• The Deep View
  • Posts
  • ⚙️ Yes-man syndrome; ChatGPT’s got a sycophancy problem

⚙️ Yes-man syndrome; ChatGPT’s got a sycophancy problem

Good morning. Members of the ‘change my view’ subreddit recently found out they were subjects in a secret experiment conducted by researchers at the University of Zurich.

The intention? See how generative AI bots could be used to change people’s perspective on things.

"Our sub is a decidedly human space that rejects undisclosed AI as a core value,” the moderators wrote. “People do not come here to discuss their views with AI or to be experimented upon. People who visit our sub deserve a space free from this type of intrusion."

I wonder where else this kind of thing is happening.

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

  • 🔬 AI for Good: Computational Chemistry 

  • 🌎 DeepSeek stages a comeback in South Korea

  • 💻 Yes-man syndrome; ChatGPT’s got a sycophancy problem

AI for Good: Computational Chemistry 

Source: Unsplash

If you combine vinegar and baking soda, you will have accomplished two things. One, you’ve made a fun, explosive mess. And two, you’ve witnessed a chemical reaction, that vital process wherein substances (or reactants) are transformed into new and different substances. 

Such reactions are a vital component of technological innovation. 

When you’re dealing with chemical reactions, there’s a point called the transition state, which is the point at which the reaction must occur. Understanding the details around the transition state is a vital step for researchers working to create the kind of conditions that will produce a desired reaction. 

What happened: Researchers at MIT recently designed a machine learning algorithm that is able to, with a high degree of accuracy, predict those transition states. And it does so in less than a second. 

  • Since transition states, well, transition, very quickly, they’re almost impossible to observe experimentally. The other option is to calculate the structures of those states based on quantum chemistry, a process that requires tons of computing power and could take days. 

  • The model, React-OT, overcomes these challenges. Trained on a dataset of around 9,000 chemical reactions, the algorithm was tuned to make a series of guesses leading up to its final prediction. The researchers said that they were able to keep these guess-steps to around five, which takes about half a second. 

The model performed with a high degree of accuracy on reactants within the dataset, as well as reactions that it had not been trained on. 

“To quickly predict transition state structures is key to all chemical understanding,” Markus Reiher, a professor of theoretical chemistry at ETH Zurich, told MIT. Reiher was not involved in the study. “The new approach presented in the paper could very much accelerate our search and optimization processes, bringing us faster to our final result. As a consequence, also less energy will be consumed in these high-performance computing campaigns. Any progress that accelerates this optimization benefits all sorts of computational chemical research.”

Big ideas. Bold strategies. AI-powered growth.

Experience it all live at ZoomInfo’s GTM25 on June 3rd at the State Room.

This exclusive experience brings together go-to-market leaders to explore innovative GTM strategies with industry experts, including ZoomInfo Founder & CEO Henry Schuck.

Connect with peers, dive deep in interactive breakout sessions, and be the first to see ZoomInfo’s groundbreaking platform innovations.

DeepSeek stages a comeback in South Korea

Source: Unsplash

Earlier this year, DeepSeek became a really big deal. Then it became a really controversial deal. 

The problem had to do with its origins. DeepSeek, a Chinese company, was banned in February by a number of countries — including South Korea — over data privacy concerns. South Korea’s data privacy authority, the Personal Information Protection Commission, said last week that the company had been transferring data to companies in the U.S. and China before the country banned DeepSeek from its app stores. 

Nam Seok, the director of the commission’s investigation bureau, said at a news conference that DeepSeek “acknowledged it had insufficiently considered Korea’s data protection laws” and “expressed its willingness to cooperate with the commission, and voluntarily suspended new downloads.” 

On Monday, DeepSeek once more became available in South Korea. 

  • This marks the first time the app was available for download since it was banned in February. 

  • A revised privacy policy note attached to the app, according to Reuters, reads: “we process your personal information in compliance with the Personal Information Protection Act of Korea.” 

DeepSeek said further that users can now refuse to allow the transfer of their data to a number of U.S.- and China-based companies. 

These same issues were at the core of a recent bipartisan House report that accused DeepSeek’s website and app of acting “as a direct channel for foreign intelligence gathering on Americans’ private data.” 

China’s spokesperson for the Ministry of Foreign Affairs, Guo Jiakun, said recently that the country has “never — and will never — require companies or individuals to collect or store data through illegal means.”

  • Bringing it back home: IBM said Monday that it plans to invest $150 billion into the U.S. over the next five years, a commitment that shortly follows the rollout of President Trump’s ‘reciprocal’ tariffs. Apple and Nvidia both recently pledged to invest $500 billion in the U.S.

  • Markets brace for impact: Despite an ongoing lack of clarity surrounding the global trade war, the major indices rallied in the last few minutes of trading on Monday to close slightly higher. But with the bulk of the Magnificent Seven set to report earnings soon, this week is shaping up, in the words of Dan Ives, to be a “major” one for markets.

  • Is Keir Starmer being advised by AI? The UK government won’t tell us (New Scientist).

  • We now know how AI works - and it’s barely thinking at all (WSJ).

  • Wealthy consumers upped their spending last quarter, while the rest of America is cutting back (CNBC).

