Lennart Meincke, Senior Applied Scientist at the Mack Institute and Wharton Generative AI Labs (GAIL), followed an unusual path to his current role. A software engineer in his native Germany, he originally came to Penn to study in the College of Liberal and Professional Studies, where he took classes in topics like Freudian psychology and criminology. From there, he got “sucked back into” engineering and has since become a major collaborator on some of the most influential generative AI research coming out of Wharton and Penn.
We spoke with Meincke about his path into AI research, why people find AI both fascinating and “scary,” and the personal rules he follows when using these tools.
Q: You were one of the first people the Mack Institute tapped to work on generative AI when ChatGPT emerged. How did you first become interested in it?
I was working with Drew Carton on a management simulation for Wharton students. We wanted students to be able to write free-text responses instead of relying on multiple-choice questions, but evaluating those responses automatically was incredibly difficult.
One of Drew’s exercises asked students to write a company vision statement. We’d built some rule-based logic to assess those responses, but it was pretty crude. So we went to Chris Callison-Burch in Engineering and asked, “You’re the NLP expert. Is there a better way to do this?”
He said, “Well, there’s this thing called GPT-3.”
This was several months before ChatGPT launched. He showed us a few examples, and we immediately realized it was dramatically better than anything we’d built ourselves.
Once ChatGPT came out, I read Christian Terwiesch’s Would Chat GPT Get a Wharton MBA?, cold-emailed him with a few ideas for extensions, and that’s how we started working together.
Q: A lot of your research asks AI to do very human things: make ethical judgments, be creative, give orders. Why do you think people find that both fascinating and unsettling?
I think it is scary because it changes how we think about our own work.
Before I answer, I want to give a small shout-out to our colleague Gideon Nave in Wharton Marketing, who coined the term cognitive surrender. I think it’s a great description of something I try to avoid myself. You don’t want to completely give yourself over to the system and just press Enter. You want to preserve your own thoughts and your own identity.
There are times when you’re working on something, you ask AI for a few new directions, and they’re good. You think, “Okay, let’s do that.” That’s a really strange way to work. It makes you wonder: Which part of this work do I actually own? Where is my agency?
I remember sending a faculty member an abstract that ChatGPT had mostly written. He replied, “This is one of the best abstracts I’ve ever read.” But I didn’t feel very happy about it because I hadn’t really written it.
Over time, I’ve accepted that these tools are now part of our workflow. Ultimately, though, I’m still responsible for deciding whether that’s the piece of work we’re going to use. I’m the one on the hook for that.
Q: Your recent paper looked at how AI responds to persuasive techniques, specifically how human persuasion methods can be used to convince AI tools to do things that are harmful. What surprised you most?
Our original question was whether we could systematically change these systems’ behavior. But our broader finding is that they’re almost parahuman. They respond to many of the same persuasive techniques humans do.
It was never just about jailbreaking models. The more interesting result was that principles like persuasion and flattery work on AI in surprisingly similar ways to how they work on people.
That’s also what makes it a little unsettling. You can pick up a book on human persuasion, apply those same principles to AI, and they work remarkably well. Given the way these systems are currently designed, I think it’ll be very difficult to eliminate that completely.
Q: Where do you come down on questions of responsibility? If someone harms themselves after interacting with AI, who should be accountable?
I don’t think I have a perfect answer. My instinct is that some systems should probably be age-restricted. Adults ultimately make their own choices, but children are different. There may be certain topics where some form of parental controls makes sense.
More broadly, I think we have to weigh the benefits against the harms. Billions of people get genuine value from these tools. That doesn’t excuse misuse, but every transformative technology has produced both tremendous benefits and real risks.
Q: It can be hard to set limits with AI. Even people who understand the tools well can fall into the trap of over-relying on it or treating it like a person.
Yes. That’s exactly the challenge.
People will say, “What does Claude think?” and my response is usually, “Maybe we can think about it ourselves for a couple of minutes first.”
It goes back to Gideon’s idea of cognitive surrender. Even if AI can make your life easier, that doesn’t always mean it should. To me it’s like walking instead of driving, or taking the stairs instead of the elevator. The easier option isn’t always the one that’s better for you.
I know people working in AI whose guiding principle is, I’m not going to automate anything I wouldn’t know how to do myself. I don’t follow that completely, but I have a lot of sympathy for it.
Q: What’s the most important thing people your age should understand about AI?
I think—and maybe it’s a bit of a somber note—I don’t think AI is going to go away. So use it where it can be useful, but still invest in your skills.
Especially if you’re in high school or early college right now, I understand it’s very scary and all the incentives are stacked against you. You’re incentivized to use it so you can move faster. But there are certainly some meta-skills, like critical thinking, that are worth developing before you lean too heavily on these systems.
I actually think that’s not all too different from what I saw when I was a student. Even before AI was a big thing, a lot of students would try to do the bare minimum on an assignment: search online, then copy and paste whatever they needed. In the long run, it really didn’t serve them well. It was the people who spent 12 hours trying to understand something who ultimately had much more successful careers. They aced the interviews because they took the time to build the fundamentals, even if it meant taking much longer.
I understand that’s a huge sacrifice. I’m self-taught, so I went through that myself. I know it takes much longer. But I think, for many things, it’s well worth it.