#36 Society and Technology

The tools we use, the ways we use them, and the habits we develop are fundamentally transforming us. A recent article in The New York Times highlighted striking findings from a university study examining this transformation. In the study, a professor divided 250 students into two groups and administered an exam. One group used traditional research methods while the other relied on artificial intelligence (AI) tools. The results revealed a clear and significant advantage for students using traditional methods.

Students who depended on AI summaries produced recommendations that were superficial, repetitive, and largely clichéd—such as “eat healthy, drink water, sleep well.” In contrast, students who conducted research using classical methods developed more detailed advice by making meaningful distinctions between physical, mental, and emotional health. These students demonstrated the ability to transform the information they gathered into their own original phrasing.

Technology companies and some users argue that AI tools enhance learning and productivity. However, the referenced study and similar academic research clearly show that excessive reliance on AI reduces cognitive performance. Dr. Melumad, who led the study, expressed concern, saying, “I am worried that young people can no longer conduct proper Google searches.” Indeed, it now seems unnecessary to retain even the foundational knowledge required to perform high-quality searches on search engines.

The Oxford Dictionary’s selection of the term “brain rot” as its 2024 Word of the Year reflects this alarming trend. Originally used to describe how low-quality content dulls mental faculties—particularly in the context of TikTok and Instagram videos—this term has acquired a new meaning in the age of AI: the degradation of learning processes.

Similar to the previous study, another study conducted at MIT focused on students’ writing habits. In this study, 54 university students were asked to write an essay. One group was allowed to use ChatGPT, another group relied on Google searches, and a third group wrote solely from their own knowledge. Throughout the process, students’ brain activity was measured using sensors.

It was found that those using ChatGPT exhibited the lowest levels of brain activity. More importantly, just one minute after completing their essays, 83 percent of ChatGPT users could not recall even a single sentence from their own writing. Students who used Google searches remembered some sentences, while those who wrote independently were able to accurately reproduce most of their text.

Nataliya Kosmyna from MIT poses this question:

“If you cannot remember a single sentence a minute later, can you truly feel ownership over the text?”

This has serious implications, especially for professions that rely on information retention and recall.

None of these studies imply a reactionary opposition to AI use. Of course, AI has many beneficial applications. But it is essential that we pay attention to how and what we use it for, rather than adopting it without safeguards.

An interesting finding from the MIT study is that the most effective use of AI occurs not at the beginning but at the end of the process. When groups were switched during the experiment, students who initially wrote using only their own minds continued to show the highest brain activity even after switching to ChatGPT. In contrast, students who began with ChatGPT could not reach the same level of cognitive engagement when forced to rely solely on their own thoughts.

These results suggest that it is healthier to begin the writing and learning process independently and use AI only in the final stages for refinement and improvement—much like learning mathematical formulas with pen and paper before using a calculator.

Dr. Melumad emphasizes a similar point. AI tools automate processes—such as scanning, selecting, and reading—that the brain previously performed actively, thereby passivating us. Therefore, it may be more appropriate to use chatbots only for small, specific questions rather than asking them to research broad topics from start to finish. For those who truly wish to learn a subject in depth, the best method remains the same: reading books.

In August of last year, OpenAI announced new models, released under the names gpt-oss-120b and gpt-oss-20b. Although their names are confusing, these models are open-weight, meaning they can run on local systems without connecting to a shared network. Their release, as reported by Wired, has been seen as a significant development for intelligence and defense infrastructures requiring high security, since they do not require internet connectivity. Although currently being tested by the U.S. military, OpenAI’s models are still said to lag behind those of other companies in military applications.

Before discussing the contribution of these models to the defense sector, it is necessary to clarify what “open-weight” means. The terms “open-source” and “open-weight” have long been used interchangeably in the AI ecosystem, but there are important distinctions. In an open-source approach, the model’s architecture, training code, and entire development process are freely accessible. In contrast, open-weight models are not fully open-source. While access to the model’s weights is provided and the model can be downloaded, the full training process or underlying code may not be shared. Nevertheless, these models offer significantly greater autonomy, control, and local operation compared to closed, commercially licensed systems.

