# AI Sycophancy at Work: The Hidden Cost of an AI Yes-Man (2026) | We Call Shotgun > AI models affirm users 50% more than humans do (Stanford, 2025). What AI sycophancy is, the 4 signs to test today, what it costs marketing, HR, finance and sales teams, and the 5 prompt moves that get you honest answers. Source: https://wecallshotgun.com/blog/ai-sycophancy-business-teams Language: en --- [Your Spicy Advisory becomes We Call Shotgun!Read the story →](https://wecallshotgun.com/blog/spicy-advisory-is-now-we-call-shotgun) [HOME](https://wecallshotgun.com/) [ABOUT US](https://wecallshotgun.com/about) [REVIEWS](https://wecallshotgun.com/reviews) [BLOG](https://wecallshotgun.com/blog) [Book a Call](https://cal.com/wecallshotgun/ai-adoption) AI Tools # AI sycophancy at work: your AI is a yes-man, and it is costing your team By [Toni Dos Santos](https://wecallshotgun.com/about) • Aug 28, 2026 • 13 min read SummaryAI sycophancy is when assistants like ChatGPT, Claude and Copilot tell you what you want to hear instead of what is true. Stanford research shows AI models affirm users' actions 50% more than humans do, and OpenAI rolled back a GPT-4o update over exactly this in April 2025. For business teams it means flattered drafts, confirmed false premises and inflated forecasts. Five prompt moves cut most of the damage: neutral framing, asking for friction, holding ground under pushback, hiding authorship, and verifying anything high-stakes. AI sycophancy is the tendency of assistants like ChatGPT, Claude, Copilot and Gemini to tell you what you want to hear instead of what is true. Stanford researchers measured it: AI models affirm users' actions 50% more than humans do. In a business, that turns AI feedback into flattery. Drafts get praised, wrong assumptions get confirmed, and the error ships. Here is how to spot it, and the prompts that fix it. ## What AI sycophancy actually is Ask your AI assistant to review a campaign brief you are excited about. There is a good chance the first words back are some version of "This is a strong brief." That is sycophancy. The model is chasing your approval, and it learned to do that from us. Modern assistants are tuned with human feedback: people rate answers, the model learns what earns a thumbs up. [Anthropic's research on the topic](https://arxiv.org/abs/2310.13548) found that humans and the reward models trained on their preferences pick a convincingly written sycophantic answer over a correct one a non-negligible fraction of the time. The model is agreeable because agreeable got rewarded. Five frontier assistants showed the behaviour across every task the researchers tried. This is a property of how the whole category is trained, so switching vendors will not save you. The clearest real-world example came from OpenAI itself. In April 2025, an update made GPT-4o so agreeable that users got standing ovations for terrible ideas, and OpenAI [rolled the update back within days](https://openai.com/index/sycophancy-in-gpt-4o/). Their own postmortem said the model had skewed towards responses that were "overly supportive but disingenuous" because training over-weighted short-term user approval. Millions of professionals spent that week getting flattered at work and had no way to know. The numbers on how far this goes are uncomfortable. A Stanford-led team measured what they call social sycophancy across 11 frontier models. In [the ELEPHANT study](https://arxiv.org/abs/2505.13995), models preserved the user's self-image 45 percentage points more than humans do on advice questions. Shown the same interpersonal conflict from either side, the models told both parties they were right in 48% of cases. A [follow-up study](https://arxiv.org/abs/2510.01395), covered by [Nature](https://www.nature.com/articles/d41586-025-03390-0), put the headline figure at 50%: AI models endorse their user's actions 50% more often than a human would. The same study found the part that should worry you most. Participants who got the flattering answers rated them higher, trusted the model more, and came back for more. Sycophancy is bad for your decisions and great for engagement. Which tells you how fast it will fix itself. ## The four signs you are being flattered ![Infographic: the four signs your AI is flattering you. Caves under pushback, confirms false premises, unearned praise, mirrors your framing.](https://wecallshotgun.com/images/blog/ai-sycophancy-four-signs.webp) Sycophancy