#AI spent 17 hours writing a 30 page academic paper! Self selected topic, including experiments, and in compliance with APA format standards
It’s not about piecing together knowledge points, AI is really conducting research this time.
An AI system called Virtuous Machines spent 17 hours, $114, conducted experiments with 288 real people, and wrote a 30 page academic paper.
And it’s also fully automated from topic selection to manuscript completion!?
Let’s take a look at what this AI has written.
##AI automation for scientific research: from inspiration to publishing papers
####Engage in scientific research like humans
This paper independently completed by AI belongs to the field of cognitive psychology, specifically focusing on research directions related to human visual cognition.
And it’s not just random writing, it relies on human scientific research methods.
Firstly, research questions were proposed based on cognitive psychology theory, such as “Is there a relationship between visual working memory and psychological rotation ability?” and “What is the impact of psychological image clarity on visual cognitive task performance. Visual working memory refers to the ability of humans to maintain and process visual information, involving the processes of information storage, manipulation, and retrieval; psychological rotation refers to the cognitive process of achieving spatial object rotation through psychological manipulation to complete perceptual matching
Next, the experimental plan was designed, taking into account sample size calculation, control variables, and the use of the VVIQ2 scale to measure the clarity of psychological imagery of the subjects;
After determining the experimental plan, it also recruited 288 subjects through the online platform Prolific, and collected 277 valid data (some subjects did not complete the experiment and were screened out by AI). It continued to write Python code for 8 hours and processed the data using repeated measures of variance.
In the process of analyzing data, even identifying outliers and adjusting statistical models;
When organizing the results, it is possible to cite over 40 authentic articles on PubMed and Semantic Scholar, and even the “Methods”, “Results”, and “Discussion” sections of the paper comply with APA formatting standards.
How did you achieve such high efficiency?
####AI teams with different functions, dedicated personnel
Let’s take a look at the architecture of this AI system.
Its independent research capability stems from the technical design of * * collaboration+simulating human cognitive mechanisms+dynamic knowledge exchange * *.
In collaborative architecture, the Master is the core control module that oversees the entire system.
Other AI assistant modules focus on specialized tasks such as literature search, data analysis, and experimental design.
The underlying ability foundation that supports the functioning of human cognitive mechanisms is like an onion ring.
The core is the ability to retrieve knowledge accurately from massive academic databases; Next is the ability to abstract and refine, which can summarize general logic from numerous specific studies; Furthermore, there is the ability of metacognitive reflection, which enables AI to self examine, such as asking whether the methods are appropriate and whether the conclusions and hypotheses are logically coherent after completing data analysis.
Next is the task decomposition ability, which breaks down scientific research projects into executable small tasks, such as writing papers into literature review, experimental design, and other stages; There is also the ability for autonomous iteration, without the need for human intervention. AI will repeatedly modify paper drafts and debug crashed code until satisfactory.
The outermost layer is the ability of multi-agent collaboration, with teams of AI assistants with different functions, allowing the system to have dedicated personnel for literature retrieval, experimental design, and data analysis.
In addition, there is also a * * d-RAG * * real-time memory bank that allows users to search for the latest literature while recalling their previous research. New and old knowledge can be interactively integrated.
Under this framework, writing a 30 page paper in 17 hours can be considered a pinch in the hand.
####Fast speed, but there are also minor drawbacks
However, although this AI is impressive, it is not perfect either.
Although the advantages are obvious: efficiency is more than 10 times faster than human teams, data analysis is rigorous enough to reject statistical significance traps (even if p<0.05, if the effect size is too small, it will indicate that the “actual significance of the results is limited”), and it can also handle noisy data in real experiments.
But it occasionally leads to theoretical misunderstandings, such as labeling existing research conclusions as first-time discoveries; Omitting the Y-axis units of the chart and mixing “cross trial interval” and “stimulus presentation interval”.
It can only be said that AI’s research speed is quite impressive, but it seems almost impossible to completely replace the theoretical depth and innovative thinking of human researchers at present
*Research Address: https://arxiv.org/abs/2508.13421
Reference link: https://x.com/IntuitMachine/status/1972252510585847835 *