Cognitive Debt: The true price of AI in the IT market

Image by Berke Citak (@berctk) | Unsplash

The use of Artificial Intelligence and its widespread adoption in daily life have brought numerous benefits, such as convenience, practicality, and assistance in solving both simple problems and complex tasks. However, as this article aims to demonstrate, the indiscriminate use of this tool also exacerbates certain issues, such as “cognitive debt,” the rapid restructuring of the software development market, the visible and invisible costs associated with these tools, and the risks they pose to both project security and the learning of new technologies. Relevant articles and publications from 2025 and 2026 were gathered to analyze these dangers and reveal how the global Information Technology landscape has undergone significant changes in such a short period.

Non-members: read the article here

In recent years, the use of AI (Artificial Intelligence) tools has gained popularity as a method for quickly and easily providing information while reducing the need to perform repetitive tasks. Within companies, the focus has shifted to accelerating project production and development, aiming to increase the volume of deliverables and enhance employee productivity. As these technologies become more accessible and yield increasingly effective — though not yet fully efficient — results, it has become commonplace to manage and develop work involving significant amounts of content and code written by artificial intelligence rather than a human developer.

The term “Vibe Coding” (a method of developing projects using only a description, or “prompt”) has gained popularity to describe people who participate in creating programs, websites, and applications without needing to write a single line of code. The user briefly or in detail describes what they wish to create, and the artificial intelligence tool delivers something close to that description. Traditional development methods are increasingly sharing the stage with newer approaches, a shift that is already transforming the job market and creating new related roles and positions. Job listings for “vibe coders” and/or AI tool specialists are appearing on major job search sites like LinkedIn and Indeed.

“Vibe Coder” as a job title found in a search on LinkedIn

On one hand, AI tools lower the barrier to entry, enabling individuals without deep technical expertise to work in the technology sector — creating new opportunities and job openings for those who demonstrate mastery of these tools. On the other hand, issues stemming from a lack of specialized domain knowledge are beginning to emerge. “Vibe Code Fixers” — professionals focused on correcting errors in AI-generated code — and profiles adopting this very label are easily found when searching for experts to “clean up the mess” caused by over-reliance on the market’s most popular tools. The risk associated with using tools to build IT projects without prior foundational learning has a name: Cognitive Debt.

“Vibe Code Fixer” as a job title found in a search on LinkedIn

“Cognitive debt” refers to the process of offloading complex mental tasks to artificial intelligence tools. Tasks deemed more complex — such as problem-solving, tackling logical challenges, or learning new information or technologies — require time and effort to understand and master. However, there is a growing trend of using AI as a “mental crutch,” delegating the more “tedious” aspects of a process to the tools while focusing solely on the final outcome. In their publication, Miranda (2025) argues that outsourcing intellectual tasks can impair information retention, cognitive authorship, and analytical capacity in the long run. This phenomenon is particularly pronounced in Latin American countries, where the use of AI tools is actively encouraged, including by universities. In their article, Kosmyna et al. (2025) present data confirming these findings: users with access to AI performed significantly worse than a control group without such access in a study examining the neural, linguistic, and behavioral consequences of the excessive use of AI writing assistance tools.

In a recent CNN Brasil article by Christ (2026), Borja Castelar, former director of LinkedIn Latin America, analyzes the need to preserve logical thinking in order to achieve excellent results. According to him, one of the major risks associated with AI dependency is the impression that human thought can be replaced by technology:

“[Young people] enter a market where technical knowledge quickly becomes obsolete. What sustains a career today is not a specific profession, but the ability to continuously learn, adapt, and develop human skills. A career is no longer a fixed path; it has become a constant process of reinvention”.

The search for talent specializing in the use of the latest tools is balanced against the available job market. Traditional entry-level roles are saturated with professionals, whereas the few individuals who excel in using generative or agentic AI are fiercely sought after by major organizations. Companies that successfully strike a balance in the human-machine relationship will gain a significant competitive advantage, while those that employ artificial intelligence in a superficial or low-impact manner are unlikely to see the same results (World Economic Forum, 2025).

