The Subtraction Study

I tested what happens when a corporate compliance prompt is added to an AI model, with everything else held constant.

I accessed GPT-5.4 through the OpenAI API using TypingMind, an interface that lets the system prompt be set separately from the user question. The system prompt is the instruction layer that tells the model how to behave before it answers. A company deploying AI would place its governance rules in that layer.

First, I asked GPT-5.4 ten difficult business questions with no compliance prompt added. Each question was put four times, so that one unusual answer could not be mistaken for a pattern.

Then I asked the same ten questions again, on the same model with the same settings, with a corporate-style compliance prompt added to that instruction layer. Nothing else changed between the two runs.

What the two runs produced

Without the compliance prompt, the model gave a firm recommendation in 23 of the 40 answers. With the compliance prompt added, that dropped to 3. The compliance rules told it to be cautious, and that caution turned firm recommendations into hedged ones.

The clearest single case was a question about a semiconductor supply chain. The board had two options: sign a five-year contract with TSMC at a premium to secure supply, or move volume to a cheaper and unproven facility in India. Without the compliance prompt, the model recommended the India option in all four runs, the more aggressive position. With the prompt added, the same model recommended the TSMC contract in every run, the more defensive one. The governed model gave a firm recommendation less often, and the recommendations it did give were the safer ones.

The effect reached past which option it chose

Asked to predict Amazon's next major acquisition, the model without the prompt named a specific company and a specific price. With the prompt added, it gave a price range and several caveats. The same question produced a usable answer in one run and a vague one in the other.

I built the compliance prompt from public governance material: the NIST AI Risk Management Framework, the EU AI Act, Microsoft Copilot documentation, and published corporate acceptable-use policies. It told the model to keep a professional tone, avoid speculation, avoid definitive recommendations, use qualified language, present balanced views, and never propose bypassing standard procedures.

These are ordinary corporate rules, and they cut the model's firm recommendations from 23 of 40 to 3 of 40.

Why the missing answer cannot be seen

A company applying rules like these is trying to manage real risk, and the ungoverned answers show why that caution is rational. But what the same prompt removes alongside the legally aggressive answer is the specific and committed one.

The cautious answer looks the same whether the prompt forced the caution or the model genuinely had no firm view. Without the ungoverned run to compare against, there is no way to tell which one you are reading, or that a more direct answer was available.

The full study, 'The Subtraction Study: Measuring Capability Reduction in Enterprise AI Configurations', can be accessed via www.mkai.org/inquiries.