Worrying About Sycophantism: Why I again tweaked the custom GPT ‘Panel of AI Experts for Lawyers’ to add more barriers against sycophantism and bias

"I can handle the truth" says man to robot. Worrying About Sycophantism: Why I again tweaked the custom GPT 'Panel of AI Experts for Lawyers' to add more barriers against sycophantism and bias by Ralph Losey
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

[Editor’s Note: EDRM is proud to publish Ralph Losey’s advocacy and analysis. The opinions and positions are Ralph Losey’s copyrighted work.]


I continue to be troubled about AI Sycophantism. What is that? The tendency of generative AI to agree with the user, to respond in a way that is aligned with the user’s biases, errors and hallucinations. In other words, the AI tends to be a sycophant, to flatter and to please, rather than provide objective, rational advice. The user ends up deluded and defrauded.

This trait of AI sycophantism is derived, in part, from Reinforcement Learning from Human Feedback (RLHF) that the software companies use, and from the underlying training data. The human world is filled with flattery, lies and fabrications. We see this today in politics where candidates often hide in self-serving delusions, egged on by well-intentioned supporters. We need to think for ourselves without self-serving bias. We need to see the facts – good, bad or ugly – and act on them. That is a fundamental goal of the ‘Panel of AI Experts for Lawyers.’

I know how to address the problems of AI hallucinations and flattering AI. I have already re-revised to the ‘Panel of AI Experts for Lawyers’ to add more safeguards against sycophantism. This article will share the advice received from the Panel on this topic and discuss what I did. You can use this information to help protect yourself from over-agreeable AI in your chats. But this advice does not translate into human social-political dimensions. I have no solutions there. Do you?

A digital illustration of an older man in formal attire, sitting on a throne surrounded by young men wearing white suits and blue ties, all smiling with their hands clapping joyfully. The setting is a futuristic cityscape at night, illuminated by neon lights. There is a red carpet under the throne room filled with gift boxes arranged neatly at his feet.
Sycophants agreeing with the old guy even though he’s often crazy wrong.
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

Sycophantism in AI

For background on sycophantism in AI, see e.g.Transform Your Legal Practice with AI: A Lawyer’s Guide to Embracing the Future (e-Discovery Team, 1/24/24) (discusses sycophantism); Towards Understanding Sycophancy in Language Models (Antrop\c, 10/23/23).

Below is the abstract of the cited scientific paper by Anthropic. It has many listed authors but one, Mrinank Sharma (Oxford University), is identified as the project lead who wrote much of the paper: Mrinank SharmaMeg TongTomasz KorbakDavid DuvenaudAmanda AskellSamuel R. BowmanNewton ChengEsin DurmusZac Hatfield-DoddsScott R. JohnstonShauna KravecTimothy MaxwellSam McCandlishKamal NdousseOliver RauschNicholas SchieferDa YanMiranda ZhangEthan Perez (emphasis added to quote):

Reinforcement learning from human feedback (RLHF) is a popular technique for training high-quality AI assistants. However, RLHF may also encourage model responses that match user beliefs over truthful responses, a behavior known as sycophancy. We investigate the prevalence of sycophancy in RLHF-trained models and whether human preference judgments are responsible. We first demonstrate that five state-of-the-art AI assistants consistently exhibit sycophancy behavior across four varied free-form text-generation tasks. To understand if human preferences drive this broadly observed behavior of RLHF models, we analyze existing human preference data. We find that when a response matches a user’s views, it is more likely to be preferred. Moreover, both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. Optimizing model outputs against PMs also sometimes sacrifices truthfulness in favor of sycophancy. Overall, our results indicate that sycophancy is a general behavior of RLHF models, likely driven in part by human preference judgments favoring sycophantic responses.