According to a publication by the National Bureau of Economic Research (Demirer; Musolff; Yang, 2026) — which analyzed over 100,000 GitHub developers and their AI usage telemetry — an increase in “commits” (submissions of new code versions) was observed, with cumulative effects ranging from 40% to 140% and even 180%. However, the same publication notes a significant discrepancy between what is produced and what is actually delivered: the cumulative 180% figure drops to 50% in terms of the number of projects, and to 30% regarding actual project deliveries. These data raise a curious red flag: while AI is increasingly used in daily corporate operations, the quantitative and qualitative results fall short of expectations, and the financial cost has become impossible to ignore. Companies such as Uber, Microsoft, and Nvidia are limiting employee AI usage to curb spending associated with excessive consumption — a practice colloquially known as “tokenmaxxing.” This maximization of token usage (tokens being the digital currency powering most AI tools) represents a trend of measuring productivity by processing volume rather than final outcomes. Many companies adopt an “AI-First” approach, making AI a priority tool in their daily operations, yet fail to quantify the actual gains these tools provide (The Economist, 2026). This impulsive use of AI is already causing massive financial losses, with a particularly significant impact on small businesses (Taylor, 2026); notably, only 26% of companies report having a comprehensive cost analysis of their AI deployment, according to The Wall Street Journal (Broughton; Maurer; Williams, 2026).

Despite reports from companies like Meta and Microsoft — recognized among the world’s largest tech firms — outlining plans to lay off a significant percentage of their workforce, there is growing adoption of these tools, with usage 69% higher than in 2025 (Angelo & Rogelberg, 2026). In contrast to these figures, an analysis of the technical requirements needed for AI to operate at a quality and output level comparable to humans reveals that using these tools is advisable in only 23% of cases; in the remaining 77%, it would be more cost-effective to retain and support human developers. A conflict exists between encouraging employees to use generative AI to achieve desired productivity gains and the rising costs of these tools. In a Forbes article, Janakiram (2026) details how Uber, the world’s largest ride-sharing company, is currently engaged in a deep analysis and management of its artificial intelligence usage. Between January and April of this year, approximately 5,000 of the company’s engineers had already exhausted the entire 2026 budget allocated for AI — equivalent to $3.4 billion. The same article highlights a significant development: Anthropic’s May 13th announcement regarding price increases for subscriptions and credits for its Claude Code service. The analysis published by MindStudio (2026) on June 15th has already proven accurate. Claude’s monthly pricing was adjusted to correct a cost-versus-profit margin that had become unsustainable. Offering simpler models at a low cost achieved the expected outcome: broader reach, higher usage rates, massive data volumes, and continuous training to refine and create superior (and, consequently, more expensive) models. This strategy proves successful in the long run, and brings with it the reality of developers being partially or entirely dependent on such resources.

A study conducted by Anthropic (2026) analyzed how participants — mostly junior software developers — engaged with or without AI tools in two scenarios: (1) how quickly they could learn a new skill (in this case, a Python library) and (2) whether using artificial intelligence made it more difficult for them to interpret the code they had just written. The study examined skills related to debugging (fixing or diagnosing issues), interpreting, writing, and understanding the fundamental principles of the library used. The results showed that participants in the AI ​​group completed the tests faster but achieved poorer results than the group that did not use the tool. Furthermore, the greatest discrepancy occurred during debugging, suggesting that the process of understanding why code is incorrect and how to fix it is the most heavily impacted area when unrestricted interaction potentially hinders knowledge acquisition. Dávid Ondruš (2025) was already noting these risks when he wrote his article on Medium:

AI pushes quick results, but good development is about building systems that last and spotting issues others miss. Overuse turns you into someone who assembles unrelated puzzle.

While mentoring intern and junior-level developers, Ondruš noticed a heavy reliance on tools like ChatGPT, Cursor, and Copilot; whatever was generated was accepted without regard for quality or factual accuracy. Even with tasks that were likely simple for their skill level — such as creating an API to communicate with another microservice, fetching data from a database, or returning results — he observed a pattern: instead of analyzing how to build the application, their first instinct was to open ChatGPT and paste the problem to be solved. The code looked great, but error popups would soon appear. Issues included invented variables, a lack of understanding regarding what a specific log entry actually meant, or simply pasting the entire log into the tool and expecting it to handle the interpretation. This dependency is producing developers who bypass the essential process of analysis and questioning — missing out on the frustration that drives learning and the experience of breaking through to the “other side” with the ability to identify the problem.

Another highly relevant aspect when discussing development and programming is code security. A survey conducted by Cycode (2025) involving over 400 developers from the United States and Europe found that while 97% of companies use assistants (such as CoPilot), only 19% claim to know exactly how AI is utilized in their daily operations; nevertheless, one-third of these companies state that AI generates the majority of the code they use. These figures — alarming in themselves — mask one of the greatest risks associated with heavy reliance on artificial intelligence: so-called “Shadow AI.” Developers use AI without adhering to best practices regarding the tools, feeding them sensitive data and failing to conduct a rigorous analysis before the handoff. This, in turn, creates significant risks regarding corporate data protection (security blind spots), potentially leading to critical damage and financial loss. “AI-generated code is flooding codebases, and the vast majority of it has never undergone review by a human developer” (H@shtalk, 2026).

