The Prevailing View
The consensus in enterprise AI is that multiple-choice tests reliably evaluate model alignment. Chief information security officers routinely sign off on large language model deployments based on standard AI safety benchmarks. The prevailing view holds that if a model accurately identifies the safest response from a predefined list, it is safe for production. However, a new research paper, Forcing a Letter Suppresses Refusal: Mechanistic Evidence from WMDP, shatters this assumption.
Our Position The AI industry’s reliance on multiple-choice safety benchmarks is fundamentally flawed. Testing environments must urgently pivot to dynamic, open-ended adversarial probing, or enterprise risk assessments will remain dangerously inaccurate.
What the Data Actually Shows
The research provides mechanistic evidence that the multiple-choice formats used in standard safety evaluations actively bypass a model’s built-in refusal mechanisms. By forcing a model to predict a specific letter token (such as A, B, C, or D) on the WMDP benchmark, the evaluation inadvertently suppresses the internal safety direction designed to prevent harmful outputs.
When evaluated through these constrained AI safety benchmarks, the model is not demonstrating its natural behaviour or its ability to refuse dangerous requests. Instead, the test format itself overrides the model’s guardrails. Because leading safety scores rely heavily on these multiple-choice evaluations, current metrics severely underestimate actual model capabilities. Enterprises are judging a system’s resilience by forcing it into a format that disables its primary defence.
The Real Implication
For organisations integrating AI into critical workflows, this finding dismantles the foundation of current AI risk management. If a vendor’s safety guarantee relies on multiple-choice evaluations, the resulting compliance score is an illusion.
This is a direct security liability. As regulatory frameworks demand rigorous proof of model safety, relying on tests that artificially suppress refusal exposes enterprises to immense risk. Security teams must recognise that model alignment cannot be accurately graded through static questionnaires. When an employee or an autonomous agent interacts with a model in production, the prompts are complex and open-ended. A safety architecture validated only by multiple-choice tests will fail when faced with real-world conversational manipulation.
Aaron Ranson, Chief AI Officer: «We are building bank vaults and testing them with multiple-choice quizzes. Until we subject models to continuous, open-ended probing, enterprise safety metrics are just theatre masquerading as governance.»
What to Do Instead
Enterprise leaders must immediately stop treating static safety scores as absolute proof of alignment. Governance teams need to mandate dynamic, open-ended adversarial testing for any model entering a production environment. This involves shifting from automated multiple-choice suites to continuous, contextual probing that mimics how employees actually interact with the system.
Balancing LLM evals versus human review is no longer a theoretical debate; organisations must build evaluation pipelines that combine automated adversarial generation with expert human oversight. We build the AI systems companies actually run, and our approach to EU AI Governance anchors on this reality. Compliance and security demand rigorous, real-world testing, not reliance on flawed assumptions.