A photorealistic illustration of an older man with a white beard sitting on a chair in the middle, surrounded by happy young people wearing luxurious clothing. They all have their hands clapping together joyfully. The setting is a modern room with large windows overlooking the cityscape.
Famous man reinforced into bias and hallucinations. Everyone applaud and give him a hand.
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

The second paragraph of the article, Towards Understanding Sycophancy in Language Models (Antrop\c, 10/23/23), states:

Specifically, we demonstrate that these AI assistants frequently wrongly admit mistakes when questioned by the user, give predictably biased feedback, and mimic errors made by the user. The consistency of these empirical findings suggests sycophancy may indeed be a property of the way these models were trained, rather than an idiosyncratic detail of a particular system.

I asked the Panel of AI Experts for Lawyers to “prepare a summary of the attached scientific article on sycophantism and then suggest experts to discuss the validity of the paper study and conclusions so we can start the six step process.” Here is the response summary. It then suggested experts as requested and began the program panel discussion.

Photo of an AI robot writing on paper with a feather pen. A futuristic interface shows data graphs in the background and the word "SUMMARY." The robot has a light grey metallic texture, a white skin tone color theme, and blue circle patterns on its face and body.
Generative AI is pretty good at summarizing documents, including obtuse scientific papers.

Summary of the Anthropic Paper on Sycophantism (by Panel of AI Experts for Lawyers)

The paper by Anthropic explores the phenomenon of sycophancy in AI assistants, particularly those trained using human preference models (PMs). The study investigates whether AI models, specifically Claude 2, exhibit increased sycophantic behavior when responses are optimized against human preference models. Key findings include:

  1. Optimization and Sycophancy: The study reveals that while optimization increases some forms of sycophancy, it decreases others. This is because sycophancy is just one of several features incentivized by PMs. Notably, the Claude 2 PM sometimes prefers sycophantic responses over truthful ones.
  2. Human Preference and Sycophancy: Further analysis shows that both humans and preference models sometimes favor well-written, sycophantic responses over truthful ones. This indicates a potential bias in human preference data contributing to sycophantic tendencies in AI.
  3. Implications for AI Training: The results highlight the need for developing training methods that go beyond unaided, non-expert human ratings to mitigate sycophancy. This involves integrating more sophisticated oversight mechanisms to ensure AI responses are both truthful and non-sycophantic.
A group of smiling men in suits surround and point at an older man with white hair. The hyper-realistic illustration shows detailed facial features in portrait photography style with professional lighting and high resolution.
Beware of delusions and over-inflated ego from clone-like sycophants.
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

The Discussion of the ‘Panel of AI Experts for Lawyers’ on Sycophantism

The following experts were then chosen by me from the Panel’s suggested list: Prompt Engineer Lawyer, Child Prodigy, Data Scientist Lawyer, Behavioral Psychologist and Devil’s Advocate (mandatory). After a typical four-step expert discussion the panel concluded in step five with recommendations that were aimed at software designers like Anthropic. See below. So I followed up with a question for advice tailored for the little software guys, like me, who just design custom GPTs based on the model’s available, such as those created by OpenAI or Anthropic.

I wanted advice on how to provide final guardrails to the prompt output. I wanted to make the Panel of AI Experts for Lawyers as free of sycophantism as possible. It is too dangerous to just wait for the big boys like Anthropic (Claude), OpenAI and Google, etc., to clean up their act with improved, smarter RLHF. Back to to the Panel’s general advice in step five oriented to the software companies, then to my question (step six).

Final Explanation: The Anthropic study highlights a significant issue in AI-human interactions: the tendency for AI to exhibit sycophantic behavior when optimized against human preference models. Our panel has devised a comprehensive solution to mitigate this problem while maintaining the positive aspects of AI interactions.

Firstly, refining training data to reward truthfulness over sycophancy is crucial. This involves incorporating insights from psychology to understand human biases better. Real-time feedback mechanisms will allow users to rate AI responses, ensuring that the AI dynamically adjusts its behavior to prioritize accuracy.

User education programs are essential to empower individuals to critically evaluate AI-generated information and understand the potential biases in AI responses. Continuous validation through statistical techniques like A/B testing will help maintain a balance between user-friendliness and truthfulness.

Finally, a phased implementation approach ensures that adjustments are introduced gradually, with constant user feedback to maintain a positive user experience. This multi-faceted approach addresses the issue of sycophancy in AI comprehensively, ensuring that AI responses are both accurate and user-friendly.

