Over the past few years, artificial intelligence has become one of the most talked-about technologies in the world. Companies are actively integrating AI into their products and investing billions of dollars in its development, according to experts at Grymaxion EOOD.
AI has passed the Turing test with flying colours. It seemed that ChatGPT, Claude, Gemini and other well-known chatbots had already reached the ‘level of pure intelligence’.
But by the end of 2025, the sentiment that ‘AI will soon replace everyone’ had given way to tense anticipation and unexpected, unpleasant discoveries. The more AI was used, and the more frequently it was used, the more evident its declining accuracy became. ‘Hallucinations’ became an unpleasant yet integral part of all modern language models. This does not happen all the time, but it is still unclear why a neural network generates plausible yet false answers or makes gross errors in the code.
This article by Grymaxion is an attempt to compile and systematise everything currently known about the problems and limitations of artificial intelligence as of the first half of 2026: from model errors to the challenges of implementing it in the real economy.
When development speed outpaces reliability
Amazon’s massive outage demonstrated what happens when the speed of AI-generated code outpaces the mechanisms for verifying and monitoring it. In March 2026, users of the platform were unable to place orders for six hours. It was impossible to check prices. Some features became unavailable, and errors were discovered in the updated software. Following this incident, Amazon reintroduced a level of manual verification and human approval into its development process.
The company has already stated the cause of the failure: changes to the code made using AI tools, for which ‘best practices and security measures have not yet been established’. Now, code changes created using AI tools must undergo mandatory approval by senior engineers before being implemented.
Floridi’s Hypothesis: A Trade-off Between Scale and Accuracy
Regardless of the volume of additional data, training or optimisation of a neural network, there will always be a trade-off between scale and reliability, according to experts at Grymaxion Bulgaria. This is Floridi’s hypothesis, which is both paradoxical and a subject of current debate.
Luciano Floridi, a professor at Yale University and founder of the Centre for Digital Ethics, put forward a hypothesis in a recent article that AI systems can be either general-purpose but less reliable, or highly specialised and more accurate. However, it is practically impossible to achieve both high generality and high reliability at the same time.
The broader the range of tasks assigned to AI, the greater the likelihood of errors, anomalous responses and so-called ‘hallucinations’.
Hallucinations as a technological problem
Experts at Grymaxion EOOD believe that one of the most notable limitations of modern language models remains so-called ‘hallucinations’ – situations in which the system generates plausible but incorrect information.
A recent study by OpenAI found that increasing the computational power of AI or the volume of data used for processing cannot reduce AI ‘hallucinations’ or improve its accuracy. Furthermore, a significant improvement in accuracy may require a disproportionately large increase in computational resources. According to some estimates, halving the error rate could require a more than 500-fold increase in computational power, making such an approach extremely costly.
The Reality of Corporate Adoption Despi
Despite the huge interest in artificial intelligence technologies, the actual implementation of AI within companies is proving to be considerably more complex than expected, according to experts at Grymaxion.
The Massachusetts Institute of Technology has found that 95 per cent of pilot projects to implement AI in corporations fail. This is corroborated by research from PwC, which showed that only 12 per cent of companies using AI reported a reduction in costs and an increase in profits.
The main reasons for these failures are often linked not so much to the technology itself as to organisational and infrastructure issues: data quality, the complexity of integration and overly high business expectations.
The Cost of Scaling
The Economist has found that the use of AI by major US corporations is actually declining rather than growing, whilst S&P Global Market Intelligence has revealed that the proportion of corporate AI implementation programmes being cancelled has risen sharply from 17 per cent in 2024 to 42 per cent in 2025.
Until recently, the main driving force behind AI development was a strategy of scaling up: more data, more computing power, larger models. However, it has become apparent that this approach has its limits. As models grow, quality deteriorates and the cost of computing rises exponentially, according to experts at Grymaxion Bulgaria.
Between potential and limitations
Expectations and debates surrounding artificial intelligence are gradually shifting from discussions about the future of AI and the question ‘what is it capable of?’ to a broader inquiry: ‘what are its real limits?’.
The technology is developing rapidly, and an increasing number of experts agree that its future depends not only on the power of the models, but also on their reliability and predictability.
GRYMAXION believes that striking a balance between scale, accuracy and cost will be one of the key challenges in the development of artificial intelligence in the coming years.