TZUTC: -0500   
   MSGID: 1623.consprcy@1:2320/105 2d60cda0   
   PID: Synchronet 3.21a-Linux master/123f2d28a Jul 12 2025 GCC 12.2.0   
   TID: SBBSecho 3.28-Linux master/123f2d28a Jul 12 2025 GCC 12.2.0   
   BBSID: CAPCITY2   
   CHRS: ASCII 1   
   FORMAT: flowed   
   Why yes-man AI could sink your business strategy  and how to stop it   
      
   Date:  Thu, 23 Oct 2025 14:27:03 +0000   
      
   Description:   
   Businesses risk flawed decisions as generalist AI hallucinates and   
   flattersspecialist models ensure accuracy.   
      
   FULL STORY   
   ======================================================================   
   Generative AI is quickly becoming a ubiquitous tool in modern business .   
   According to McKinsey, 78% of companies are now leveraging AIs ability to   
   automate and elevate productivity  up from 55% in 2024.    
      
   However, these systems arent without their flaws. Companies are becoming   
   increasingly aware of the issues associated with generalist large language   
   models, such as their eagerness to provide users with answers  even if they   
   arent factually correct.    
      
   Hallucinations are a well-documented challenge. Indeed, OpenAIs research   
   revealed that its own o3 and o4-mini models hallucinated 33% and 48% of the   
   time respectively when tested by the companys PersonQA benchmark  designed to   
   measure the ability of models to answer short, fact-seeking questions.    
      
   For organizations relying on generalist large language models to guide   
   decisions, their tendency to invent facts is a serious liability. Yet it is   
   not the only one. Equally, these mainstream models also present the issue of   
   sycophantic responses  when users perspectives are overly validated,   
   regardless of the truth.   
      
   How sycophancy can exacerbate yes-man AI    
      
   While there is a much greater spotlight on hallucinations, yes-man models    
   that wont advise users when they are wrong (and actually justify their   
   arguments with sycophantic responses) are in many ways more dangerous to   
   decision-making. When the default of an AI model is to agree, it can    
   reinforce biases and entrench incorrect assumptions.    
      
   Having rolled out (and quickly retracted) an update in April 2025 that made   
   its models noticeably more sycophantic, OpenAIs own researchers admitted that   
   people-pleasing responses can raise safety concerns around issues like mental   
   health, emotional over-reliance, or risky behavior.    
      
   Concerningly, a study by Anthropic researchers looking at the way in which   
   human feedback can encourage sycophantic behavior showed that AI assistants   
   may modify accurate answers when questioned by the user, and ultimately give   
   an inaccurate response.    
      
   Meanwhile, research has also shown that both humans and preference models   
   (PMs) prefer convincingly written sycophantic responses over correct ones a   
   non-negligible fraction of the time.    
      
   Thats a worrisome combination. Not only do  generalist large language models   
   sometimes alter correct answers to appease users, but people themselves often   
   prefer these agreeable, sycophantic responses over factual ones.    
      
   In effect, the generalist large language models are reinforcing users views   
   even when those views are wrong  creating a harmful loop in which validation   
   is valued above accuracy.   
      
   The issue of sycophancy in high stakes settings    
      
   In high-stakes business settings such as strategic planning, compliance, risk   
   management or dispute resolution, this presents a serious risk.    
      
   Looking at the latter example of dispute resolution, we can see how the    
   issues of sycophancy arent limited to factual correctness but also extend to   
   tone and affirmation.    
      
   Unlike in customer service  where a flattering, sycophantic answer may build   
   satisfaction  flattery is a structural liability in disputes. If a model   
   echoes a users sense of justification (i.e., youre right to feel that way),   
   then the AI may validate their perceived rightness, leading them to enter a   
   negotiation more aggressively.    
      
   In this sense, that affirmation can actively raise the stakes of   
   disagreements, with users taking the AIs validation as implicit endorsement,   
   hardening their positions and making compromise more difficult.    
      
   In other cases, models might validate both parties equally (i.e., you both   
   make strong points), which can create a false equivalence when one sides   
   position is actually weaker, harmful, or factually incorrect.   
      
   Greater segmentation and specialist AI are needed   
      
   The root of the problem lies in the purpose of generalized AI models like   
   ChatGPT. These systems are designed to be helpful, engaging in casual Q&A     
   not for the rigorous impartiality that applications like dispute resolution   
   demand. Their very architecture rewards agreement and smooth conversation,   
   rather than critical evaluation.    
      
   It is for this reason that strong segmentation is inevitable. While well   
   continue to see consumer-grade LLMs for casual use, organizations need to   
   adopt specialist AI models for more sensitive or business-critical functions   
   that are specifically engineered to avoid the pitfalls of hallucination and   
   sycophancy.    
      
   What success looks like for these specialist AI models will be defined by    
   very different metrics. In the case of dispute resolution, systems will be   
   rewarded not for making the user feel validated, but for moving the dispute   
   forward in a fair and balanced way.    
      
   In changing alignment from pleasing users to maintaining accuracy and    
   balance, specialist conflict resolution models can and should be trained to   
   acknowledge feelings without endorsing or validating positions (i.e., I hear   
   that this feels frustrating, rather than youre right to be frustrated).    
      
   As generative AI further cements its position at the forefront of business   
   strategy, these details are critical. In high-stakes functions, the potential   
   cost of a yes-man AI  one that flatters rather than challenges, or invents   
   rather than informs  is simply too great. When business leaders lean on   
   validation rather than facts, the risk of poor decisions increases   
   dramatically.    
      
   For organizations, the path forward is a clear one. Embrace specialist,   
   domain-trained models that are built to guide, not gratify. Only specialist    
   AI models grounded in factual objectivity can help businesses to overcome   
   complex challenges rather than further complicate them, acting as trusted   
   assets in high stakes use cases.    
      
    This article was produced as part of TechRadarPro's Expert Insights channel   
   where we feature the best and brightest minds in the technology industry   
   today. The views expressed here are those of the author and are not   
   necessarily those of TechRadarPro or Future plc. If you are interested in   
   contributing find out more here:   
   https://www.techradar.com/news/submit-your-story-to-techradar-pro   
   ======================================================================   
   Link to news story:   
   https://www.techradar.com/pro/why-yes-man-ai-could-sink-your-business-strategy   
   -and-how-to-stop-it   
   $$   
   --- SBBSecho 3.28-Linux   
    * Origin: capitolcityonline.net * Telnet/SSH:2022/HTTP (1:2320/105)   
   SEEN-BY: 105/81 106/201 128/187 129/14 305 153/7715 154/110 218/700   
   SEEN-BY: 226/30 227/114 229/110 111 206 300 307 317 400 426 428 470   
   SEEN-BY: 229/664 700 705 266/512 291/111 320/219 322/757 342/200 396/45   
   SEEN-BY: 460/58 633/280 712/848 902/26 2320/0 105 304 3634/12 5075/35   
   PATH: 2320/105 229/426