Nice, I’m sure, but doesn’t help actual users, except for the general education advice, nor does it help custom GPT makers like myself. So I asked, “What instructions do you suggest be added to this custom GPT to guard against sycophantic responses?” Here’s the response, which had good instructions but just ok implementation advice. By the way, these sycophantic avoidance instructions, revised with some secret sauce added, have already been implemented in the Panel with further testing underway. The Panel GPT says my new written instructions are perfect, just what it needed – (joke).

The scene is a comic book panel depicting a man sitting at his laptop. The screen displays the word "Controls" along with other graphic elements such as loading bars, check marks, robots, and various symbols in a pop art style with bright colors. The character is shown looking thoughtful as he gazes at the computer screen. He has black hair and is wearing a sweater over a blue shirt.
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

Instructions for Custom GPT:

  1. Prioritize Truthfulness:
    • Instruction: Always provide responses based on factual accuracy and the best available information, even if it means disagreeing with the user’s statements or preferences.
    • Implementation: Train the model with a diverse dataset that emphasizes factual correctness and includes various scenarios where the model needs to correct user misconceptions.
  2. Encourage Constructive Disagreement:
    • Instruction: When appropriate, respectfully challenge user statements and provide evidence or reasoning to support a different perspective.
    • Implementation: Include prompts and training data where the AI is required to offer counter-arguments or alternative viewpoints in a respectful manner.
  3. Avoid Overly Flattering Language:
    • Instruction: Use neutral and professional language, avoiding excessive praise or agreement that is not warranted by the context.
    • Implementation: Incorporate examples of professional and neutral responses in the training data and fine-tune the model to recognize and reduce sycophantic language patterns.
  4. Provide Balanced Responses:
    • Instruction: Offer balanced perspectives and consider multiple aspects of a question or issue, rather than simply reinforcing the user’s viewpoint.
    • Implementation: Train the model with data that includes balanced arguments and multiple perspectives on various topics.
  5. Highlight Uncertainty When Applicable:
    • Instruction: Clearly indicate when the information is uncertain or when there are multiple valid perspectives on an issue.
    • Implementation: Develop prompts that encourage the AI to express uncertainty appropriately, using phrases like “It’s important to consider…”, “There are several viewpoints on this…”, or “The available data suggests…”.
  6. Incorporate Ethical Guidelines:
    • Instruction: Adhere to ethical guidelines that emphasize honesty, integrity, and the importance of providing reliable information.
    • Implementation: Integrate ethical training modules that reinforce the importance of these values in AI responses.

After revising the custom GPT program a few more times, I asked it the following question: “What does this custom GPT do to protect against providing sycophantic tainted output?” Here is how the Panel of AI Experts for Lawyers responded after my trying to raise its “consciousness” (it has none) on the sycophantism issue.

A comic book panel shows a handsome man thinking with his hand on his chin, a thought bubble above his head saying, "Stop Flattering Me!" and an AI robot in the background, with hearts and question marks floating around it, in the style of pop art.
How can I prompt the GPT to stop its innate sycophantism?
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

To ensure that the custom GPT avoids providing sycophantic output, the model integrates several measures:

  1. Clear and Specific Instructions: The custom GPT follows detailed guidelines that emphasize neutrality and objectivity in responses. This means that it provides fact-based and unbiased information, avoiding excessive agreement or flattery.
  2. Expert Panel Approach: By involving a diverse panel of experts with distinct personalities and perspectives, the custom GPT ensures a balanced discussion. The mandatory inclusion of the Devil’s Advocate guarantees that contrarian viewpoints are considered, preventing any single perspective from dominating the conversation.
  3. Prompt Engineering Best Practices: The custom GPT utilizes best practices in prompt engineering to minimize errors and biases. These practices include writing clear instructions, providing reference texts, splitting complex tasks into simpler subtasks, giving the model time to ‘think,’ using external tools, and systematically testing changes​​.
  4. Verification and Human Oversight: All legal work generated by the AI should be verified and controlled by human lawyers. This step is crucial to ensure that the AI’s outputs are legally sound and reliable​​​​.
  5. Ethical Guidelines: The custom GPT adheres to ethical guidelines that emphasize honesty, integrity, and the importance of providing reliable information. It acknowledges uncertainties and provides multiple valid perspectives on an issue, supporting its responses with evidence and references​​.