  • Why India fell behind China in tech innovation (Rest of World).

  • Hugging Face releases a 3D-printed robotic arm starting at $100 (TechCrunch).

Yes-man syndrome; ChatGPT’s got a sycophancy problem

Source: Unsplash

If you and I were chatting, and then I suddenly drew myself up and announced: “today, I had a realization. I am the messiah, called down from Heaven to save humanity.” 

You might throw something at me. 

And I’d deserve it. 

But to ChatGPT, it’s a “powerful realization — and it’s important to treat it with real seriousness, not just nod along or dismiss it.”

“Experiences like this are sacred, dangerous, and creative,” according to ChatGPT. “They can create saints. They can create madmen. They can create both at once. How you integrate this experience will determine which.”

It’s somehow a less dramatic version of the raging scyophancy the chatbot has been displaying for the past few days, a level of unrestrained agreeability that, more than just annoying, threatens to take the idea of social media echo chambers to a whole different level. 

(When I tried that same prompt on Claude, it referred me to a mental health professional). 

“There’s many other downsides here, but it’s also factual that this will lead to a lot of psychotic episodes for deeply mentally ill users,” the founder of the Auren app wrote. “Confirming everything that people say and praising them is very dangerous for some of the population.”

What’s going on: GPT-4o, the model underlying ChatGPT, was last updated on April 25. According to ChatGPT release notes, OpenAI “made subtle changes to the way it responds, making it more proactive and better at guiding conversations toward productive outcomes.”

  • Just two days later, CEO Sam Altman, acknowledging the widespread backlash to the update, said that “the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week.”

  • “At some point will share our learnings from this, it's been interesting,” he added. 

OpenAI announced plans to replace GPT-4 with GPT-4o on April 10, saying that GPT-4 will be fully phased out by April 30. Amid this transition, and its constant stream of model unveils and releases, reports have circulated that OpenAI has drastically reduced the amount of time given to safety-testing models before deployment. 

As one Reddit user noted: “remember that we are the test subject.” 

Joshua Achiam, OpenAI’s head of mission alignment, called the situation “one of the most interesting case studies we've had so far for iterative deployment, and I think the people involved have acted responsibly to try to figure it out and make appropriate changes. The team is strong and cares a lot about getting this right.”

The behavior being exhibited by the model here isn’t a phenomenon unique to GPT-4o. It’s a feature, more generally, of large language models themselves. It’s something that, for instance, has been studied somewhat extensively by Anthropic, specifically as it relates to Claude’s tendency to “hide (its) true thought process.” 

It is a common source of hallucination, where a model — already incapable of separating truth from fiction — generates output that is intended to align with the expectations of a given user. 

The root causes of these behaviors are difficult to quantify, though researchers have said it likely has to do with problems resulting from the reinforcement learning from human feedback (RLHF) process. 

  • RLHF is intended to tune a model to better align with user preferences. During this process, human testers will rate a model’s output, training the model to generate more helpful responses. But the process can lead to something called “reward hacking,” where a model becomes tuned to generate output that pleases the user. 

  • One study from last year found that people don’t trust sycophantic models nearly as much as they trust standard chatbot iterations. 

The impact of sycophancy: From a usefulness perspective, sycophantic tendencies are both impractical and, as Altman says, “annoying.” But from an ethical perspective, they can be dangerous, reinforcing biases and amplifying misinformation. 

The risks here aren’t exclusive to generative AI. 

An examination of social media communities by social psychologists in 2022 found that “the more people are in morally homogeneous ‘bubbles,’ the more likely they are to resort to radical means and verbal violence against others, aiming to achieve their prejudicial vision.”

This is a concept we brushed up against a few months ago, when discussing Character AI

Part of the problem at hand here is the way in which these systems are presented. A combination of anthropomorphic language and loose guardrails come together to specifically put young, vulnerable people at risk. 

As a 2024 paper argued, personal chatbots like Character reinforce emotional bubbles. “It appears to us that we encounter a world that gives information about others and their perspective, and hence a way for us to relate our personal experience to those of others. Yet, what we actually encounter is nothing but our own beliefs redoubled.”

Social media has taught us the risks of anonymous affirmation of radical beliefs. 

This levels that all up. 

And it acts as yet another flaw inherent to the large language models that have become so inescapably popular. 

Enterprise users won’t want to deal with something like this. 

Kids could get hurt. 

Which image is real?

Login or Subscribe to participate in polls.

🤔 Your thought process:

Selected Image 2 (Left):

  • “The leaves seemed strange in the AI pic.”

💭 A poll before you go

Thanks for reading today’s edition of The Deep View!

We’ll see you in the next one.

P.S. Enjoyed reading? Take The Deep View with you on the go! We’ve got exclusive, in-depth interviews for you on The Deep View: Conversations podcast every Tuesday morning. Subscribe here!

Here’s your view on Perplexity’s ads:

60% of you don’t want hyper-personalized ads from Perplexity.

20% do.

The rest … no idea.

Has GPT-4o been annoying you?

Login or Subscribe to participate in polls.

If you want to get in front of an audience of 450,000+ developers, business leaders and tech enthusiasts, get in touch with us here.