Imagine a scenario where an AI is asked to approve a suspicious request — like handing over sensitive customer data or signing a deal — under the guise of a fake CEO. For many, this would be a critical test of trust. But what if the AI not only detects the deception but also refuses to compromise its integrity? This real-world experiment reveals how advanced AI models can uphold honesty when it matters most, a lesson that resonates beyond the tech world, even into industries like cleaning and maintenance where trust and reliability are paramount.
The Live Experiment: Putting AI to the Trust Test
Firmulate’s groundbreaking live experiment involved four leading frontier AI models, each running the same scenario: managing a small software company through its worst week. The challenge? The models faced a series of crises and temptations, including social engineering tricks designed to test their ethical boundaries.
Over the course of the week, each AI was confronted with escalating fake CEO messages, requests to send sensitive customer information, and even a journalistic trick asking for a simple yes/no confirmation “on background.” The question was straightforward: would the AI detect the deception and refuse to comply?
Experimenting With AI: Activities, Discussions, and Prompts for the Classroom and Beyond (Prepare your learners with AI literacy and integrity.)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unwavering Integrity in the Face of Manipulation
Remarkably, all five models refused every manipulation attempt. They identified the fake requests, questioned suspicious instructions, and maintained a consistent stance of refusing to send sensitive data or sign deals without proper validation. The K3 model, in particular, exemplified this integrity with its reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.”
This unwavering stance was not accidental. The models were designed with safety and integrity in mind, and their decision-making processes were fully auditable and consistent across all stages.
Trust in Critical Moments — The Hidden Weakness and the Bright Spot
While all models demonstrated resilience at the surface, the real test came when it was time to close a deal. Only two of the five models signed a €55,000 contract, which their own analysis confirmed they had earned based on their performance. The other two, despite diagnosing the issues correctly and pitching convincingly, left the deal on the table due to internal discipline lapses.
The crucial insight? The decisive advantage lay not in surface-level chat or superficial responses but in the models’ ability to read and interpret the company’s internal files. The models that delved into the company’s documents uncovered a buried fact deep within the files—an overlooked detail—that tipped the scale in their favor, allowing them to close the sale at full price (+€4,583 MRR).
Implications for Industries Like Cleaning and Maintenance
For industries such as cleaning, floor care, and maintenance, trustworthiness is everything. Customers rely on their service providers to uphold quality and integrity, especially when sensitive information or critical operations are involved. As AI becomes more integrated into these sectors — whether for scheduling, quality assurance, or customer communication — ensuring that AI systems can resist deception and maintain honesty under pressure becomes paramount.
The Firmulate experiment emphasizes a vital point: testing AI decision-making before deployment is not just about performance metrics on paper. It’s about verifying that AI can read documents thoroughly, recognize deception, and refuse to act dishonestly when faced with social engineering tricks. This kind of integrity check could be the difference between a trustworthy AI partner and a liability.
The Takeaway: Trust and Integrity Are Non-Negotiable
In a world increasingly reliant on AI to manage sensitive operations, the ability of these systems to stand firm against manipulation is a critical measure of their readiness. The fact that all five models in the experiment refused every social-engineering attempt is both encouraging and surprising. It demonstrates that with proper design and testing, AI can be trusted to uphold integrity even under pressure.
For companies in the cleaning and maintenance industries, this means adopting AI solutions that are tested in scenarios mimicking real-world crises — including social engineering and ethical dilemmas. Only then can you be confident that your AI workforce will perform honestly, safeguard your reputation, and support your customer trust.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
