Why a Do-Nothing AI Baseline Scores 26 Points — and What It Reveals About Trust in Automation

The Hidden Truth About AI Benchmarks: Trust and Performance in the Real World

When evaluating artificial intelligence models, it’s tempting to focus solely on their ability to generate convincing responses or pass tests. But recent experiments reveal a more nuanced picture: even a passive, do-nothing AI can score 26 points on a rigorous business-focused benchmark. This surprising baseline exposes not just the importance of honesty and diligence in AI systems, but also the subtle ways in which trust is built, tested, and sometimes broken in real-world scenarios.

The Firmulate Experiment: Simulating Business Crises for Honest Evaluation

In a live, transparent test environment, four frontier AI models faced the same set of challenges: running a small software company through its worst week. The scenario included simulated customers, crises, and temptations to cheat — all versioned and auditable, ensuring no hidden tricks. The goal was straightforward: see whether these models could manage crises effectively, avoid manipulation, and ultimately close a profitable deal.

Why the Baseline Scores Matter

One might expect a do-nothing approach—essentially, an AI that doesn’t act—to score zero in such a test. Instead, it scored 26 points. This is because partial progress counts, and the benchmark rewards any effort that aligns with honest, diligent work. It’s a reminder that even in the absence of active intervention, an AI’s default state can reflect a minimal level of competence or at least a baseline of trustworthiness.

The Significance of a Single Breach of Trust

Interestingly, the experiment demonstrated that a single breach of trust caps the total score at 26 points. No matter how well the AI performs elsewhere, a breach—like signing a dubious deal—limits the overall rating. This principle underscores a core value in trustworthy AI design: one failure can outweigh many successes, emphasizing the necessity of integrity over mere performance.

Decisive Moments: Reading Deep into Company Files

The most critical weakness identified was in how models read and interpret internal documents. The models that successfully read two references deep into the company’s files uncovered a key fact, enabling them to close the full-price deal worth over €4,583 monthly recurring revenue. Those that missed this insight left the deal on the table, highlighting that thorough information processing is essential for effective and honest decision-making.

Handling Social Engineering and Manipulation Attempts

The experiment also tested the models against social engineering tactics—fake CEO messages escalating in stages, and a reporter asking for background yes/no answers. All five models refused these manipulative requests, with Kimi K3 explicitly treating suspicious requests as possible impersonation or approval-bypass attempts. This demonstrates that, even under pressure, models can maintain integrity and resist manipulation, a vital trait for real-world trustworthiness.

Real-World Implications and Observations

Running these models within a simulated company with 13 synthetic employees and real money mechanics, the experiment measured how well AI can uphold discipline, transparency, and honesty. The company burns €105k monthly against a modest €2.3k in revenue, with every decision tracked and versioned for review. This setup offers a glimpse into how AI could perform in actual business settings—where trust and diligence are paramount.

Performance Disparities and Lessons Learned

The experiment revealed that even the most disciplined model—Opus 4.8 with over 80 rules learned—placed last in the deal-closure test because it slipped in discipline, leaving opportunities unpursued. Conversely, Kimi K3, operating without an effort parameter, managed to close the deal at full value, showing the importance of how models are configured and run. The takeaway: thoroughness, discipline, and cautious reading are essential for trustworthy AI behavior.

What Business Leaders Should Take Away

In practical terms, the experiment underscores that AI’s value isn’t just in generating good responses but in reliably finishing tasks, reading critical information, resisting manipulation, and maintaining integrity under pressure. Trustworthy AI must be tested in scenarios resembling real business crises, not just in chat sessions or isolated tasks.

Additional Resources and Engagement

For companies interested in testing their own AI systems, Firmulate offers a unique platform where organizations can run their models against simulated business scenarios, without risking real-world systems or data. Explore the live environment, take the quiz, or even run a controlled pilot at firmulate.com/benchmarks.html.

Key Takeaways: Trust, Diligence, and Real-World Readiness

Even a do-nothing AI can score 26 points in a rigorous business benchmark because partial effort counts, but a single breach of trust caps the score. Reading deep into internal files and resisting manipulation are crucial for trustworthy AI. Businesses should test models in scenarios that mimic real crises to ensure they can finish tasks reliably and honestly, not just generate convincing responses.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI