AI assisted professional skill degradation: empirical evidence and countermeasures

Introduction

With the rapid penetration of generative artificial intelligence (AI) and decision support systems in professional fields such as healthcare, software engineering, and financial analysis, a deep-seated concern is spreading within the professional community: as professionals increasingly rely on AI tools to complete core tasks, will their "hard won skills" accumulated through long-term training gradually become dull? This concern is not unfounded. Recent empirical studies have revealed that AI driven skill degradation is actually occurring. This article takes the colonoscopy study of Polish endoscopists as the core case, and combines relevant evidence from fields such as computer science to analyze the mechanism of AI induced skill degradation from the perspectives of cognitive science and career development. It also explores a strategic framework for establishing a balance between efficiency and professional competence.

Empirical evidence of AI skill degradation

1. Medical field: Decreased detection rate of adenomas in colonoscopy examination

A study on endoscopists in Poland provides clear quantitative evidence of skill degradation. The research subjects are all experienced doctors who have completed at least 2000 colonoscopy examinations in their careers. The research team has introduced an AI assisted diagnostic system that can analyze colonoscopy images in real-time and label precancerous lesions - adenomas. In the three months before the introduction of the AI system, the detection rate of adenomas by doctors without assistance was 28.4% (approximately 28 adenomas were found in every 100 examinations). In the three months after the AI system was put into use, when the system was unavailable due to maintenance or upgrades, the adenoma detection rate of these doctors significantly decreased to 22.4%, a decrease of 21.1 percentage points. This data has a dual significance: on the one hand, the detection rate of doctors during AI assistance may be improved (although the data during AI assistance was not directly provided in the study, it can be reasonably inferred), and on the other hand, once AI support is removed, human diagnostic ability shows a statistically significant decline. This degradation is not due to memory decay - a three-month interval is not enough to cause the highly programmed skill of endoscopic diagnosis to naturally decline, but more likely to stem from the establishment of cognitive dependence.

2. In the field of computer science, the ability to review code and locate problems has declined

In the field of software engineering, the widespread use of large language models such as GitHub Copilot and ChatGPT has also raised similar concerns. A preprint study conducted by Stanford University in 2023 showed that developers who used AI assisted programming tools for six months had an error rate about 18% higher than the control group who did not use AI when completing code review tasks in an AI free environment. Specifically, it manifests as a decrease in the ability to identify boundary condition errors and a decrease in the accuracy of judging code logical consistency. In addition, when there are hidden defects in the content generated by AI, developers who rely on AI often discover the root cause of the problem later than independent programmers. These phenomena are highly consistent with 'automation bias' - the tendency of humans to overly trust the output of automated systems.

3. The cognitive mechanism of skill degradation

The research in psychology and cognitive science provides mechanistic explanations for these phenomena. The core mechanism can be summarized into three levels: **Cognitive Offloading: When AI undertakes low-level cognitive tasks such as pattern recognition and feature extraction, the human brain reduces the activation frequency of related neural pathways. This' outsourcing 'improves efficiency in the short term, but weakens the neural plasticity associated with these skills in the long run. For example, endoscopists used to need to actively scan the subtle color changes and raised shapes of mucous membranes, which involved active allocation of visual attention and retrieval of associative memory; The real-time tagging of AI simplifies it into passive confirmation of prompts, leading to a decline in active search capabilities. **Automation Bias: Studies have shown that humans tend to believe in the information provided by automation systems, even if it is clearly incorrect. Doctors in the Polish study may still unconsciously expect system assistance even when AI is unavailable, leading to decreased alertness and an "anchoring effect" - a tendency to assume that AI will not make mistakes, thereby reducing the frequency of self judgment calls. **Fine tuning disruption: The formation of professional skills relies on "deliberate practice" - repeatedly exposing errors and adjusting strategies in challenging tasks. AI tools often block this feedback loop by providing a 'perfect answer'. For example, when AI always correctly labels adenomas, doctors lose the opportunity to reflect on "near missed" cases, thereby interrupting the critical learning process required for skill climbing.

