publish what a full editorial department used to. So they did. And so did everyone else. Content volume … AI removed the operational ceiling and raised the quality bar at the same time. Most companies only solved the first problem, but the second one became their biggest liability. And it's the one you can't automate your way out of. Another unsolved problem is brand alignment. Every team knows they need AI to sound like them, but almost none have a system for it. Some teams rely on detailed prompts or brand guidelines pasted into every session. Some teams have a single editor who "knows the voice." The irony: the teams spending the most on content production are the ones with the least infrastructure to protect what makes their content theirs. New challenge (post-AI)

AI isn’t replacing all human content—but it is shifting the volume curve fast. … In 2025, AI content volume jumped to 42 while human content fell to 58. - AI is projected to surpass human content volume by 2026 (56 vs. 44). - Human-written content declined every year from 2021 to 2025.``` AI isn’t replacing all human content—but it is shifting the volume curve fast. Essential Statistics - In 2021, AI content accounted for just 0% of indexed content volume. - By 2023, AI content rose to 9 while human-written dropped to 91. - In 2025, AI content volume jumped to 42 while human content fell to 58. - AI is projected to surpass human content volume by 2026 (56 vs. 44). - Human-written content declined every year from 2021 to 2025. - AI content grew consistently each year since 2020. - The lines cross in 2026, marking the tipping point. Key Takeaways