Some things never change: how far generative AI can really change software engineering practice

Initially, we selected a sample of 30 well-established IT companies known for their extensive industry reach. Selected companies either had global operations or had well-established operations in Brazil. These companies all have at least 200 employees, with the majority exceeding 10,000 collaborators and boasting an average existence of 20 years. We contacted professionals from these companies because we focused on recruiting participants with significant experience in SE. Our target audience included individuals in roles such as Chief Technology Officers (CTOs), Technical Leaders, Software Architects, Developers, and System Analysts. By targeting these professionals, we ensured data collection from individuals directly encountering the potential ramifications of GenAI integration within their work. Most of the research participants (53%) are between 36 and 45 years old. Regarding experience working in software engineering, 47% of the participants have between 11 and 20 years of experience. Notably, 30% of the participants have over 20 years of experience. This distribution shows a good level of maturity and work experience. Decades of experience indicate that some participants have already experienced other periods of potentially significant changes in the field and their consequences. This scenario is representative of the software engineer population and provides a range of perspectives about GenAI practices and impressions.

We developed a structured online questionnaire hosted on the AirTable⁵⁵5https://airtable.com platform to gather quantitative and qualitative data. The questionnaire ensured anonymity for participants and has three sections. The first section collected demographic information, including job titles, years of work experience, and age ranges. The second section assesses quantitative data collection of participants’ familiarity with GenAI, its application within SE activities, and the specific GenAI tools they have tried or are using. The third section explored the perceived correlations between GenAI and SE. We referenced a series of processes in which respondents could identify which had the highest potential for GenAI integration based on their experience. We created four open-ended questions to address our objective of gathering qualitative data. We asked about their perceptions of the impacts, challenges, and ethical considerations of utilizing GenAI within SE. However, our central question was focused on aspects of SE processes that are unlikely to change significantly due to GenAI adoption.

We included an informed consent statement in the questionnaire to guarantee the participants’ reliability. This statement outlined the research objectives, intended data utilization, researcher details and contact information, and data privacy assurances. The statement emphasized that collected data would be used solely for aggregated analysis and would not be disclosed individually within research outputs. Furthermore, adherence to the General Data Protection Regulation (GDPR) for data security and privacy was explicitly stated. Also, according to article 26 of CNS Resolution No. 674 of May 6, 2022 presented by the Brazilian National Research Ethics Commission, research of this nature does not require ethical assessment by the CONEP System.

We conducted a pilot survey with five individuals to gather feedback on the instrument’s design and content. This process was necessary to identify areas for improvement and ensure the quality of the data. The objective was to assess the following aspects: (a) comprehension: are the questions clear and easy to understand?; (b) bias: are there questions that could lead to biased responses?; (c) objectivity: are the questions phrased objectively? (d) cognitive load: is the questionnaire too long?; (e) language: does the language used have a pattern? The participants were asked to complete the questionnaire and provide feedback on their experience by sending text messages to the researchers, allowing immediate and direct communication. Therefore, the pilot application produced the following key findings: (a) the participants suggested that the questionnaire adopt a consistent approach, focusing on the participant’s perspective; (b) there was an open-ended question regarding the benefits, however, participants indicated that we should approach GenAI’s impacts to allow a more comprehensive understanding of both the positive and negative perspectives; (c) the participants expressed concerns about the potential of their responses to be linked to their employers. Based on these findings, we revised the questionnaire. We removed the mandatory question about the participant’s company, considering the sensitivity of GenAI in the workplace. As suggested, we also rephrased some questions to enhance clarity and conciseness and changed the question about benefits to impacts.

Following the validation and refinement of the questionnaire, we initiated the data collection process by disseminating the survey to our target population. We opted to contact several potential respondents who matched our intended audience directly via online channels. Therefore, we avoided broad and unrestricted dissemination on social media, messaging apps, and email lists to ensure a more representative sample. We sent a brief introductory text with instructions about the questionnaire and the purpose of the data collection, emphasizing our commitment to data security and privacy. The text included a link to the online questionnaire, which was designed to be self-administered and asynchronous. The survey was available for nine days.

