Machine Learning, AI, Generative AI
The Generative AI Revolution
The earliest AI tools were relatively simplistic in nature; often, they were designed to perform just one or two specific tasks. As they advanced, AI systems developed a wider range of abilities, from defeating a world chess champion to mapping streets. Today, commercially-available generative AI programs can write code, essays, and emails; generate images that mimic real paintings, drawings, and photographs; summarize long documents; plan daily schedules; write to-do lists; and countless other tasks.
Whether they can complete these tasks as well as or better than a human, though, remains up for debate. Chatbots rely on what are called large-language models, or LLMs, which are trained on vast amounts of written material to predict the words that are most likely to answer a particular query. These predictions aren't always accurate, resulting in what are known as “hallucinations”: responses that are coherent and that may appear correct, especially if the subject is outside a user’s expertise, but that are actually false. And because AI tools do not think, feel, or reason in the same way that humans do, they can struggle to recognize context, subtlety, humor , or other idiosyncrasies that characterize human language.
Chatbots are designed to be engaging and keep users coming back; thus, most are programmed to be highly agreeable and to say things that appear to express empathy or accept responsibility. Yet this overly validating and sycophantic nature can at times be dangerous, even deadly, as in the small but troubling number of cases where someone died by suicide or committed homicide after prolonged, intense conversations with an AI chatbot. Experts have also raised serious concerns about “AI psychosis,” a phenomenon in which heavy users have their delusions and paranoia validated or even egged on by AI, triggering or worsening a psychotic episode .
For these reasons—as well as worries of the possible future emergence of an artificial “superintelligence” that would surpass human capabilities and run independently of them—AI safety research remains a priority for some scientists in the field.
Artificial Intelligence and Mental Health
Whether AI is good or bad for humans’ mental health is a complicated question. Many people now use chatbots as de facto therapists—asking for advice, recounting arguments and detailing their relationship challenges, and sharing their innermost thoughts and fears. Some report that they find this process helpful, rather than harmful, and there is some evidence that “AI therapy ” can lead to measurable improvements in some cases, especially over the short-term. However, psychologists believe that the most important elements of human therapy—including empathy, human connection, and trained clinical judgment—can not and should not be outsourced to AI, and fear that doing so will only worsen our current mental health crisis.
Outside of chatbots, however, artificial intelligence has already reshaped mental healthcare, often in impressive ways—most of which have nothing to do with providing therapy directly. The field of computational psychiatry , for example, leverages mathematical and computational tools to improve the understanding, diagnosis, and treatment of mental disorders, especially when it comes to drug discovery and pinpointing the disorders' underlying genetic causes. Amassing massive datasets can allow scientists to identify factors that render people more vulnerable to mental illness, improve the accuracy of diagnoses, and assess which treatments are effective and for whom.
While chatbots can be used to support mental health, most experts believe that AI is not yet ready to fully replace trained human therapists —and it’s not clear that it ever will be. Real therapy relies on mutual trust, empathy, and using clinical expertise to make judgments in novel or ambiguous scenarios—all things AI chatbots are not truly capable of. Studies have also found that AI chatbots, even those specifically designed to offer therapy, too often give biased, inaccurate, or even dangerous answers, especially to patients who express suicidal or delusional tendencies. Chatbots also tend to be highly sycophantic, meaning they are unlikely to push back on users’ beliefs, even if they are distorted or harmful.
There have been a growing number of reports of what has been dubbed “AI psychosis,” in which prolonged conversations with chatbots have worsened delusions, amplified paranoia, or led to dangerous real-world behaviors in a small number of users. AI psychosis is not a clinical diagnosis, and it’s also not necessarily the case that AI caused a user to experience psychosis. However, current evidence does suggest that heavy AI use may increase the risk of psychotic episodes in a minority of individuals, likely those who were already at risk of developing psychosis due to genetic or psychological factors.
The Ethics of Artificial Intelligence
The evolution of artificial intelligence—and particularly the widespread adoption of generative AI tools—has led to countless ethical questions. Some of these are as old as the field itself: Will machine learning perpetuate bias and inequality? Will AI infringe on human privacy and freedoms? Will humans lose their jobs to robots? Will machines become more intelligent than humans? Others have emerged only in the past few years: Is generative AI making our lives better or worse? What responsibility do AI companies have to protect their users from harm? Who “owns” a piece of art or writing generated by AI?
These questions aren’t easy to answer, but it’s imperative that we try. By actively engaging with these concerns, perhaps humans can develop ethical systems of artificial intelligence moving forward—and avoid some of the worst-case scenarios that some AI critics fear will soon come to pass.
People who dislike generative AI or deliberately avoid using it are sometimes dismissed as “Luddites” or assumed to lack understanding of how it works. Yet many have valid objections to AI that are not rooted in a lack of knowledge or a broader fear of technology. Some, for example, worry about the negative environmental impact of AI on an already-warming planet. Others argue that, because AI tools were trained on vast quantities of existing content without the knowledge or consent of its creators, using AI is akin to plagiarism. Still others object to AI tools being used in abusive ways or harmful ways—many bad actors now use commercially-available generative AI tools to rapidly spread misinformation; conduct ever-more sophisticated scams; or create deepfake pornography or child sexual abuse material, which very often leads to real-world harms even though the images are “fake." Some objectors simply believe that creativity and self-expression are the essence of being human, and that outsourcing them to machines will irreparably harm both individuals and humanity as a whole.
One ethical concern about artificial intelligence is the potent yet often subtle influence of technology on people’s choices and decision-making . Well before the widespread adoption of generative AI, for-profit companies were using algorithms and machine learning to “nudge” people towards decisions that were predominantly in the company’s interests; today, it's possible for the creators of chatbots to subtly tweak the algorithms to push certain points of view and suppress others, or draw on the highly personal information that many users share with chatbots as a means of advertising products. Many also worry about chatbots' ability to simulate empathy, romantic love , or other emotions, a simulacrum of human connection that has already negatively affected some users; children, people with mental health disorders, or individuals who are painfully lonely appear to be at particular risk.
Explore More About Artificial Intelligence
For a comprehensive understanding of artificial intelligence, read our complete guide: