When AI becomes a strategic threat
Over the past few years companies across different industries have heavily invested into AI. The motivation to do so are multifacetted: increasing productivity, reducing costs, automating workflows or developing entirely new products and services.
In doing so, companies face a wide range of challenges such as technical, financial, organizational, regulatory or ethical challenges. But there is one category of challenges that receives comparatively little attention, which are strategic risks.
Long before a model or framework is selected, organizations need to answer more fundamental questions such as what role should AI play within your company? Which business problems are you actually trying to solve? Which capabilities should you build by yourself and which should you source externally? What are the long-term trade-offs of your AI strategy and which risks are acceptable and which are not?
These strategic decisions define the base upon which all subsequent technical decisions are built. For instance, developing an in-house language model involves completely different technical challenges, cost structures or organizational problems than integrating an external API into an existing application. Neither approach is inherently superior. In fact, the right decision depends, at least to a large extent, on the company's strategic objectives.
Short-term thinking versus long-term strategy
One major strategic challenge that I have encountered in recent years are conflicts between short-term thinking and long-term strategy. Especially managers at the mid-level management or project managers frequently operate under extreme pressure to maximize short-term productivity, which often contradicts the strategic long-term goals and values of their company.
This is particularly true for companies in classic engineering disciplines, whose long-term values are usually based on quality, precision, rigor, reliability or robustness. And this problem does not only apply to AI, but also to other areas such as cyber security. Similar to insurance payments, security investments often appear expensive because their benefits are largely invisible until something goes wrong.
AI represents a similar challenge, especially when short-term thinking overshadows strategic long-term goals. Over the past 2-3 years numerous companies announced ambitious plans to replace large parts of their software engineering workforce with AI systems. Thousands of engineers were laid off based on the assumption that generative AI would soon perform comparable work.
For someone like me, who is building AI systems for almost 10 years and using AI copilots for software development, it was clear that these expectations were difficult to achieve. And in the last few months more and more evidence shows that many companies are rehiring software engineers because they were not able to fully replace them. Just as Excel sheets did not replace accountants, AI systems will not replace software engineers.
What is AI slop?
The term AI slop is commonly used to describe low-quality code generated by AI. A recent example illustrates this problem. The CEO of YCombinator Garry Tan just bragged in a Twitter post that his company ships 37,000 lines of code per day using agentic AI. In response a Polish developer, who goes by the username Gregorein, analyzed one of Tan's web apps and identified numerous inefficiencies and implementation issues.
For instance, the application generated 169 database requests during page load, by which approximately 6.4 MB of data were transferred. Moreover, the browser loaded 78 different JavaScript controllers that were never actually used. In addition to this, the browser downloaded multiple redundant versions of the company logo, all of which served unnecessarily large image files with at least 2 MB in size each.
Additionally, his analysis shows that Tan's website contains duplicate content, empty CSS style sheets, a text-editor on a read-only page or missing image descriptions. As Gregorein pointed out, his analysis was limited to the frontend code because he did not have access to the backend code.
Another crazy story around AI agents involved the company PocketOS, which sells software for car rental services. PocketOS used the coding pilot Cursor, which is based on Anthropic's Claude model and which wiped out its entire database in production. The company ended up in chaos with car reservations of three months and new customer sign ups being completely gone.
The bigger picture
AI slop often involves overengineered and bloated software, which is not scalable and which becomes increasingly difficult to maintain. As systems grow, new features begin to break existing functionalities and debugging becomes more expensive. When AI code goes into production without proper quality assessment, it can lead to functional failures.
Moreover, AI code often includes security vulnerabilities. In many cases, problems arise at a later point in time, which forces developers to reverse engineer the entire code that nobody understands. The underlying trade-off is straightforward: AI can generate code much faster than humans can review, understand and maintain. Higher code output therefore does not automatically translate into higher productivity. The larger the quantity of code the lower its quality.
Whether AI slop becomes a strategic threat depends on the specific industry, in which AI code is applied to. In some industries, the consequences of AI slop may remain relatively limited. For instance, if a software company provides an online learning platform based on AI slop, the worst case scenario would be a bunch of dissatisfied users and higher maintenance costs.
In classical engineering disciplines such as the automotive industry, aviation or medical technology, however, the consequences can be far more severe. These industries are built upon principles such as reliability, robustness, traceability, validation and rigorous quality assurance. AI slop directly conflicts with these engineering principles and in the worst case may even compromise human safety.
It still baffles me how easily decision makers at the management level fall for media hypes and often fail to realize that early adopters also become the first to accumulate technical debt, expose hidden weaknesses and absorb the costs of immature technologies. AI can be an exceptional copilot, but in many domains, however, it should not yet be trusted as the captain.