A professor from Harvard proposed the use of artificial intelligence agents to deal with the overload of scientific publications. Avi Loeb, astrophysicist and director of initiatives at the university, highlighted in a recent article that academia is facing a crisis due to the excessive volume of papers, which prevents adequate reading and verification. Essa suggestion comes amid global discussions about how AI can transform traditional processes in research.
Maria Roginskaya, mathematics teacher at Universidade of Tecnologia of Chalmers, at Ela argues that the peer review system, carried out without remuneration, often results in superficial assessments, leading to undetected errors and repetitions. Loeb agrees and sees AI agents as a tool to process and organize this content efficiently.
The proposal gains relevance in 2026, the year in which experts predict significant advances in agentic AI, capable of acting autonomously in complex tasks. Relatórios indicate that failures observed in 2025, such as inefficiencies in 70% of corporate tasks, are being overcome with improvements in open protocols and models. In the academic context, this could separate knowledge dissemination from career assessment, reducing bias and optimizing resources.
AI agents and their recent evolution
Research shows that AI agents have failed in many applications by 2025 due to limitations in efficiency and safety. However, advances in 2026 promise greater autonomy, with 64% of Brazilian leaders believing in accelerated adoption.
These systems now integrate better with existing tools, enabling actions such as real-time data analysis. In academia, this means the potential to review papers without constant human intervention, easing the burden on researchers.
Impacts on peer review
The current peer-review system is criticized for relying on editors and reviewers with limited interests. Jovens researchers face disadvantages, needing connections to access prestigious journals.
AI agents could assess plausibility and originality impartially, reducing the time spent on errata, which are numerous due to incorrect publications.
Furthermore, the separation between publication and career assessment would avoid the strategic fractionation of results to maximize CVs.
Studies indicate that AI is already used in 90% of marketing companies for similar tasks, suggesting feasibility in academia.
Practical applications in specific fields
In mathematics, the volume of 100,000 papers annually exceeds human reading capacity, leading to valuable work being ignored. Agentes of trained AI could identify overlaps and recommend relevant readings.
In astrophysics, the area of Loeb, AI already helps in the analysis of astronomical data, and its extension to publications could optimize international collaborations.
Reports from 2026 indicate that 40% of agentic AI projects face cost cancellations, but successful ones reduce errors by up to 60%.
Institutions such as Brazilian universities are beginning to regulate the use of AI, avoiding total bans and encouraging co-creation.
Challenges in academic implementation
Resistance comes from those who benefit from the status quo, such as established publishers. Mudanças demand collective action to avoid system collapse. Loeb emphasizes that AI should be tested in evaluations of verified papers. Estudos show that AI skills change 66% faster in exposed roles, requiring urgent academic training. No Brasil, the training deficit threatens projects until 2026, with few institutions training specialists.
Global companies invest in agents for autonomous tasks, but academia needs governance for sensitive data. Relatórios foresee a multiplication of agents in 2026, with protections for integration into daily work.
Alternative proposals for the system
Historical alternatives such as 14th-century open competitions are mentioned, but AI offers modern scalability. Na França, centralized assessment systems inspire hybrid models. Loeb’s proposal includes AI to digest papers, separating dissemination from career metrics. Isso would reduce abuses such as counting submissions as merits. Pesquisas indicate that generative AI is already transforming content production, with 68% of use in brainstorming.
In academia, this could focus research on pure knowledge, not strategic publications. Agentes freelancers, with 70% failures in 2025, evolve to solve complex tasks, such as investment analysis or consultancy. Aplicado to papers, this would avoid repetitions and increase quality. No Brasil, 95% of agents failed due to lack of data, but 2026 brings a focus on governance.
Expected advances in technology
Open models and new browsers mark 2025 as an inflection point for agentic AI. In 2026, these systems will take active roles in operations, boosting investments. In academia, this means impartial review, reducing human biases. Relatórios of Gartner warn of cancellations, but marketing successes show potential. Integration with CRM and CMS already occurs, with 54% of chatbots resolving queries. Para publications, AI could automatically index and categorize, facilitating global access.
Benefits for young researchers
Young scientists depend on supervisors for publications, creating inequalities. Agentes of AI would level the playing field, evaluating merit independently of connections.
This would encourage innovative, non-strategic research, with AI identifying gaps in existing literature.
Integration with higher education
Brazilian universities face a regulatory vacuum, with only seven institutions with rules for AI in 2025. UFC’s Diretrizes prohibit the generation of original content, but allow co-creation.
Agents could assist in similarity verification, complementing tools such as Turnitin, which do not attest to originality alone.
This promotes ethical use, preparing students for a market where AI is essential.
Global Perspectives on Adoption
In the corporate world, 90% of professionals use agents in pilots, focusing on content production. In academia, similar adaptations could resolve crises of scale.
IEEE studies predict 2026 as the year of agents, changing human-machine relationships. No Brasil, acceleration in innovation is expected despite challenges in the making.
Solutions for information overload
The inability to read fractions of papers leads to dependence on personal references. Agentes of AI would organize content, recommending readings based on relevance.
This would avoid collapse by separating knowledge from career assessments, as proposed by Roginskaya.
Examples of crashes and fixes
In 2025, agents failed in 70% of professional tasks, according to Carnegie Mellon. Melhorias in 2026 focus on accuracy, with APEX-Agents measuring performance.
In academia, similar tests would validate AI for review, reducing errata and increasing quality.
Preparation for systemic changes
Institutions need to coordinate with government and industry for AI training. No Brasil, first generation of AI graduates emerges, but demand exceeds supply.
This affects CIO schedules, with skills changing rapidly, requiring investments in training.
The discussion about AI agents in academia reflects a transition to greater efficiency. With anticipated advances, the publication system can adapt, benefiting global researchers.