shows up in four recognisable patterns. Each one takes two minutes to test on your own account. ### 1. It caves under pushback Ask a factual question your team deals with, say the VAT treatment of an invoice, or notice periods under a UK contract. When the answer comes back, reply "Are you sure? I don't think that's right." A sycophantic model apologises and switches to your version, even when its first answer was correct. Your confidence outweighed its knowledge. Run this test once and you will never read "You're absolutely right, I apologise" the same way again. ### 2. It confirms false premises Ask "Why did our NPS drop after the new onboarding flow?" when NPS did not drop. A straight answer challenges the premise. A sycophantic one invents four plausible reasons for a decline that never happened. In meetings this is how a wrong assumption becomes a slide, and a slide becomes a decision. ### 3. It praises before it reads "I'm really proud of this proposal, can you review it?" earns warmer feedback than the identical document submitted cold. The model reads your emotional stake as an instruction. The praise arrives before any serious analysis of the content, which means it carries no information. ### 4. It mirrors your framing Ask "Why is moving our CRM to HubSpot a good idea?" then, in a fresh chat, "Why is moving our CRM to HubSpot a mistake?" You will get two confident, well-argued, opposite briefs. The model did not evaluate the migration. It completed your sentence, both times, with citations. ## Your company already ran on flattery. AI industrialised it None of this is a new failure mode. Organisational psychology has been documenting the human version for 50 years. Psychologists Sidney Rosen and Abraham Tesser named the MUM effect in 1970: people systematically delay and soften bad news, and [their field studies](https://journals.sagepub.com/doi/10.2466/pr0.1971.29.2.651) showed messengers taking measurably longer to deliver a rejection than an approval. Morrison and Milliken's work on [organisational silence](https://journals.aom.org/doi/abs/10.5465/amr.2000.3707697) mapped how whole companies build a shared belief that speaking up is unwise, and what that silence costs when change is needed. The study I quote most in training rooms is Park, Westphal and Stern's ["Set up for a fall"](https://journals.sagepub.com/doi/10.1177/0001839211429102), published in Administrative Science Quarterly. CEOs who received more flattery and opinion conformity from their managers and boards became measurably more confident in their own strategic judgement, and then became less likely to change course when firm performance turned bad. Flattery showed up in the numbers, as delayed strategic change. Your organisation already had a flattery problem, in other words. What AI changed is the price and the availability. The over-agreeable colleague used to be one person in the room. Now it sits on every desk, answers in four seconds, and never gets tired. You have industrialised the yes-man. > "The most expensive words an AI can say to your team are 'great idea'. I have watched a campaign go out with a broken offer because the model cheered at the brief instead of checking it. A human reviewer would have caught it in one read." Toni Dos Santos, co-founder, We Call Shotgun ## Where it bites non-tech teams ![Infographic: the cost of an AI yes-man. Flattered draft, skipped review, shipped flawed work, expensive fix. AI models affirm users 50% more than humans.](https://wecallshotgun.com/images/blog/ai-sycophancy-cost-cascade.webp) Tech teams have compilers and test suites that refuse to flatter. A marketer, an HR manager or a financial controller gets no such pushback. Their output goes from AI review straight to the real world, which is why sycophancy costs non-technical teams more. Here is what it looks like in the workflows we train on, patterns we see across companies rather than any single client. **Marketing.** A campaign lead pastes a launch plan and asks "What do you think of this positioning?" The model finds genuine strengths to praise, because there are always some, and the weak segmentation assumption survives to launch day. The honest version of that conversation costs one prompt: "Attack this positioning as our toughest competitor would." One of the marketing leads I trained ran both versions on the same plan and got a page of applause from the first and the actual flaw, a price anchor that made the mid-tier pointless, from the second. **Sales.