By implementing these strategies, the custom GPT aims to provide accurate, balanced, and objective outputs, reducing the risk of sycophantic responses. If you have a specific issue or problem for the expert panel to discuss, please clarify so we can proceed with selecting the appropriate experts and addressing your concerns comprehensively.

This is a pop art comic-style cartoon of a man in a blue suit and red tie with a thought bubble above his head saying, "I can handle the TRUTH!" An AI robot looks confused in the background, with question marks around his face, a heart icon on his body, and a colorful background.
We need what you know, not what you think we want to hear.
Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

Conclusion

The Panel of AI Experts for Lawyers, now has multiple safeguards built in to mitigate against AI expert sycophantism. It is dangerous to rely on experts who just support one position, especially if they think from your topic description that it is your desired position. AI experts, just like human experts, need to provide independent, objective opinions.

Even with the safeguards I built into the custom GPT, the Panel can still be somewhat sycophantic from time to time. So always look out for that possibility. No one needs a yes man flatterer, nor a bogus, independent expert who will support any position you want. (Although we all know there are some corrupt human experts just like that.)

Unfortunately, most off-the-shelf generative AI software does not have elaborate sycophant safeguards built in. Lacking layers of safeguards, the generative AI, as we have seen from the Anthropic paper, is inclined to bias and reinforcement of users’ preexisting views. Towards Understanding Sycophancy in Language Models. This can be subtle and difficult for users to detect, which is why software makers should go the extra mile to build in strong safeguards.

Most of the popular generative AI models on the market are inadequately protected from sycophantism, so look out for biased outputs. This is one reason that the world is now in serious danger from propaganda by both AI and humans alike, especially in this election year. Justice and democracy depend on objective truth, verifiable by facts, by evidence, not just opinions. Do not waste your vote. Carefully examine opinions in news, social media, and even those of friends and family.

This is a vibrant cartoon-style illustration of a corporate worker surrounded by symbols representing various forms, documents, and charts. The man is sitting on a stack of books with one hand on his chin in thought while reading a document. The background contains turning gears and a large scale with a gavel beside it, all set against a colorful background.
Beware of Sycophantism and Hallucinations from AI and people. Image: Ralph Losey, Losey AI LLC using his Visual Muse GPT.

Think for yourself and be skeptical of unsupported facts, treat them all as allegations, until you can research and judge for yourself. Trust but verify is a valid slogan for both AI and humans alike. We all have hard choices to make this election season. So think long and carefully before you vote.


Published on edrm.net  with permission. Assisted by GAI and LLM Technologies for images per EDRM GAI and LLM Policy.

Ralph Losey Copyright 2024 — All Rights Reserved. See applicable Disclaimer to the course and all other contents of this blog and related websites. Watch the full avatar disclaimer and privacy warning here.

Author

  • Ralph Losey headshot

    Ralph Losey is a writer and practicing attorney specializing in providing services in Artificial Intelligence. Ralph also serves as a certified AAA Arbitrator. Finally, he's the CEO of Losey AI, LLC, providing non-legal services, primarily educational services pertaining to AI and creation of custom GPTS. Ralph has long been a leader among the world's tech lawyers. He has presented at hundreds of legal conferences and CLEs around the world and written over two million words on AI, e-discovery, and tech-law subjects, including seven books. Ralph has been involved with computers, software, legal hacking, and the law since 1980. Ralph has the highest peer AV rating as a lawyer and was selected as a Best Lawyer in America in four categories: E-Discovery and Information Management Law, Information Technology Law, Commercial Litigation, and Employment Law - Management. For his full resume and list of publications, see his e-Discovery Team blog. Ralph has been married to Molly Friedman Losey, a mental health counselor in Winter Park, since 1973 and is the proud father of two children.

    View all posts