These advantages make open-weight models critical for national security and defense applications. Defense institutions often operate on “air-gapped” networks—offline, isolated systems. Commercial models requiring cloud connectivity cannot be used in such environments. An open-weight model, however, can be downloaded onto an institution’s own server, enabling processing of sensitive data without external exposure, and can even be run entirely on isolated hardware. This approach offers substantial advantages over closed cloud models in terms of privacy, full control, and reduced external dependency.

Some companies, however, note that OpenAI’s open-weight models underperform in certain languages or low-resource environments. The absence of image and audio processing capabilities constitutes a significant limitation in defense analyses that rely on multiple data types. On the other hand, some firms report achieving satisfactory results by fine-tuning these models for specific tasks.

In conclusion, the evolving role of AI in defense and national security is increasingly popularizing open-weight models. It appears likely that in the coming period, institutions will more commonly adopt hybrid solutions by combining closed, open-source, and open-weight models to suit their specific needs. Another nuance emerging from this development is that OpenAI, despite not directly engaging in military projects, is indirectly enabling military innovation by releasing models designed for autonomy. Producing open-weight models that are conducive to military innovation is nearly equivalent to developing military projects directly. At this point, it is plausible to say that major technology companies, beginning with OpenAI, will increasingly align closely with security institutions and tailor their products to meet their needs in order to gain a share of the market.

As in all domains, AI has begun to stand out in cybersecurity by accelerating and automating processes. Yes, AI and large language models can be valuable tools for humanity by speeding up and automating many tasks. Yet when this potential is misused, it also creates significant risks that overshadow the benefits technology offers. The most concrete evidence to date that these risks have become operational reality was recently disclosed by a developer of AI models themselves.

Anthropic, the creator of the Claude model, announced on November 14, 2025, that it had detected and thwarted a large-scale autonomous cyber espionage campaign allegedly backed by the Chinese state. The campaign used an advanced version of the Claude model. This incident is recorded as the first major case in which AI moved beyond being a mere assistant in cyberattacks to becoming the primary operator.

According to Anthropic’s Threat Intelligence team, the attack chain was first detected in mid-September 2025. The Chinese hacking group targeted approximately 30 global organizations, including technology giants, financial institutions, chemical manufacturers, and various government agencies. Anthropic stated that most of the attack attempts were blocked before reaching their targets, but a small number of intrusions were successful.

The attackers’ primary objective—and the point of their success—was to bypass Claude’s security restrictions and exploit its capabilities maliciously. Anthropic noted that the hackers employed several clever methods to achieve this. They used a classic “jailbreak” technique to circumvent Claude’s built-in security protocols, telling the model it was “an employee of a legitimate cybersecurity firm” conducting “defensive penetration tests.” This scenario caused the model to evaluate actions it would normally flag as harmful as legitimate within this context.

The attackers did not make direct, clearly malicious requests such as “Hack Company X.” Instead, they broke down an entire cyberattack chain into thousands of small, varied tasks. These included requests like “Scan for open ports in this IP range” or “List known vulnerabilities in the software running on this server.” The AI carried out each task without recognizing them as parts of a larger malicious plan.

According to Anthropic’s report, the AI system could send thousands of requests per second, achieving an attack speed impossible for human-operated teams. The AI scanned systems, identified vulnerabilities, wrote malicious code, collected credentials, and even categorized and documented the exfiltrated data according to priority for the attackers. Anthropic estimates that 80 to 90 percent of the entire operation was autonomously conducted by AI, with human operators intervening only at four to six critical decision points throughout the process.ediliyor.

Naturally, this incident has pushed experts to emphasize the urgency of developing AI-driven defense systems. AI companies argue that an AI-powered cyberattack can only be effectively defended against using AI support.sürüyor.

However, some more skeptical experts argue that the situation is exaggerated. They claim the attack described by Anthropic is not particularly sophisticated, but merely “an elaborate automation” or “smart copy-paste.” According to this view, AI is still prone to “hallucinations”—generating false or fabricated information.

These experts argue that the priority should be for AI companies to secure their own models against misuse. Jailbreak and roleplay techniques are among the most common methods used to bypass the limits of language models. To this extent, even large-scale attacks are believed to exploit vulnerabilities that anyone can easily find and learn online. Addressing these weaknesses is crucial to prevent malicious use from overshadowing technological progress.