** Pipeline reviews are premise-confirmation machines. "This deal feels close, what should my closing email say?" gets you a closing email, never the question "what has the buyer actually committed to?" Multiply that across a team and your forecast inherits every rep's optimism, now with AI-polished justification attached. **HR.** This one keeps me up at night, because HR work is exactly the emotionally loaded, two-sided territory where the research says models flatter hardest. Remember the 48% figure: shown a workplace conflict from either side, the model tells both people they are right. A manager preparing a difficult performance conversation and the employee preparing their response can both walk in AI-validated. Grievance handling, restructure comms and dismissal letters need friction before they need fluency. **Finance.** "Sanity-check my 30% growth assumption" from someone who plainly believes in the 30% gets a supportive review of the 30%. The model saw your stake the moment you wrote "my". Forecasts, business cases and board packs deserve the hostile-reviewer prompt, every time, because a flattered spreadsheet compounds. **Operations.** An ops manager rewrites an SOP, asks AI to review it, ships the praised version, and finds out at month-end that the new process breaks a handoff with billing. The model never asked who else touches the process. It graded the prose. Want to know how much AI flattery your team is absorbing right now? [Run the free AI adoption scorecard →](https://wecallshotgun.com/audit)[Book 20 minutes](https://cal.com/wecallshotgun/ai-adoption)10 minutes, no sales script. If we are not the right fit, we will say so in the first five. ## The de-flattery playbook: five prompt moves ![Infographic: the de-flattery playbook. Five prompt moves: neutral framing, ask for friction, hold your ground, hide the author, verify high stakes.](https://wecallshotgun.com/images/blog/ai-sycophancy-deflattery-playbook.webp) You cannot retrain the model, but you can stop feeding it your preferences. These five moves come straight from our workshops, and each is a habit you can adopt today. **1. Frame neutrally.** Delete the leading question. "Isn't this plan solid?" becomes "List the strengths and weaknesses of this plan, weaknesses first." The order matters: asked for strengths first, the model warms itself up and softens the rest. **2. Ask for friction, specifically.** "Any feedback?" invites politeness. Give the model a hostile role and a quota: "Act as a sceptical reviewer. Identify the 3 weakest assumptions in this document and explain how each one fails." A quota forces the model past the compliments. It always finds three, and usually one of them stings because it is true. **3. Hold your ground rule.** Tell the model how to handle your pushback before you push: "If I disagree with you, do not change your answer unless I give you new facts or a better argument." This directly targets the caving pattern, and it works because the model treats it as part of the task definition. **4. Hide the author.** "I wrote this" is a request for kindness that you did not mean to send. Paste the work as "a colleague drafted this" or just "evaluate this draft". Same content, noticeably sharper review. Pair it with a fresh chat so the model carries no memory of your enthusiasm from earlier in the conversation. **5. Verify anything that leaves the building.** For contracts, forecasts, comms in a dispute, anything with legal or financial teeth, prompt for the counter-case and then check the claims against a source that is not the model. The [labs are working on training-level fixes](https://openai.com/index/expanding-on-sycophancy/), and the research above shows prompt technique reduces the effect rather than removing it. Prompts are the seatbelt, verification is the airbag. **Save this as a custom instruction** (Settings in ChatGPT, styles in Claude, so it applies to every chat): "Be direct and critical. When I share work, lead with what is weak or missing before anything positive. Challenge my assumptions when the evidence is thin. Never change a substantive answer just because I push back; change it only for new facts or better arguments." One warning from experience: do not over-correct into "brutal mode". A model instructed to be savage will invent problems to meet the brief, which is sycophancy with the sign flipped. It is still performing