AI vs AI: Scientists Develop Neural Networks to Detect Generated Text Insertions
A research team, including Alexander Shirnin from HSE University, has developed two models designed to detect AI-generated insertions in scientific texts. The AIpom system integrates two types of models: a decoder and an encoder. The Papilusion system is designed to detect modifications through synonyms and summarisation by neural networks, using one type of models: encoders. In the future, these models will assist in verifying the originality and credibility of scientific publications. Articles describing the Papilusion and AIpom systems have been published in the ACL Anthology Digital Archive.
As language models like ChatGPT and GigaChat become more popular and widely used, it becomes increasingly challenging to distinguish original human-written text from AI-generated content. Artificial intelligence is already being used to write scientific publications and graduation papers. Therefore, it is crucial to develop tools capable of identifying AI-generated insertions in texts. A research team, including scientists from HSE University, presented their solutions at the SemEval 2024 and DAGPap24 international scientific competitions.
The AIpom model was used to identify the boundaries between original and generated fragments in scientific papers. In each paper, the proportion of machine-generated text to the author's text varied. To train the models, the organisers provided texts on the same topic. However, during the verification stage, the topics changed, making the task more challenging.
Alexander Shirnin
'Models perform well on familiar topics, but their performance declines when presented with new topics,' according to Alexander Shirnin, co-author of the paper and Research Assistant at the Laboratory for Models and Methods of Computational Pragmatics, HSE Faculty of Computer Science. 'It's like a student who, having learned how to solve one type of problem, struggles to solve a problem on an unfamiliar topic or from a different subject as easily or accurately.'
To improve the system's performance, the researchers combined two models: a decoder and an encoder. At the first stage, a neural network decoder was used, with the input consisting of an instruction and the source text, and the output being a text fragment presumably generated by AI. Next, in the original text, the area where the model predicted the beginning of a generated fragment was highlighted using a special <BREAK> token. The encoder then processed the text marked up in the first stage and refined the decoder's predictions. To do this, it categorised each token—the smallest unit of text, such as a word or part of a word—and identified whether it was written by a human or generated by AI. This approach improved accuracy compared to systems that used only one type of model: AIpom ranked second at the SemEval-2024 competition.
The Papilusion model also distinguished between written text and generated text. Using Papilusion, sections of the text were classified into four categories: written by a human, modified with synonyms, generated, or summarised by a model. The task was to accurately identify each category. The number of categories and the length of insertions in the texts varied.
In this case, the developers used three models, all of the same type: encoders. They were trained to predict one of the four categories for each token in the text, with each model trained independently of the others. When a model made an error, a cost was applied, and the model was retrained with the lower layers frozen.
'Each model has a different number of layers, depending on its architecture. When training a model, we can leave the first ten or so layers unchanged and adjust only the parameters in the last two layers. This is done to prevent losing important data embedded in the first layers during training,' explains Alexander Shirnin. 'It can be compared to an athlete who makes an error in the movement of their hand. We only need to explain this part to them, rather than resetting their entire learning and retraining them, as they might forget how to move correctly overall. The same logic applies here. The method is not universal and may not work with all models, but in our case, it was effective.'
The three encoders independently determined the category for each token (word). The system's final prediction was based on the category that received the most points. Papilusion ranked sixth out of 30 in the competition.
According to the researchers, current AI detection models perform reasonably well but still have limitations. Primarily, they struggle to process data beyond what they were trained on, and overall, there is a lack of diverse data to train the models effectively.
'To obtain more data, we need to focus on collecting it. Both companies and laboratories have been doing this. Specifically for this type of task, it is necessary to collect datasets that include texts modified using multiple AI models and modification methods,' the researcher comments. 'Instead of continuing a text using just one model, more realistic scenarios should be created, such as asking the model to add to the text, rewrite the beginning for better coherence, remove parts of it, or generate a portion of the text in a new style using a different prompt. Of course, it is also important to collect data in different languages and on a variety of topics.'
See also:
Scientists Discover Why Parents May Favour One Child Over Another
An international team that included Prof. Marina Butovskaya from HSE University studied how willing parents are to care for a child depending on the child’s resemblance to them. The researchers found that similarity to the mother or father affects the level of care provided by parents and grandparents differently. Moreover, this relationship varies across Russia, Brazil, and the United States, reflecting deep cultural differences in family structures in these countries. The study's findings have been published in Social Evolution & History.
