AI chatbots could help deliver care, but with caution

On Reddit forums, many users discussing mental health have been enthusiastic about their interactions with ChatGPT. ChatGPT is an artificial intelligence chatbot from OpenAI that conducts human-like conversations by predicting the next word in a sentence. “ChatGPT is better than my therapist,” wrote one user, adding that the program listened and responded when the person said he was having trouble controlling his thoughts. rice field. “Very scary, I feel like I’m hearing ChatGPT.” Other users talk about asking ChatGPT to act as their therapist because they can’t afford a real one.

The excitement is understandable, especially given the shortage of mental health professionals in the United States and around the world. People seeking psychological help often face long waiting lists, and treatment and other mental health care are not always covered by insurance. Sophisticated chatbots such as ChatGPT and Google’s Bard can help manage therapy, even if they can’t ultimately replace therapists. “There is no such thing in medicine” [chatbots] Like mental health, it will be very effective,” said former director of the National Institute of Mental Health and co-founder of Vanna Health, a startup that connects people with severe mental illness to caregivers. says Thomas Insel. In the mental health field, “There are no procedures. There are chats. We are in communication.”

But many experts wonder whether tech companies respect the privacy of vulnerable users, whether they program appropriate safeguards to prevent AI from providing false or harmful information, or whether they have severe mental illness. I worry about prioritizing treatment for the wealthy and healthy at the expense of others. “While we appreciate the improved algorithms, they ultimately deal with the more confusing social realities people face when they seek help,” said Julia Brown, an anthropologist at the University of California, San Francisco. I don’t think I can,” he says. .

therapist’s assistant

The concept of a “robot therapist” has existed since at least 1990, when computer programs began offering psychological interventions that guided users through scripted procedures such as cognitive-behavioral therapy. These days, popular apps like those from Woebot Health and Wysa employ more advanced AI algorithms that can talk to users about their concerns. Both companies claim that their apps have been downloaded more than 1 million times. Also, chatbots are already being used to screen patients by conducting standard questionnaires. Many mental health providers in the UK’s National Health Service use chatbots from a company called Limbic to diagnose certain mental illnesses.

However, newer programs such as ChatGPT are much better at interpreting the meaning of human questions and responding in a realistic way than previous AIs. These Large Language Model (LLM) chatbots are trained on vast amounts of text from across the internet, adopt different personas, ask questions of users, and draw accurate conclusions from the information users provide. can be derived.

As assistants to human health care providers, LLM chatbots have the potential to significantly improve mental health services, especially among marginalized and critically ill people, Insel said. A critical shortage of mental health professionals, especially those willing to work with those incarcerated or experiencing homelessness, is exacerbated by the time providers need to spend on paperwork. Yes, Insel says. Programs such as ChatGPT make it easy to summarize a patient’s session and create the reports you need, allowing therapists and psychiatrists to spend more time treating their patients. “By offloading paperwork and reporting to machines, we can increase our workforce by 40% of his,” he says.

But using ChatGPT as a therapist is a more complicated matter. While the idea of leaking secrets to a machine might baffle some, LLMs sometimes respond better than many human users, says Tim Alsoff, a computer scientist at the University of Washington. His group is an LLM program that studies how crisis counselors express empathy in text messages and gives writers feedback based on the strategies people use that are most effective in getting people out of crisis. have trained

“There are many other [to therapy] Rather than put this in ChatGPT and see what happens,” says Althoff. His group has collaborated with the nonprofit Mental Health America to develop an algorithm-based tool that enhances ChatGPT. When the user enters a negative thought, the program suggests ways to reconstruct that particular thought into a positive one. More than 50,000 people have used the tool so far, and Altoff says users are more than seven times more likely to complete the program than similar tools that return canned responses.

Empathetic chatbots can also benefit peer support groups such as TalkLife and Koko, where people without professional training send helpful, uplifting messages to other users. In a study published in nature machine intelligence When Altov and colleagues asked peer supporters to create messages using empathetic chatbots in January, nearly half of the recipients were written with the help of chatbots rather than human-only texts. They found that they liked texts and rated their empathy as 20 percent higher.