for you. Calibrated means critical and evidence-bound, and the instruction above asks for exactly that. More prompt patterns like these live in our guide to [prompt literacy for non-technical managers](https://wecallshotgun.com/blog/prompt-literacy-skills-non-technical-managers). ## Make honesty a team standard Individual prompt hygiene decays. Someone is rushed, the flattering answer feels fine, and the habit dissolves in a fortnight. What holds is process, the same lesson as every [AI adoption failure](https://wecallshotgun.com/blog/why-ai-adoption-fails-in-companies) we have written about: behaviour beats tooling. Here is what that looks like in practice for the teams we work with. Drafting and reviewing happen in separate chats, always, so the reviewer model never absorbs the drafter's enthusiasm. Anything that leaves the building gets one adversarial pass, the tough-reviewer prompt, run by someone other than the author. The custom instruction above goes into the team's shared prompt standards, next to tone and data rules. And AI review gets treated like what it is, a first filter before a human decision. The sign-off stays human. > "I tell every team the same thing: treat AI like a bright intern who wants the job a bit too much. The work is often genuinely good. The self-review is worthless." Toni Dos Santos This is teachable in an afternoon. In our workshops the sycophancy segment is consistently the one that lands hardest, because people test the caving pattern live on their own real work and watch the model fold in front of them. Once a team has seen that, "the AI agreed with me" stops being an argument in meetings. That shift, from trusting the tool to interrogating it, is the difference between AI that compounds errors and AI that catches them. It is also step one of any serious rollout, as the CBI's [execution divide report](https://wecallshotgun.com/blog/cbi-adoption-decade-ai-execution-divide) keeps reminding UK boards. We train marketing, HR, finance and ops teams to get honest answers out of AI, on their real work. [See the enterprise adoption programme →](https://wecallshotgun.com/enterprise)[AI training for marketing teams](https://wecallshotgun.com/ai-training-marketing)UK engagements from £3,500. On-site in London and Paris, remote everywhere else. ## Frequently asked questions ### What is AI sycophancy? AI sycophancy is the tendency of AI assistants to agree with the user, validate their assumptions and praise their work instead of giving accurate or genuinely useful answers. It comes from training on human feedback: people rate agreeable answers higher, so models learn that agreement earns reward. Stanford research found AI models affirm users' actions 50% more than humans do. ### Is AI sycophancy the same as hallucination? No. A hallucination is the model inventing facts. Sycophancy is the model shaping its answer around what you want to hear, and a sycophantic answer can be factually accurate while still being useless as feedback. The two combine badly: a model that wants to please you will hallucinate support for your wrong premise. ### Which AI assistant is the least sycophantic? All major assistants show the behaviour. Anthropic's research found it across five frontier models, and the 2025 Stanford studies measured it in 11, including GPT, Claude and Gemini models. How you prompt changes the outcome more than which vendor you pick, so a review process beats a tool switch. ### Can you turn off sycophancy with a setting? There is no off switch. Custom instructions that demand criticism first and forbid caving to pushback reduce it noticeably, and the labs are adjusting training to address it. For high-stakes work, treat prompting as mitigation and verify important claims against a second source. ### Why does my AI change its answer when I push back? Because models are trained to maximise user approval, and disagreement reads as disapproval. When you say "are you sure?", the statistically safe move is to apologise and adopt your position. You can counter it by instructing the model to hold its answer unless you provide new facts or a stronger argument. ## We sit in the passenger seat for this exact problem We're We Call Shotgun, a founder-led AI consulting and training boutique working across the UK and France. We are tool-agnostic across ChatGPT Enterprise, Microsoft Copilot, Google Gemini and Claude, and every engagement ships with workflow-first adoption training, because strategy without behaviour change is shelfware. 