When a Virus Steps on a Mine: Ancient Mechanism of Infected Cell Self-Destruction Discovered
When a virus enters a cell, it disrupts the cell’s normal functions. It was previously believed that the cell's protective response to the virus triggered cellular self-destruction. However, a study involving bioinformatics researchers at HSE University has revealed a different mechanism: the cell does not react to the virus itself but to its own transcripts, which become abnormally long. The study has been published in Nature.
Researchers Identify Link between Bilingualism and Cognitive Efficiency
An international team of researchers, including scholars from HSE University, has discovered that knowledge of a foreign language can improve memory performance and increase automaticity when solving complex tasks. The higher a person’s language proficiency, the stronger the effect. The results have been published in the journal Brain and Cognition.
Artificial Intelligence Transforms Employment in Russian Companies
Russian enterprises rank among the world’s top ten leaders in AI adoption. In 2023, nearly one-third of domestic companies reported using artificial intelligence. According to a new study by Larisa Smirnykh, Professor at the HSE Faculty of Economic Sciences, the impact of digitalisation on employment is uneven: while the introduction of AI in small and large enterprises led to a reduction in the number of employees, in medium-sized companies, on the contrary, it contributed to job growth. The article has been published in Voprosy Ekonomiki.
Lost Signal: How Solar Activity Silenced Earth's Radiation
Researchers from HSE University and the Space Research Institute of the Russian Academy of Sciences analysed seven years of data from the ERG (Arase) satellite and, for the first time, provided a detailed description of a new type of radio emission from near-Earth space—the hectometric continuum, first discovered in 2017. The researchers found that this radiation appears a few hours after sunset and disappears one to three hours after sunrise. It was most frequently observed during the summer months and less often in spring and autumn. However, by mid-2022, when the Sun entered a phase of increased activity, the radiation had completely vanished—though the scientists believe the signal may reappear in the future. The study has been published in the Journal of Geophysical Research: Space Physics.
Banking Crises Drive Biodiversity Loss
Economists from HSE University, MGIMO University, and Bocconi University have found that financial crises have a significant negative impact on biodiversity and the environment. This relationship appears to be bi-directional: as global biodiversity declines, the likelihood of new crises increases. The study examines the status of populations encompassing thousands of species worldwide over the past 50 years. The article has been published in Economics Letters, an international journal.
Scientists Discover That the Brain Responds to Others’ Actions as if They Were Its Own
When we watch someone move their finger, our brain doesn’t remain passive. Research conducted by scientists from HSE University and Lausanne University Hospital shows that observing movement activates the motor cortex as if we were performing the action ourselves—while simultaneously ‘silencing’ unnecessary muscles. The findings were published in Scientific Reports.
Russian Scientists Investigate Age-Related Differences in Brain Damage Volume Following Childhood Stroke
A team of Russian scientists and clinicians, including Sofya Kulikova from HSE University in Perm, compared the extent and characteristics of brain damage in children who experienced a stroke either within the first four weeks of life or before the age of two. The researchers found that the younger the child, the more extensive the brain damage—particularly in the frontal and parietal lobes, which are responsible for movement, language, and thinking. The study, published in Neuroscience and Behavioral Physiology, provides insights into how age can influence the nature and extent of brain lesions and lays the groundwork for developing personalised rehabilitation programmes for children who experience a stroke early in life.
Scientists Test Asymmetry Between Matter and Antimatter
An international team, including scientists from HSE University, has collected and analysed data from dozens of experiments on charm mixing—the process in which an unstable charm meson oscillates between its particle and antiparticle states. These oscillations were observed only four times per thousand decays, fully consistent with the predictions of the Standard Model. This indicates that no signs of new physics have yet been detected in these processes, and if unknown particles do exist, they are likely too heavy to be observed with current equipment. The paper has been published in Physical Review D.
HSE Scientists Reveal What Drives Public Trust in Science
Researchers at HSE ISSEK have analysed the level of trust in scientific knowledge in Russian society and the factors shaping attitudes and perceptions. It was found that trust in science depends more on everyday experience, social expectations, and the perceived promises of science than on objective knowledge. The article has been published in Universe of Russia.