But human involvement is still important.In an experiment conducted by Coco co-founder Rob Morris, explained on twitter, company leaders have found that users can often determine whether a response is from a bot or not, and they dislike those responses once they know the message was generated by AI. (The experiment sparked backlash online, but Morris said the app included a note informing users that some of the messages were written by AI.) Apparently, “Efficiency We like the complexity of human interaction that exists, even though it comes at the expense of quality and quality.” before,” says Morris.

Researchers and companies developing mental health chatbots claim they are not trying to replace human therapists, but rather complement them. After all, people can talk to chatbots whenever they want, not just when they have an appointment, said Joseph Gallagher, Woebot’s chief program officer. This speeds up the treatment process and encourages people to trust bots. The bond between therapist and client, the therapeutic partnership, is believed to be a large part of therapy effectiveness.

Woebot researchers without ChatGPT, in a study of 36,000 users, compared it to human chatbots based on standard questionnaires used to measure therapeutic engagement. found that users were able to build trust with their chatbot within four days, compared to weeks. therapist. “I hear people say, ‘How can you tell a human being to do this,'” Gallagher said. “It reduces risk and reduces vulnerability.”

The risks of outsourcing care

But some experts worry that trust could backfire, especially if chatbots aren’t accurate. According to a theory called automation bias, people are more likely to trust advice from machines than advice from humans, even if it’s wrong. “Even if it’s beautiful nonsense, people tend to accept it,” says Evian Van Dis, a clinical psychology researcher at the University of Utrecht in the Netherlands.

Also, there are still limits to the quality of advice that chatbots can provide. They may not perceive information that humans would time as indicating a problem, such as asking how a severely underweight person can lose weight. Van Dis argues that if the medical literature (believed to be from wealthy Western countries) in which the AI program was trained contains biases, the AI program could be biased against certain groups of people. I am afraid that it will be lost. They may miss cultural differences in how mental illness is expressed, or draw the wrong conclusions based on how users write in their second language.

The biggest concern is that chatbots could harm users, for example by suggesting that they stop treatment or recommend self-harm. In recent weeks, the National Eating Disorders Association (NEDA) has closed its previously human helpline in favor of a chat called Tessa, which is not based on generative AI and instead provides users with scripted advice. It has been criticized for introducing bots. Some users have posted on social media that Tessa occasionally gives weight loss tips that could trigger people with eating disorders. NEDA shut down the chatbot on May 30 and said in a statement that it is reviewing what happened.

Limbic CEO Ross Harper said of AI chatbots not adapted for medical purposes, “The current format is not suitable for clinical settings where reliability and accuracy are paramount. ‘ said. He fears mental health app developers will accidentally develop harmful ones if they don’t change the underlying algorithms to incorporate proper scientific and medical practices. there is “It can set the whole field back,” Harper said.

Wysa’s head of clinical development and research, Chaitali Sinha, said that while the government is looking for ways to regulate AI programs like ChatGPT, Wysa’s industry is at a kind of dead end. “If it can’t be regulated, it can’t be used in clinical settings,” she says. Van Dis raised concerns about how tech companies collect and use the information users enter into chatbots, raising concerns about potential breaches of confidentiality, and how chatbots were created in the first place. He adds that the general public knows very little about what they were trained to do.

Limbic, which is testing a ChatGPT-based therapy app, is looking to address this issue by adding another program that limits ChatGPT’s responses to evidence-based therapy. Harper said health regulators could still evaluate and regulate this program and similar “layered” programs as medical products, even though the law on the underlying AI program is still pending.

Wysa is currently applying to the U.S. Food and Drug Administration to approve a chatbot that delivers cognitive-behavioral therapy as a medical device, which Sinha said could be approved within a year. Wysa uses an AI that he does not ChatGPT, but Sinha said the company could consider generative AI once regulations become clearer.

Without regulation, Brown fears it will be left to emotionally vulnerable users to decide whether chatbots are trustworthy, accurate and helpful. She also believes that commercial chatbots are primarily aimed at “concerned people”, i.e. treatment or app subscriptions, rather than isolated individuals who may be most at risk but don’t know how to get help. I am concerned that it will be developed for people who can afford to pay premiums.

Ultimately, Insel says the question is whether any treatment is better than none. “Treatment is best when there is a deep connection, but many people often don’t, and it’s hard to get quality care,” he says. Training enough therapists to meet the demand is nearly impossible, and partnerships with experts and carefully developed chatbots could greatly reduce the burden. “Having these tools available to the masses is the way out of this situation,” Insel said.

Source link