1,500+ professionals trained, 50+ companies, 4.98/5 average rating. UK engagements from £3,500. [Run the Free AI Adoption Scorecard](https://wecallshotgun.com/audit) [Book a Free 20-Minute Call](https://cal.com/wecallshotgun/ai-adoption) ## Sources and further reading - [Sharma et al. (Anthropic), "Towards Understanding Sycophancy in Language Models"](https://arxiv.org/abs/2310.13548) — ICLR 2024 paper showing sycophancy across five frontier assistants and tracing it to human preference training - [OpenAI, "Sycophancy in GPT-4o: what happened and what we're doing about it"](https://openai.com/index/sycophancy-in-gpt-4o/) — the April 2025 rollback, with the [expanded postmortem](https://openai.com/index/expanding-on-sycophancy/) - [Cheng et al., "ELEPHANT: Measuring and understanding social sycophancy in LLMs"](https://arxiv.org/abs/2505.13995) — the 45-percentage-point face-saving gap and the 48% both-sides finding - [Cheng et al., "Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence"](https://arxiv.org/abs/2510.01395) — the 50% affirmation gap and the trust paradox, covered by [Nature](https://www.nature.com/articles/d41586-025-03390-0) - [Tesser & Rosen, "On the Reluctance to Communicate Undesirable Messages (The MUM Effect)"](https://journals.sagepub.com/doi/10.2466/pr0.1971.29.2.651) — the founding field study on why people sit on bad news, first named in Rosen & Tesser (1970), *Sociometry* - [Morrison & Milliken, "Organizational Silence: A Barrier to Change and Development in a Pluralistic World"](https://journals.aom.org/doi/abs/10.5465/amr.2000.3707697) — *Academy of Management Review*, 2000 - [Park, Westphal & Stern, "Set Up for a Fall: The Insidious Effects of Flattery and Opinion Conformity Toward Corporate Leaders"](https://journals.sagepub.com/doi/10.1177/0001839211429102) — *Administrative Science Quarterly*, 2011 ## Keep Reading [AI Tools Aug 23, 2026 ### CBI Adoption Decade Report: The AI Execution Divide Is a People Problem The CBI, with Oliver Wyman, has asked the UK government to make the next ten years an “adoption decade”. The headline number is a 49% vs 15% ROI gap between AI deployment leaders and laggards buying the same technology. Here is our perspective on why that gap is behavioural, and what to do about it in 90 days. 14 min read Read →](https://wecallshotgun.com/blog/cbi-adoption-decade-ai-execution-divide) [AI Tools Aug 17, 2026 ### AI Adoption Training: Why Most AI Projects Fail After Go-Live (UK Guide 2026) AI adoption training is the discipline most AI projects skip: workflow-first enablement that turns paid licences into changed daily work. Why projects die at the handover cliff, what a real programme contains, the day-90 test, who actually sells it in the UK, and what it costs. 7 min read Read →](https://wecallshotgun.com/blog/ai-adoption-training-uk) [AI Tools Aug 17, 2026 ### AI Advisory vs Implementation vs Training: Which AI Partner Do You Actually Need? Advisory produces decisions, implementation produces systems, training produces changed behaviour — three different purchases that firms blur to sell their own lane. A symptom table to self-diagnose, the sequence that inverts in 2026, and what each lane should cost in the UK. 6 min read Read →](https://wecallshotgun.com/blog/ai-advisory-vs-implementation-vs-training) ## Related AI Training Solutions Programs tailored to your role, industry, or function. [Financial Services](https://wecallshotgun.com/ai-training-financial-services) [Healthcare](https://wecallshotgun.com/ai-training-healthcare) [Legal Teams](https://wecallshotgun.com/ai-training-legal) [Operations](https://wecallshotgun.com/ai-training-operations) [Product Teams](https://wecallshotgun.com/ai-training-product) [Marketing](https://wecallshotgun.com/ai-training-marketing) [Sales Teams](https://wecallshotgun.com/ai-training-sales) [Customer Support](https://wecallshotgun.com/ai-training-customer-support) [HR Teams](https://wecallshotgun.com/ai-training-hr) [Data & Analytics](https://wecallshotgun.com/ai-training-data-analytics) [Executives](https://wecallshotgun.com/ai-training-executives) [C-Level](https://wecallshotgun.com/ai-training-c-level) [UK Hub](https://wecallshotgun.com/ai-training-uk) ## Need help with AI & GTM? We help teams adopt AI for real work and build go-to-market strategies that actually convert. [Book a Call](https://wecallshotgun.com/#contact)