The Bluffing Machine
Why AI Hallucinates and Why RevOps Discipline Matters More Than Ever
AI is rewriting how go-to-market (GTM) teams operate. From SDR automation to RevOps forecasting, large language models (LLMs) now sit in the revenue engine. However, there is a catch: they still hallucinate.
Hallucinations are not bugs in the matrix. They are statistical inevitabilities. A recent paper from OpenAI researchers Adam Kalai, Ofir Nachum, Edwin Zhang, and Santosh Vempala explains why:
- Pretraining error is unavoidable. Even with flawless training data, the math behind cross-entropy loss guarantees some errors. Models especially fail on infrequent facts (like a CEO’s birthday or a niche customer’s contract terms) (Kalai, Nachum, Vempala, & Zhang, 2025).
- Post-training makes it worse. Because benchmarks and leaderboards overwhelmingly use binary scoring (right/wrong), models learn that guessing confidently is better than saying “I do not know.” Like students on multiple-choice exams, bluffing beats blank answers (Kalai et al., 2025).
- Put simply: LLMs are optimized to sound right, not to be right.
Why This Matters for GTM & RevOps Leaders
Most RevOps and GTM teams are evaluating AI SDRs, forecasting assistants, or customer-facing copilots. The risk is not that these systems go off-script; they produce confident-sounding but false data points inside your revenue machine.
- Pipeline Forecasting: A hallucinated win probability or account signal can cascade into misallocated spend, false board narratives, or investor misalignment.
- Customer Outreach: AI SDRs that “guess” details can burn trust in seconds. One wrong industry fact in an outbound message is all it takes to tank credibility.
- Revenue Data Integrity: CRMs already struggle with insufficient data. Feeding them hallucinations compounds noise instead of reducing it.
At Mir Meridian, we have seen firsthand how bad assumptions in GTM strategy derail growth. Hallucinations are the AI version of this problem: attractive and confident—but wrong.
How Leaders Should Respond
- Redesign Metrics to Reward Truth, Not Bluffing.
Just as the researchers propose new scoring systems that credit uncertainty (Kalai et al., 2025), RevOps teams must reframe internal metrics; an SDR bot that flags “uncertain, needs human validation” is more valuable than one that fills the CRM with confident fiction.
- Operationalize Behavioral Calibration.
The future does not ask for a percentage confidence score—it is building GTM systems where the AI only acts when it is 75%+ certain and otherwise escalates to a human. That is how you protect pipeline fidelity.
- Build Feedback Loops That Learn.
Kalai and colleagues argue that hallucinations persist because test incentives reinforce them (2025). In GTM, you can’t just plug in AI and forget it. Every outbound touchpoint, forecast miss, and customer interaction must feed back into the system to refine—not just scale—the motion.
The Bigger Picture
Hallucinations remind us of a timeless GTM lesson: motion ≠ progress (Mir Meridian, 2025). More emails, dashboards, and AI-generated insight don’t equal growth; alignment, trust, and execution do.
AI will be a core part of revenue engines—but only if we design systems that reward truthful uncertainty instead of confident guessing.
At Mir Meridian, we believe the winners will not be the companies that adopt AI fastest, but the ones that operationalize it wisely—embedding guardrails, calibration, and RevOps discipline so that AI amplifies reality, not illusion.
References