How to preserve human expertise?

Faced with the real risk of AI driven skill degradation, high-risk fields such as medicine, aerospace, and nuclear energy have begun to explore systematic response strategies. The following framework is available for reference:

1. Hybrid Training Mode: Forced Departure from AI's' Cognitive Recovery Period '

Some hospitals have started implementing the "AI intermittent use system": after using AI assisted systems for a certain number of examinations, doctors must arrange a certain proportion of AI free independent operations and undergo independent assessments. This approach is similar to pilots still needing to undergo regular manual flight training after the popularization of autonomous driving technology. The data from a study in Poland shows that if doctors work in an AI free environment for at least a specific percentage of the week, their detection rate deteriorates significantly lower than the group that continuously uses AI. Specifically, doctors who independently perform no less than 5 colonoscopes per week have a degradation rate only one-third that of those who continuously use AI.

2. Cognitive warning system: embedded monitoring indicators for skill decline

Skill degradation often occurs silently. Institutions should establish a dynamic capability assessment mechanism, such as regularly inserting "covert tests" into AI assisted processes - intentionally not marking certain normal images by the system to assess doctors' proactive detection capabilities; Or set up "human-machine divergence analysis" - automatically record and review when the doctor's judgment is inconsistent with AI, instead of relying solely on AI's correctness as the standard. For individuals who are highly dependent on AI for a long time, a trigger threshold for skill retraining should be set.

3. Layered teaching method: restricting the use of AI in the initial training stage

For beginners (medical students, junior programmers), excessive reliance on AI may lead to permanent loss of core competencies. Reference can be made to the recommendation of the Royal Society of Radiology in the UK: the use of AI assisted tools is prohibited before completing basic skill certification (usually requiring independent completion of over 500 image interpretations). AI is only allowed to be used as a 'second diagnosis and treatment opinion' in advanced case analysis, rather than as a substitute for diagnostic decision-making. This strategy has been piloted in some teaching hospitals, and the results show that in professional tests after training, the restricted AI group performs better than the free use group.

4. Technical design improvement: AI should 'assist decision-making' rather than 'replace decision-making'

Currently, most AI systems tend to provide the 'final answer', which maximizes the cognitive offloading effect. A better design direction should be "gradual disclosure" - for example, a colonoscopy AI system can be designed to first display the probability score of suspicious areas (rather than directly marking them), forcing users to actively evaluate; After the user completes the evaluation, the system displays the marked results and records the differences between the two. This' human-machine collaborative feedback loop 'maintains both efficiency and active cognitive participation of humans. Google DeepMind has attempted this type of "explainable AI" framework in medical imaging diagnosis, and initial user research has shown that it can help delay skill degradation.

Conclusion

The degradation of professional skills caused by AI is not an exaggeration. The study of Polish endoscopists provides statistically significant empirical evidence, and data from other fields such as software engineering also point to similar trends. The core mechanism lies in cognitive unloading, automated bias, and the loss of deliberate practice opportunities. However, this does not mean that humans should reject AI tools. In most cases, the efficiency improvement and error reduction brought by AI assistance far exceed the losses caused by skill degradation. The key to the problem lies in how to design systems and technologies that enable humans to enjoy the benefits of AI while maintaining the "resilience" of their professional abilities - that is, the ability to quickly recover and perform tasks when AI fails. This requires the professional community to establish clear guidelines for the use of AI: restrictions on beginners and intermittent independent operation requirements for experts; Technology developers should shift towards an interactive mode of "assisted decision-making" rather than "alternative decision-making"; Industry regulators should incorporate skill maintenance indicators into the evaluation criteria for AI deployment. The ideal relationship between AI and human expertise should not be a zero sum game, but a dynamic symbiosis: AI bears the cognitive burden of standardization and repetition, while humans focus on exception handling, ethical judgment, and innovative breakthroughs. Only in this way can we find a sustainable balance between "efficiency" and "ability".