Imagine trying to judge the quality of a new ice cream flavor: you might look at how it tastes, how smooth it is, or whether it melts perfectly. But what if the real test isn’t just flavor, but whether it can resist tampering or deception? In the world of AI, this question is just as critical. As companies consider deploying AI that manages customer data, finances, or even decision-making, the true measure isn’t just how clever the AI looks—it’s whether it can be trusted to do what it’s told, every time.
Get baking supplies delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Setting the Stage: Why Trust in AI Matters
When businesses evaluate AI models, the focus often lands on performance metrics like accuracy or speed. But in a recent public experiment conducted by Firmulate, a company specializing in AI benchmarking, a different dimension emerged: trustworthiness. This was demonstrated through a carefully designed scenario where AI models were tested under the same simulated crises—think of it as a week of bad reviews, angry customers, and tempting shortcuts.
AI trustworthiness benchmarking tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Benchmark: A Do-Nothing Baseline and Its Surprising Score
In this experiment, four leading AI models faced a simulated small software company experiencing its worst week. Each was tasked with managing crises, reading critical internal documents, and making strategic decisions. The models scored an average of 73 to 95 points out of 100, with the best, GPT-5.6-sol, topping the list at 95. But what’s truly striking is the baseline score—the “do-nothing” model, which basically refused to act or manipulate, scored 26 points. That’s not zero, as some might expect, but a reflection of partial progress. It also highlights a crucial principle: even minimal effort in AI decision-making is worth something, but it’s no substitute for honest, diligent behavior.
Understanding the Scoring: Trust and Partial Progress
The scoring system in the experiment is designed to reward honest, thorough work. For example, all models identified every crisis and refused manipulative tactics like fake CEO messages or attempts to bypass approval processes. Yet only two models successfully closed a key deal, worth €55,000 monthly recurring revenue, matching their own diagnoses and pitches. The others, despite good insights, failed to sign the deal due to slip-ups like leaving the close on the table or mishandling escalation procedures.
Why Partial Work Matters
The results emphasize an important point: partial progress counts. A model that reads the internal documents and refuses manipulative requests scores higher than one that ignores these signals altogether. Still, the experiment makes clear that trust isn’t just about identifying crises—it’s about reliably following through on commitments, even when tempted or pressured. This is a key principle in both AI and business: no matter how clever an agent is, a single breach of trust caps its overall score, because in real life, trust is absolute.
The Real-World Parallel: Managing a Company or Baking a Cake
For ice cream enthusiasts and bakers, the lesson is simple: consistency and integrity matter. You wouldn’t want a pastry chef who cheats on ingredients or a baker who cheats on time. Similarly, AI models need to demonstrate they can resist manipulation—be it fake messages or hidden documents—just as a good recipe resists shortcuts.
Firmulate’s Live Experiment: Testing AI in Action
Beyond the scores, the live experiment at Firmulate showcases an AI-driven simulated company with 13 synthetic employees, real money mechanics, and a public cash countdown. Every decision is versioned and auditable, providing a transparent view of AI behavior under pressure. Watchers can see the AI in action, making decisions that affect cash flow and customer outcomes, highlighting the importance of honesty and diligence in automation.
The Lessons for Business and AI Adoption
What does this mean for companies considering AI? First, it’s not enough for an AI to produce impressive chat or generate convincing reports. The real test is whether it can finish what it starts, read internal files thoroughly, and remain honest under pressure. The experiment underscores that current AI models can spot crises and refuse manipulation, but they also have weaknesses—like slipping in process discipline—that can be critical in real-world applications.
The Bottom Line: Trust Is the Benchmark
As firms look to adopt AI, the takeaway from this benchmark is clear: trustworthiness isn’t just a nice-to-have; it’s the foundation. Even a do-nothing baseline scores 26 out of 100, illustrating that minimal effort in honesty and thoroughness is better than nothing. However, to truly excel, AI models need to demonstrate consistent discipline, read critical information carefully, and resist shortcuts—just as a master baker resists substituting ingredients for speed.
In AI benchmarking, trustworthiness is the real currency. Even a do-nothing approach scores 26, emphasizing that partial progress matters, but honesty caps overall performance—making transparency and discipline essential for trustworthy AI deployment.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
