AI Runs First Fully Automated Scientific Conference

Artificial intelligence has successfully managed an entire scientific conference, from paper writing to peer review and presentation, marking a major step toward machine-driven scholarship. The HIKMA experiment demonstrated that AI can handle the full research-to-publication pipeline while maintaining transparency and intellectual property safeguards, offering a blueprint for how AI might support—not replace—human researchers in the future.

Researchers developed a multi-agent framework called HIKMA that autonomously conducted a semi-autonomous scientific conference. The system processed 60 datasets, generated corresponding research papers, managed peer review through AI agents, revised manuscripts based on feedback, created presentation slides, and delivered talks using AI avatars. Thirty manuscripts progressed to full acceptance and archival, showing that AI can replicate many stages of academic communication under human oversight.

The methodology followed a structured eight-stage workflow mirroring traditional academic publishing. It began with dataset intake from platforms like Kaggle, where 60 datasets were registered with license checks and intellectual property tags. The AI then generated manuscripts using constrained prompts derived from dataset metadata, ensuring content remained grounded in the source material. Each paper underwent dual peer review by two AI agents—one providing constructive feedback and another adopting a skeptical, stress-testing approach—producing 120 total reviews. Manuscripts then entered a revision loop where the AI incorporated reviewer comments, tracked changes in red text, and generated formal response letters. Accepted papers were finalized into camera-ready versions with standardized formatting, watermarks, and fictional author attributions to maintain anonymity. The system automatically converted these into presentation slides and scripts, then used HeyGen's platform to create avatar-narrated videos with synchronized speech and lip movements. All artifacts were archived with cryptographic hashes for verification.

Results showed the pipeline successfully processed 30 manuscripts to completion, with each stage documented in a central workbook for transparency. The AI-generated papers averaged 7-9 pages and followed conventional academic structure. Review scores varied—for example, one paper scored 8.0 and 5.0 from the two reviewers, leading to a 'Weak Accept' decision—demonstrating the system's ability to handle divergent evaluations. The revision process expanded manuscripts by approximately 20% to address methodological clarifications and robustness requests. All outputs, including papers, slides, and presentation videos, were publicly released on a conference website and through a podcast series, ensuring full accessibility and reproducibility.

This experiment matters because it showcases how AI could handle time-consuming academic tasks while preserving research integrity. For regular readers, it means scientific knowledge could be generated and shared faster, with automated systems assisting in literature review, writing, and even conference presentation. The framework maintains intellectual property protection through license verification and watermarking, addressing concerns about unauthorized data use. It also demonstrates practical applications for educational content creation and multilingual science communication through AI-generated presentations.

However, limitations remain. The system cannot establish true authorship accountability since all content is AI-generated with fictional attributions. Bias correction is incomplete, as AI models may reproduce linguistic and cultural biases from their training data. While license parsing helps flag intellectual property risks, it cannot fully prevent misuse of sensitive material without manual oversight. The experiment also highlighted epistemic risks—over-automation might reduce reflective judgment that defines human scholarship. These unresolved issues underscore the need for continued human-AI collaboration rather than full automation.

The HIKMA conference provides a tangible model for how AI could integrate into scholarly workflows, balancing efficiency with ethical safeguards. Its lessons emphasize that trust in AI-generated scholarship depends on rigorous documentation, transparency, and clear boundaries between machine and human contributions.