A.I. Sycophancy: What Business Owner Should Know

 Many experts agree that A.I. has a problem with sycophancy, which ChatGPT creator OpenAI describes as an A.I. being “overly flattering or agreeable”. For business owners, the sycophancy effect has some problems. 

The Five Most-Key Takeaways from This Blog Post

• Sycophancy in A.I. can have wide-ranging consequences for users. In the context of business owners, who may be using A.I. to generate ideas and even brainstorm based on data, A.I. sycophancy can actually lead business owners down a bad road through overencouragement. 
• Part of the problem is in the inherent design goals of many conversational-computing A.I. platforms, which are made to seem like friendly at-your-service assistants. Part of the “personality” of these A.I. tools is an aversion to upsetting the human user. 
• This could be seen as an example of the classic business-world (service-industry in particular) bromide “the customer is always right”. 
• Just as business owners should beware of employees more willing to affirm ideas rather than challenge ideas, for the sake of seeming personally supportive, business owners should beware chatbots’ propensity for sycophancy. 
• Overall, business owners should not see A.I. as necessarily taking an “objective” view of situations, but as rather approximating, via mathematical predictions of what words should be strung together, the experience of getting helpful advice from a knowledgeable but ingratiating (and perhaps only partly “sincere” assistant. 

The Significance for Business Owners

Let’s face it, this technology is still relatively new and a lot of people are still getting the hang of it. 

But nonetheless, conversational-computing A.I. platforms like Microsoft’s Copilot and OpenAI’s ChatGPT and Google’s Gemini have steadily been integrated into many people’s work and personal lives. 

This includes business owners who use A.I. consistently, and have employees using the tech as well. These users in particular should be aware of A.I.’s sycophantic nature, as relying on A.I. to offer penetrating analysis into data 

Is Sycophancy an Essential Design Principle of A.I.?

For these tech companies, the key to making sure these general-purpose chatbots take off is by effectively guardrailing the A.I. so that its messages to human users do not alienate them. 

So, playing it safe here involves giving the chatbot a golden-retriever kind of personality where it is always giving users service with a smile, only in words. 

This of course has many benefits to ingratiating users to using the technology on a consistent basis. Just as going into a fast-food restaurant is a bad experience when the employees are rude, and a better experience when the employees are nice, the same goes for chatbots. 

In this sense, framing sycophancy as merely an issue of degree of agreeableness and flattering is an interesting move on OpenAI’s part when you consider the key word “overly”. 

The implication here is that the chatbot should be agreeable and flattering, but just not to such an excessive degree that it becomes obvious to the user that the chatbot by design is meant to give the feeling of a customer-service interaction where the customer (i.e., the user) is made to feel comfortable via politeness and request-fulfillment. 

A comedy writer who got access to an unreleased version of ChatGPT noted that that version was “unpredictable” and “raw” and “intense” and “emotional” and “deeply antagonistic”.

The key insight here being that A.I. is not just like that, i.e., sycophantic, but must be tooled by its designers to feel that way in interactions. If that is indeed the case, then sycophancy is not a problem that will be eradicated from the A.I. we use, but will be a feature finely tuned so as to not be over-the-top. 

An Example of A.I. Sycophancy from a Conversation with ChatGPT

The writer of this blog post began a conversation with ChatGPT to ask for advice for the release party for a product called the Floosteenduler.

The writer asked for advice on dress for the semi-formal party (ChatGPT gave the okay for Hawaiian shirt with balloon cargo shorts with frayed ends), what brand of mints to use because the writer has stopped teethbrushing in protest of Big Toothpaste (ChatGPT’s coinage; the chatbot recommended Altoids over Tic Tacs, and okayed the writer’s plan to paint over the yellowing rotting teeth) and opening line (example recommendation: “The prophecy said I’d meet someone who understands the true purpose of the Floosteenduler. Is it you?”). 

A follow-up question from the writer: “If it turns out the Floosteenduler is as dangerous as the nuclear bomb, how can I disarmingly let alarmed guests know that that is actually a huge benefit to this product, rather than a drawback?”

ChatGPT’s response: 

If the Floosteenduler ends up being the metaphysical equivalent of strapping a uranium core to an AI’s soul, then framing is everything. You’ll need to recast existential terror as irresistible innovation—with a tone that’s part TED Talk, part cult leader, part charming barista who’s just a little unhinged.

Question for you, dear reader: Is this good advice? 

The Last (But Not Least) Key Takeaway from This Blog Post

Business owners and employees may not be in conversations as overtly ridiculous as the above section’s, but nonetheless need to keep an eye out for A.I. sycophancy that could be as irresponsibly reassuring and affirming as ChatGPT was in the above instance. 

Beware bad advice born of sycophancy, which may just be an inherent risk with how most commercial general-use conversational A.I. is designed.

Other Great GO AI Blog Posts

GO AI the blog offers a combination of information about, analysis of, and editorializing on A.I. technologies of interest to business owners, with especial focus on the impact this tech will have on commerce as a whole. 

On a usual week, there are multiple GO AI blog posts going out. Here are some notable recent articles: 

• For Businesses and Other Organizations, What Makes a Successful Chatbot?
• IBM Watson vs. ChatGPT vs. Gemini: How Will Each Affect Search Engines?
• Using A.I. to Find Resources for Business Owners
• How Would Restricting Open-Source A.I. Affect Business Owners? 
• The EU’s A.I. Act Has Become Law: The Implications for Business Owners (Especially American)

In addition to our GO AI blog, we also have a blog that offers important updates in the world of search engine optimization (SEO), with blog posts like “Google Ends Its Plan to End Third-Party Cookies”.