The use of artificial intelligence (AI) explodes through many branches of science. Between 2012 and 2022, the average percentage of scientific papers that interact with artificial intelligence, through 20 fields, quadruple (see “the rise of artificial intelligence in research”), including economics, geology, political science and psychology1.
Hopes are high that artificial intelligence can It appears to be slowing down: Although there is more funding, publications and employees, we achieve progress at a slower pace.
But the rush to adopt artificial intelligence has consequences. With the spread of its use – in predicting diseases, predicting the results of people’s lives and anticipating civil wars – there is a justification of a degree of caution and meditation. While statistical methods in general have the risk of using them incorrectly, artificial intelligence carries greater risks due to its complexity and black nature. The errors have become increasingly common, especially when using ready -made tools by researchers with limited experience in computer science. It is easy for researchers to exaggerate the predictive capabilities of the artificial intelligence model, and thus create an illusion of progress with the interruption of real developments.
Here, we discuss the risks and suggest a set of treatments. Create clear scientific guidance on how to use these tools and technologies sooner.
There are many ways in which artificial intelligence can be spread in science. It can be used to comb effectively through previous work, or to search for a problem space (, for example, drug candidates) to obtain a solution that can then be verified through traditional means.
Another use of artificial intelligence is to build a model of machine learning for a phenomenon of attention, and interrogating it to acquire knowledge around the world. The researchers call this science based on the machine learning. It can be considered an upgrade of traditional statistical modeling. Modeling machinery is a saw for the saw to the manual ax of statistics-stronger and more automatically, but it is dangerous if it is incorrectly used.
Source: Reference. 1
Our interest is mainly about these modified methods, as artificial intelligence is used to make predictions or test hypotheses on how the system works. One of the common sources of error is a “leakage”, a problem that arises when information from the form evaluation data incorrectly affects the training process. When this happens, the machine learning model may save patterns in evaluation data rather than pick up the meaningful patterns behind the phenomenon of attention. This limits the ability to apply in the real world of these models and does not produce little in terms of scientific knowledge.
Through our investigations and assembly of current evidence, we found that the leaves across at least 30 scientific fields – starting from psychiatry and molecular biology to computer safety – that use machine learning suffers from leakage2 (He sees go.nature.com/4ieawbk). It is a kind of “teaching for testing”, or worse than that, give answers away before the exam.
For example, during the Covid-19s, hundreds of studies have mentioned that artificial intelligence can diagnose the disease using only X-rays on the chest or CT scans. A systematic review of 415 such studies found that only 62 meets the basic quality standards3. Even among them, the defects were widespread, including bad evaluation methods, repeated data and lack of clarity about whether “positive” cases of people had a certain medical diagnosis.
In more than ten studies, the researchers used a training data set in which all positive cases were in adults, and the negatives were between children between one and five. As a result, the artificial intelligence model has learned to distinguish between adults and children, but the researchers mistakenly concluded that they developed a Covid-19 detector.
It is difficult to systematically look like these errors because assessment of predictive accuracy is very difficult and so far it lacks standardized. Computer symbol rules can be thousands of long lines. It may be difficult to discover errors, and it can be one costly error. Thus, we believe that the cloning crisis in science -based science is still in its early days.
With some studies that now use large language language models – for example, using it as an alternative to human participants in psychological experiences – there are more ways that research may be unacceptable. These models are sensitive to inputs. Small changes in the formulation of claims can cause remarkable changes in the outputs. Because the models are often owned and managed by private companies, access can be restricted at any time, making the repetition of these studies.
Deceive
A greater risk of hasty adoption in artificial intelligence and machine learning lies in the fact that the flow of results, even if free of errors, may not lead to real scientific progress.
To understand this danger, consider the influence of a greatly influential paper from 2001, where the statistical Leo Breman described brilliantly cultural and methodological differences between the fields of statistics and machine learning.4.
Will Amnesty International strengthen scientific photography? There is still time to create an ethical code
He strongly called on the latter, including the adoption of automated learning models on simpler statistics, with a priority prediction accuracy about issues about the sincerity of the model that represents nature. In our opinion, this invitation did not mention the well -known restrictions of the machine learning approach. The paper does not distinguish enough to distinguish between using machine learning models in engineering and natural sciences. Although Breiman found that black box models can work well in engineering, such as classification of submarines using sonar data, they are difficult to use in natural sciences, where nature explains (for example, the principles behind the spread of sound waves in water) are the basic point.
This confusion is still widespread, and many researchers are lured due to Amnesty International’s commercial success. But just because the modeling approach is useful for engineering, this does not mean that it is good for science.
There is the old maximum that “every model is wrong, but some models are useful.” It takes a lot of work to translate outputs from models to claims around the world. The machine learning tool box makes it easy to create models, but it does not necessarily make it easier to extract knowledge around the world, and it may make it more difficult. As a result, we face the risk of producing more, but the understanding is less5.
Science is not just a set of facts or results. Actual scientific progress occurs through theories, which explain a set of results and models, which are conceptual tools to understand and investigate the field. While we move from the results to theories to models, things become more abstract, wider and less automated. We doubt that the rapid spread of artificial intelligence -based scientific results – and may have prevented – these levels are higher than progress.
If the researchers are in a field of concern about faults in individual papers, we can measure their spread by analyzing a sample of the leaves. But it is difficult to find evidence of the defender of smoking that scientific societies as a whole exaggerate the emphasis on predictive accuracy at the expense of understanding, because the antibiotic world cannot be reached. However, historically, there were many examples of fields that stumbled into a rut even when they were distinguished in the production of individual results. Among them is alchemy before chemistry, astronomy before the Copernican revolution and geology before a tectonic plate.
A shout for artificial intelligence studies does not issue headlines
Astronomy story is especially related to the prosecution. The model of the universe with the Earth in its center was very accurate in predicting planetary suggestions, due to tricks such as “bikes” – the assumption that planets are moving in circles whose centers revolve around the Earth along a larger circular path. In fact, many modern celestial dome display devices use this way to calculate the tracks.
Today, artificial intelligence excels in the production of bicycles. All that is equal, the ability to press more predictive juice from defective theories and insufficient materials will help them circumvent true scientific progress.
The paths forward
We have pointed out two main problems in the use of artificial intelligence in science: defects in individual studies and cognitive issues with the wide adoption of Amnesty International.