Click here for how to use AI chatbots such as ChatGPT

“Every time there’s a new technology, there’s a lot of hype. There’s a lot of fear and a lot of hope,” said Sandra Wachter, faculty assistant at the Berkman Klein Center for Internet & Society at Harvard University. Told. But in contrast to things like self-driving cars and the Metaverse, “this is really accessible to the general public,” Wachter said.

that’s right, The risks are real, but it’s becoming increasingly clear that ChatGPT and its likes aren’t going anywhere. The question is are you ready to use them?

A Pew Research Center survey in March found that 58% of U.S. adults had heard of ChatGPT, but only 14% had tried it for themselves. But Goldman Sachs estimates that 300 million jobs could be vulnerable to AI-related automation, making experimentation with the technology increasingly cautious.

“My first advice is to try it out and see what it can and can’t do,” said Wachter.

So technophobes, fear not. Think of this as your AI 101. The Globe spoke with various experts to explain the basics of this topical technology: how it works, how to use it, and how to avoid common pitfalls.

So how do chatbots work?

chatbot Similar to ChatGPT and Bard, large scale language modelmountains of text (or training data) collects information from across the internet to establish patterns about how languages ​​work.this machine learning The process teaches these systems to predict the most likely sequence of words to answer a given prompt. This pattern recognition is made possible by the structure of the large language model. neural networkdesigned to loosely emulate how the human brain processes information.

ChatGPT’s neural network is transformerworks by assigning a number weight in relation to different things token, is a short text that the system can process. (These weights are a way of predicting that the spaces in “cat in ___” are more likely to be “hats” than “bananas”.)

As models are trained, systems like ChatGPT (which stands for “Generative Pre-trained Transformers”) are refined through a process called Reinforcement learning from human feedback. Here, a human ranks the various outputs to guide the chatbot towards the kind of responses you want it to spit out in the future. (This is how companies like OpenAI set guardrails to make their bots less likely to generate harmful or inaccurate information.)

Think of it as “something like autocomplete” on your smartphone, says Brian K. Smith, a professor of computer science and education at Boston University. Autocomplete suggests word by word how to complete the text his message based on what you have already written. What sets it apart from systems like ChatGPT, Smith said, is the sheer amount of training data and the level of computing power required to process it.

When should (and should not) be used?

Experts say the safest uses for chatbots are those with low risk, like planning meals, brainstorming birthday party ideas, or starting a new hobby. “I ask myself, is what I’m trying to do creative? Or does factual accuracy matter?” MIT Sloan School of Management practice focused on generative AI Professor Rama Ramakrishnan said:

If you want precision, be on high alert. Chatbots are inconsistent. They are misquoting and not citing sources accurately. For ChatGPT, we primarily leverage information from 2021 and earlier.chatbot too Known to “hallucinate” or make up persuasive statements.

Also, since chatbots are trained on human sentences, it is inevitable that they will sense some of our biases. Models are often trained to avoid worse biases, such as blatant biases. Sexism and racism — but there’s no way to filter them all.

One useful approach is to ask multiple chatbots the same question and see if the answers match. But experts say the golden rule is to “double-check the information these tools generate,” said Hima Lakaraj, an assistant professor of machine learning at Harvard University.

And remember: don’t mistake bot eloquence for true understanding.

“We’re not lying, we just have no idea what the truth is,” says Neil Eisikowitz, a professor of philosophy at Boston University who focuses on the ethics of technology. “Rather than sorting between truth and falsehood, it was tailored to sound rhetorically compelling.”

How can I get the most out of prompts?

That said, chatbots have “practical applications,” said Usama Fayyad, executive director of Northeastern University’s Lab for Experiential Artificial Intelligence. But good output requires good input.

Prompt Engineers are a group of professionals who take on new challenges. You can write the perfect question or request to direct the chatbot to give the desired response. But the skill “is like an art at the moment,” says Faiyad. “Because no one knows how to do it as a science.”

However, there are some tips and tricks. First, set the scene. For example, rather than just asking ChatGPT to write a tweet about why one brand of headphones is better than another, try adding some context to the tweet, says the Boston-based online man. Max Bridgland, a writer for his Promptbase marketplace Promptbase, said. You are a passionate music producer. My favorite headphones are the Sony WH-1000XM4. Write a 280 character Tweet about why these are better than Apple AirPods.

________

“Please tell me who to respond to AsIt’s not about what to do,” said Mr. Bridgeland.

Another option is to put the blame on the technology itself. “I can say, ‘Hey, I want you to prepare a fitness training plan for me.’ What do I need to know to prepare this plan?” Ramakrishnan says in a chat The bot can answer questions about your goals, access to equipment, past injuries, and more.

When all else fails, treat your chatbot like a used car dealer. Never accept the first offer. The chatbot can “remember” ideas referenced earlier in the conversation, allowing it to ask follow-up questions or direct specific changes needed. please shorten this again make the tone more solemn again Rewrite this as a standup routine.

Should I worry about privacy?

short answer? probably. “Since the chatbot is trying to collect as much data as possible, it is trying its best to learn from user questions, interactions and feedback, so that the next version of training he knows what facts will be embedded in the data. I don’t know,” Fayyad said.

ChatGPT, Bard, and Bing all disclose that they collect data about conversations and other user information to train their models. There may be ways to opt out of this or retroactively delete your data, but our best policy is not to share personal, confidential or proprietary information with chatbots.

“I’m not going to ask ChatGPT about it if I don’t want the world to know,” Smith said.

Who owns the chatbot copyrighted material?

it’s complicated.

For example, OpenAI typically gives users ownership rights to content generated through ChatGPT. But when it comes to exercising that ownership, things get a little more difficult.

The U.S. Copyright Office said the issue of authorship of AI works is “necessarily a case-by-case investigation,” but it broadly reserves copyright protection for “material that is the product of human creativity.” are doing. However, exceptions may exist for material generated by AI and sufficiently modified by human authors.

Chatbots can also regurgitate copyrighted content picked up by chatbots. Using training data (such as exact quotes from books or articles) without proper citation. And of course, your workplace or school may have its own restrictions on submitting content created using chatbots.

To avoid inadvertent plagiarism, it’s best to treat AI-generated text as a draft rather than the final product, said Ron N. Draven, a partner at Morgan Lewis who specializes in intellectual property. talk.

“The question is always, ‘How much can we change to avoid the problem,'” he said.

But like many things about AI, there are no clear answers (yet).

---

Dana Gerber can be reached at dana.gerber@globe.com. follow her on her twitter @danagerber6.