Along with the flow of capital pouring into AI, there is a long and mostly unresolved list of real obstacles to mass adoption. Among them is recursive data contamination. Large language models generate huge amounts of content, which is then used as training material for the next generation of models. Errors and hallucinations are amplified with each cycle. This is reminiscent of multiple copying of a copy: quality steadily declines, and eventually it becomes impossible to determine the original source. The industry is already turning to synthetic data to compensate for the lack of high-quality human content — but this risks accelerating degradation rather than eliminating it.

The even more serious problem is data poisoning. Malicious actors can intentionally distort the training set, and once embedded, the “poison” remains in the model forever. The military scenario is especially dangerous: an AI trained to recognize allies and enemies based on compromised data will only discover its hidden vulnerability during a real conflict. It has been documented that poisoning language models of any size requires only 250 malicious documents — making attacks on training data not a hypothetical threat but a very real cybersecurity issue.