This category highly references increased productivity, reduced manual/repetitive tasks, and enhanced process agility. A positive perception emerges regarding the utilization of GenAI as a tool to delegate processes currently performed by humans, yet not requiring higher-order cognitive skills. In this context, AI is perceived as a workforce capable of supporting project productivity and agility. Accordingly, some participants mentioned time optimization due to the broader adoption of GenAI. One mentioned that ”there is potential yet to be explored, but it can strongly speed up the entire life cycle of software”. Regarding potential time optimization approaches, respondents mentioned aspects of coding speed and activity automation. The potential for rework reduction was highlighted less frequently, yet still relevant. This impression shows the necessity of freeing humans to focus on more strategic tasks that require critical thinking, creativity, and complex problem-solving skills.

Numerous routine activities in systems analysis and development were cited as strongly impacted by GenAI. The most significant number of mentions was in writing tests, indicating that GenAI can help generate unit and integration tests and even identify other scenarios for improved test coverage. Activities such as generating documentation and basic code were also pointed out. GenAI can assist in creating API documentation, insert code comments, and create user manuals easily. An interesting activity indicated was assistance with proof of concept (POC). GenAI can rapidly create prototypes based on initial requirements and automate basic functionality for POC demonstration. A participant expressed that his team ”have carried out some proofs of concept and usage validations and it has been very beneficial”. Other tasks briefly mentioned included software maintenance, migrations acceleration, data analysis support, requirements writing, and basic design activities.

Issues were raised such as the indication of improvements in the various stages of software construction processes, especially in code review, specifications review, and fault identification. GenAI can analyze potential coding errors, style inconsistencies, and security vulnerabilities. It can also suggest best practices and provide code refactoring recommendations. A participant mentioned that they already have been using it for ”code reviews and suggestions for improving commits”. Upon analyzing the various GenAI tools that emerge daily, this has been an approach. Numerous Integrated Development Environments (IDEs) already incorporate GenAI features, such as JetBrains IntelliJ and Microsoft Visual Studio Code using Copilot. Regarding specifications, it can offer suggestions for more clarity, completeness, and consistency, potentially identifying ambiguities.

Considering the tendency towards an increasingly collaborative process between humans and AI, general actions that assist humans were mentioned, such as aiding in human learning, providing interactive tutorials and code examples tailored to their specific needs and skill levels. In this context, a participant mentioned that ”as the barrier to writing code will be lower, programmers will specialize more in other areas of computer science than in specific languages/platforms.”. Furthermore, solving simple problems and supporting the understanding of AI capabilities and limitations, will promote knowledge transfer between humans and AI. It was also mentioned that this leads to the emergence of new roles for humans and AI. Some roles can include specialists focusing on training and fine-tuning models for specific domains, AI-human interaction designers (Xu et al., 2023), prompt engineers (White et al., 2023), AI content reviewers (Leung et al., 2023), and AI security specialists (Gupta et al., 2023).

Although we asked about impacts, not influencing negative or positive positions, we could verify that most responses implied more positive and beneficial impacts. This perspective can lead to understanding that the practitioners are open-minded to this scenario.

In this respect, the most frequently mentioned aspect was the need for human training to expand knowledge and understanding of GenAI, including model capabilities, limitations, and ethical implications. Furthermore, the necessity for human acceptance was mentioned, given that there is already a contest regarding the uses of AI and the outcomes derived from it. It is a significant challenge since the fear of job displacement and lack of trust in AI outputs have been widely discussed over the last few years. In this sense, the adaptation of developers was also highlighted, indicating the need to adjust to paradigm shifts from the increased integration of AI in SE. It will be necessary to develop new skills and adapt to new workflows. Another mention was corporate culture. Organizations that are open to innovation could be more receptive to GenAI. An interesting element cited was the potential tendency towards complacency. Other technological resources have impacted history in different ways. GenAI can lead to conformity if humans directly accept what is created without critical reasoning. Therefore, it is necessary to emphasize the importance of lifelong learning and human development enhanced by these resources and not overlapped by them.

The most cited issue was information security. Concerns were expressed about the potential for sensitive data breaches, the generation of code containing confidential information, and the risk of manipulating AI models to generate malicious processes. Furthermore, the accuracy of results and the need to understand the construction of correct prompts were also mentioned. Fewer mentions were about interoperability and the potential for over engineering in cases where developers could spend excessive time and effort fine-tuning GenAI models for specific tasks. Some pointed out concerns about AI’s performance and reliability in scenarios of low transparency of how the generative model works. Integration among resources was presented as an element that needs further exploration, given that a framework of technologies, platforms, and tools already exists, and there is a need to incorporate AI in convergence with these elements. As AI becomes more adopted in corporate scenarios, the increasing possibility of personalization will be necessary. Working with the specificities of scenarios and complex problems may bring significant issues regarding constructing new models and using large language models (LLMs).

