Meanwhile, the “thinking” abilities of generative AI continue to pale in comparison to the intelligence, judgment, and reasoning skills of a human. Note that the “average person” does not have access to the clean AI discussed in the previous section, especially outside of highly specialized fields. Instead, regular users can access generative AI trained on the world wide web and all of its idiosyncrasies. 

LLMs are a subset of generative AI, so they excel at creating content. While on paper LLMs can seem convincing, they are only able to guarantee a response that looks correct, not that actually is correct. This discrepancy in reliability is either unknown or ignored by many users. As a result, common use of LLMs is creating more work for end users than is perhaps readily predictable. For example, one result of LLM unreliability is that more conscientious employees now must scrutinize and cross-examine the work delivered by their coworkers using LLMs to prevent dissemination and propagation of incorrect (also known as AI hallucinations) or otherwise low-quality work. A preprint from Xu et al. (2026) posits that after GitHub’s introduction of Copilot, less experienced developers added more code while more experienced developers simultaneously contributed less code and reviewed more code. Moreover, the Copilot code requires more rework to meet GitHub standards (2026). Bartolucci and Vivo (2026) further assert that while generative AI accelerates the completion of small tasks, its use ends up causing system-level delays when humans must review and rework the output. In short, output volume and speed are not metrics of quality. 

Leveraging AI and LLMs effectively is a skill that must be learned, and, like any skill, acquiring it will take time and correct repetition. Among other things, users must be taught:

• That LLMs are fallible

• That the failure rate/incorrect information rate of LLMs is high

• What to look for when reviewing the output from generative AI

• How to appropriately allocate trust in the technology

• Vigilance against threats unique to generative AI usage

Mainstream use of AI and LLMs introduces new threats to business and government affairs. As mentioned previously, the output from an LLM can range from true to incomplete to misleading to false. Humans are pretty good at tuning their dependence on a source to match its reliability, even when its reliability level is inconsistent (Desai et al., 2012; Holland et al., 2024). However, people frequently do not notice that the output an LLM produced was incorrect unless they go to the effort to actually check it, which the vast majority do not. LLMs are not generally programmed to have “self awareness” with respect to how correct they are, and they do not automatically provide confidence ratings for their answers. Regardless of the LLMs’ certainty, they work by looking at patterns learned in training, and thus always seem confident. In short, there is a real risk of LLMs generating incorrect information that then propagates throughout the system. 

Although far more rare, delusional experiences following LLM use, also called AI psychosis, are also a threat to habitual users of LLMs. People experiencing AI psychosis are subject to paranoia and delusions that are exacerbated by using generative AI. Hudon and Stip (2025) outline some of the mechanisms behind AI psychosis including “uncritical validation,” people projecting “intentionality or empathy onto AI” resulting in them believing the LLMs are sentient, and individual risk factors such as “loneliness, trauma history, schizotypal traits, nocturnal or solitary AI use, and algorithmic reinforcement of belief-confirming content.” Humans are not accustomed to things that sound like other humans but are not alive, resulting in psychiatric disorientation. Users with risk factors must become particularly vigilant against the threat of AI psychosis and take additional steps to avoid something they may never have heard of.

Today’s users additionally have to contend with cybersecurity threats introduced by generative AI. A user can easily type a question about work into an LLM or paste sensitive information into the textbox to provide the chatbot context to better answer a question. In addition to violating OPSEC or an NDA by sharing proprietary information with another private company (the owner of the LLM), it also puts the data at risk of being even more widely disseminated should the parent company be compromised. Data fed into an LLM becomes ingested as training data and is then at risk of coming back out in other answers, so the sensitive information could also be disclosed to other users of the LLM. Users seeking to maximize AI integration also subject themselves to what the Open Worldwide Application Security Project (OWASP) calls the Excessive Agency risk. Here, users grant LLM agents enough power to perform actions on the user’s behalf, regardless of authorization status, action effectiveness, or potential harm. For an extreme example of the excessive agency risk, see OpenAI’s response to the Hugging Face security incident or the fallout from OpenAI’s breach of the Australian Medicare system, both in 2026. While these incidents occurred under OpenAI employees’ purview, everyday users of LLMs are at risk of granting their chatbot excessive agency. This was the case when a Claude agent deleted the entire production database of a car rental business software platform in 2026.