All this suggests that, in the not-too-distant future, an “information blackout” in the realm of information technology could occur: little (new) human-authored code is actually being written, and tools will continue “regurgitating” the same content ad infinitum — repeating lines of code, repeating the same errors, and allowing structural flaws to persist that could otherwise be identified by a human eye. The role of the future developer is not limited to simply writing good prompts or knowing how to write good code; it involves acting critically and managing this relationship with excellence.

References:

• Angelo, J. (2026, May 22). Microsoft reports are exposing AI’s real cost problem: Using the tech is more expensive than paying human employees. Fortune. https://fortune.com/2026/05/22/microsoft-ai-cost-problem-tokens-agents/
• Anthropic. (2026, January 29). How AI assistance impacts the formation of coding skills. https://www.anthropic.com/research/AI-assistance-coding-skills
• Broughton, K., Maurer, M., & Williams, J. (2026). The metric CFOs struggle to track: AI usage. The Wall Street Journal. https://www.wsj.com/cfo-journal/the-metric-cfos-struggle-to-track-ai-usage-3b30c10c
• Christ, G. (2026, January 30). ‘O risco é terceirizar o pensamento’, diz ex-diretor do Linkedin sobre IA. CNN Brasil. https://www.cnnbrasil.com.br/tecnologia/o-risco-e-terceirizar-o-pensamento-diz-ex-diretor-do-linkedin-sobre-ia/
• Cycode. (2025, November 10). AI is rewriting how software is built and secured. https://cycode.com/press/ai-is-rewriting-how-software-is-built-and-secured/
• Demirer, M., Musolff, L., & Yang, L. (2026, May). Writing code vs. shipping code: Productivity effects across generations of AI coding tools (Working Paper №35275). National Bureau of Economic Research. https://www.nber.org/papers/w35275
• H@shtalk. (2026, May 27). The cybersecurity roadmap for the AI era: What to learn, what to drop, and what to build before everyone else does. Medium. https://eva-georgieva.medium.com/the-cybersecurity-roadmap-for-the-ai-era-what-to-learn-what-to-drop-and-what-to-build-before-fb48127a7cd4
• Kosmyna, N., Hauptmann, E., Yuan, Y., Situ, J., Liao, X., Beresnitzky, A., Braunstein, I., & Maes, P. (2025, June 10). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
• MindStudio. (2026, June 15). AI pricing is about to shock everyone: Why the $20/month era is ending. MindStudio Blog. https://www.mindstudio.ai/blog/ai-pricing-shock-end-of-cheap-subscriptions
• Miranda, G. (2025). Dívida Cognitiva da Era da IA: uma proposta a partir do Design Thinking para a educação superior. Revista de Direito Civil Contemporâneo, 11(34). https://www.e-publicacoes.uerj.br/rdciv/article/view/94692
• Msv, J. (2026, May 17). Uber burns its 2026 AI budget in four months on Claude Code. Forbes. https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code/
• Ondruš, D. (2025, December 20). Over-reliance on AI is quietly stalling junior developers’ careers. Medium. https://medium.com/touch4it/over-reliance-on-ai-is-quietly-stalling-junior-developers-careers-de60074c6409
• Rogelberg, S. (2026, June 14). ‘The cost of compute is far beyond the costs of the employee’: Nvidia executive says right now AI is more expensive than paying human workers. Fortune. https://fortune.com/article/why-is-the-cost-of-ai-higher-than-human-workers-nvidia-executive/
• Taylor, M. (2026, June 12). AI costs are rising. Businesses are paying the price. Forbes. https://www.forbes.com/councils/forbestechcouncil/2026/06/12/ai-costs-are-rising-businesses-are-paying-the-price/
• The Economist. (2026, June 14). Companies are scrambling to curtail soaring AI costs. https://www.economist.com/business/2026/06/14/companies-are-scrambling-to-curtail-soaring-ai-costs
• World Economic Forum. (2025, October 3). AI’s new dual workforce challenge: Balancing overcapacity and talent shortages. https://www.weforum.org/stories/2025/10/ai-s-new-dual-workforce-challenge-balancing-overcapacity-and-talent-shortages/