Some challenges addressing management and strategic contexts to incorporate AI were pointed out. The most mentioned issue is aligning the need for AI adoption with internal corporate contexts related to organizational structure, financial costs, and information security. However, there is also the difficulty of a lack of strategy in repositioning, causing a company’s processes not to keep up with market movements towards the adoption of AI. Other issues also relate to the code’s accountability process, as we would begin to have code produced by AI in conjunction with code produced by humans.

Although various technical issues have been highlighted, the most critical challenges perceived by the participants are centered on human aspects, especially in developing skills to make good use of AI resources. This scenario opens discussions regarding required skills and the potential change in the profile of SE professionals. Also, it is interesting to note that these challenges presented by the adoption of GenAI could impact areas not directly related to technology, such as education, law, business, psychology, and sociology.

Concerns about information security are not a new topic, but the rise of GenAI has intensified this discussion. Using sensitive information in training models can create security vulnerabilities and expose confidential data. Furthermore, the evolving threat landscape requires robust security measures to protect against malicious actions that could manipulate models to produce harmful codes or outcomes. Privacy remains essential, and anonymization techniques must be employed to safeguard user data. Another recurring problem is bias in outcomes. AI outputs can reflect social biases because models are trained on data that reflects human society and culture. About this, a participant said, “some models will be created with the sole objective of being the most effective of all, without any ethical filter [..] this can become a very powerful tool in the wrong hands”. This can lead to the perpetuation and amplification of biases, potentially resulting in discriminatory outcomes. Data transparency is also a concern, as the origin and structure of the data used to train models are usually not openly available. Moreover, GenAI can sometimes fabricate information that appears plausible but is factually incorrect. This kind of behavior has been called “hallucination” (Ebert and Louridas, 2023) and can mislead people to believe in inaccurate information.

In this regard, the most mentioned concern was Intellectual Property. The ownership of AI-generated content is a complex topic. However, it is necessary to find ways to protect proprietary AI models and the innovations they represent. Patent laws may need to adapt to address AI-specific intellectual property concerns. Another consideration pointed out was human responsibility, indicating the need to be accountable for the decisions or any harm caused by AI-generated outputs. Liability must also exist in cases where AI models are misused or intentionally manipulated to produce harmful results. Some studies (Farina et al., 2024) about ethical guidelines establish principles for developing and using GenAI, essential to ensure responsible and ethical practices. The importance of compliance was also cited. For some participants, GenAI systems must comply with existing laws and regulations,“the appropriate action is to adhere to the norms established by civil and criminal codes, proceeding with the investigation and punishment of individuals who use these tools to gain illicit advantages”. However, in some cases, it will also be necessary to adapt existing frameworks or develop new ones to supervise the development and use of GenAI technologies effectively. Another topic is the indication of human supervision throughout the AI development process, from data selection and model training to deployment and monitoring. Human judgment and intervention should also remain essential in critical decisions or potential harm.

The most mentioned topic was the potential social implications of the widespread use of GenAI. Job displacement was cited, as automation could lead to job displacement in various sectors. One participant thinks this context could “require the planning of public and union policies to prevent the precarization of work and for those who remain after the mass adoption of technology”. This could reshape the nature of work, and it will be essential to consider how humans and AI can best collaborate to achieve optimal outcomes. An interesting topic pointed out by the participants was the tendency towards complacency. The potential over dependence on AI for tasks that could be done with human creativity and critical thinking could lead to a decline in these abilities, encouraging passive consumption rather than active engagement and critical information analysis.

Requirements engineering plays a critical role in the successful development and deployment. The most mentioned topic in this category was functional and non-functional requirements gathering. Capturing this information requires a serious understanding of the domain, project goals, and user needs. Also, requirements should be consistent, feasible, and verifiable, and the need for requirements disambiguation was pointed out. This activity often involves dealing with ambiguous or uncertain statements that must be clarified and translated into clear and actionable requirements. Emphasizing user-centered design is also important, ensuring that the software meets the needs and expectations of its intended users, especially when a project is in the discovery phase.

Other topics included creating or following a value proposition definition for a project or product and prioritizing features and tasks. One of the participants corroborates this idea, expressing that it will be difficult to entirely change “the core development of the product itself, such as the value proposition and requirements, things that require more creative input”. These require understanding the organization’s business goals, target market, and competitive landscape, and contextual understanding and strategic insight are difficult processes for AI to perform.

