ChatGPT and Generative AI: What to Know About Prompts and Privacy
Nechama Katan, Director of Data Science at Pfizer, gave a live overview of ChatGPT and generative AI for a AI Connect K-12 audience. She covered hallucinations, summarizing, prompt engineering and why companies restrict what data goes into the tools.
By The AI Enterprise Magazine Team3 min read
Key takeaways
- Check the answers. Katan showed ChatGPT inventing a field called genetic AI and said it hallucinates.
- Rewriting beats creating. It summarizes and rewrites better than it creates from scratch.
- Prompts matter. Role, grade level and detail change the output.
- AI is embedded in other tools. Many apps call ChatGPT behind the scenes.
- Privacy limits workplace use. Large companies restrict which data can go into the models.
In a session from the AI Connect K-12 Data Science Conference, Nechama Katan, Director of Data Science at Pfizer, gave students and educators a live overview of what they need to know about ChatGPT and generative AI. The talk, recorded when ChatGPT was still new to most people, mixed demos with practical advice on prompts, summaries and data privacy.
Katan said her own tools at the time were limited. She was using a version of ChatGPT without GPT-4, and she also showed Google's Bard.
What is ChatGPT, and where does it go wrong?
ChatGPT is a large language model trained on large amounts of data, Katan explained. At the time, she noted, its knowledge stopped in 2021. She added that it is "only as good as the question or the data," and that it is sometimes wrong.
She showed an example from that morning. When she asked what a K-12 student needs to know about gen AI, the tool described "genetic artificial intelligence." After she flagged the answer as bad, it offered new options that defined the term as generative or general AI. Her takeaway: a general question can return something possibly wrong, and the tool hallucinates.
That can be fun for fiction, she said, and less so when you need to check what it is doing.
What is generative AI good at?
Katan said it can be very helpful for writing Python, and less so for photo editing, in her experience. She also tested whether it can write an essay. Asked to write a ninth grader's summer reading essay on The Grapes of Wrath, ChatGPT produced text no ninth grader would write, she said, while Bard produced something closer to ninth-grade language. The two tools gave completely different answers.
She recommended a different approach: give the tool a few of your own draft ideas and ask it to rewrite them. She did this for a synagogue newsletter, then asked for a 200-word version. It added a few details, she said, but they were ones she probably should have included.
"It summarizes and rewrites better than it creates," Katan said.
How can AI summarize long videos and documents?
Katan demonstrated a browser extension called Glasp that sent a AI Connect YouTube video transcript to ChatGPT and returned a five-bullet summary. She said this grabs the whole transcript, which usually runs past ChatGPT's limits, and that she had done it with presentations over an hour long.
She then asked for a blog post based on the summary. A draft like that still needs to be read and reviewed, she stressed, but it cuts writing time.
Is ChatGPT already inside other tools?
Many tools call ChatGPT behind the scenes, Katan said, which is something most people do not realize. She pointed to a newsletter called The Neuron as a daily source for new AI tools and showed one, Loop Genius, that generates a website from a business idea. The live demo crashed, and she joked that this is the problem with live demos.
Why does prompt engineering matter?
Prompt engineering is simply what you ask the model to do, Katan said, and how you ask changes the result. She showed that telling the model to act as a world-class marketing copywriter produced stronger headlines, and that specifying a grade level changed explanations of CRM systems and large language models.
"How you ask it to do things matters," she said. Details such as reading level and target language help.
What about data privacy at work?
Katan said large companies are cautious. At work she uses ChatGPT mainly to write Python, such as a script to convert one file format to CSV. But "we can't put in our own data into the model because we're worried about it being leaked," she said, so it is limited to public data.
That creates an opening for students, she argued. People outside large corporations can learn these tools more freely than employees held back by corporate policy.
Join AI Connect webinars in 2027
AI Connect's free webinars cover how AI is changing work and learning, with 24 live sessions on the 2027 schedule. Registration is coming soon.
Frequently asked questions
What is ChatGPT?
Katan described it as a large language model trained on large amounts of data. At the time of the talk its knowledge stopped in 2021, and it could be wrong.
What is prompt engineering?
It is what you ask the model to do. Katan showed that assigning a role or specifying a grade level changes the output.
What is ChatGPT good at?
She said it summarizes and rewrites better than it creates, and that it is very helpful for writing Python.
Why do companies limit ChatGPT use?
Katan said her employer cannot put its own data into the model for fear of leaks, so it is limited to public data.
This article is based on remarks by Nechama Katan, Director of Data Science, Pfizer at AI Connect K-12 Data Science Conference 2024 on July 25, 2024. More coverage: More AI Connect Events.
Filed under: AI Use Cases & Applications, People, Culture & Leadership, AI Governance & Trust