The participants demonstrate the belief that GenAI will not be flexible enough to handle dynamic situations or resolve inconsistencies in requirements. Technical constraints, resource limitations, and project timelines must be considered, and it may also not be able to capture subtleties of user behavior, preferences, and pain points.

The need for experts with deep domain knowledge is especially true in complex or specialized fields where experience, critical thinking, and intuition are important. These procedures go beyond simply processing and generating data. Therefore, we must make decisions based on multiple criteria besides our expertise and experience, adapt to new circumstances and challenges, and apply knowledge and skills in various situations. This idea appears when a participant says: “the way we operate today will be assisted by GenAI, especially in repetitive tasks [..], but the main decisions will continue to be performed by humans”.

Another topic was human creativity and the capacity to break out conventional thinking patterns and develop original ideas. The nature of human creativity is often personal and reflects the individual’s experiences. That is why it is necessary for empathy to help us understand and connect with other people’s emotions.

The relationship with clients and customers was mentioned since, in the SE field, many processes will still require building a relationship, understanding customer needs, and providing empathetic support. Besides, dealing with complex customer issues, negotiating solutions, and resolving conflicts often require human judgment and emotional intelligence. Also, it was pointed out that collaborative human processes involve effective communication, teamwork, building consensus among different parties, and trust and relationships among team members.

The subjective nature of some processes was also a topic. A participant mentioned that “sometimes AI-generated ideas are too generalist, and even when explaining and trying to delve into a specific subject with the tool, it’s not quite the same as a person writing and explaining the ideas.” Subjective processes often involve understanding human values, beliefs, and cultural norms, making ethical decisions, and responding to human emotions.

The participants highlighted the importance of some quality assurance and validation processes in software development. Some believe that, even with GenAI’s capabilities, it will not replace human roles, especially in acceptance testing, which involves evaluating software against real-world scenarios and user behavior, which could involve subjective evaluation of usability, aesthetics, and user experience.

Another situation mentioned is that even with the high adoption of GenAI in the Quality Assurance field, well-structured prompts will be needed to improve the quality of AI-generated outputs. Specialists must still craft effective prompts that accurately capture the testing requirements and context to evaluate the generated outputs’ correctness, completeness, and consistency. A participant said, “even with many possible automations, the human still needs to assess whether the outcomes are as expected, and [..] evaluate whether the generated code meets the requirements and is functional.”In this sense, human review and expert validation will be important elements to ensure compliance with relevant standards, regulations, certifications, and specific business requirements.

The most mentioned questions in these processes were task definition and scope management. According to the participants, the need for project managers who can break down complex projects into manageable tasks will be preserved. Understanding project goals, defining clear scope and technical requirements, and ensuring alignment among stakeholders are increasingly important.

It was mentioned that the broad adoption of agile management methods often brings unexpected challenges and quick adaptation. Another aspect of this context is the need to manage agile teams that require effective communication, conflict management, and consensus building among diverse perspectives. This scenario is already challenging to people, and we assume that it is improbable that GenAI will replace humans in this management role.

Also, assigning roles and responsibilities in projects requires considering people’s skills, experience, and availability. It is challenging to assess these individual capabilities, delegate effectively, and ensure team members are accountable for their work.

Participants’ responses evidenced some specific aspects of software implementation and maintenance. The development of complex, distributed, or multi-entity systems presents non-standard situations. This intricacy demands a deep understanding of complex interactions, mixed data flows, dependencies, and potential failure scenarios to anticipate possible issues and create strong solutions. These tasks are more likely to continue as a human responsibility.

Another frequently cited aspect was the definition of the software architecture. Participants highlighted that defining a satisfactory architecture requires a comprehensive understanding of system requirements, performance, and scalability.

GenAI might suggest code blocks and create simple solutions. However, responses revealed other activities that still prioritize experience and human judgment, such as design construction and code customization for more complex systems, especially to preserve a well-structured codebase.

Configuration and change management also emerged as areas where human expertise is important. Effectively managing configuration files and dependencies across complex systems demands a deep understanding of system interactions and the ability to anticipate potential conflicts and risks during system updates. Finally, participants mentioned system integration challenges due to the necessity of aggregating different software components and knowledge of specific APIs, data formats, and communication protocols.