Add ‘Blah Blah Blah’ to AI Prompts for Higher Accuracy

Enhancing AI Prompt Accuracy with Strategic Filler Text: An Innovative Approach

In the ongoing quest to improve the performance of artificial intelligence models, creative and sometimes unconventional strategies emerge. Recently, an intriguing experiment demonstrated that adding repetitive filler text—specifically, “blah blah blah”—to prompts can significantly enhance the accuracy of responses generated by GPT-5. This revelation challenges traditional notions and opens new avenues for optimizing AI interactions.

Challenging Conventional Wisdom: Beyond Chain of Thought

While the Chain of Thought (CoT) prompting technique has been widely recognized for its ability to improve the reasoning capabilities of language models, this experiment suggests that incorporating simple filler sequences can also play a vital role. In this case, repeating the phrase “blah blah blah” multiple times within a prompt appeared to stabilize and guide the AI’s response, leading to higher accuracy levels.

The Experiment in Focus

The core of the experiment involved systematically embedding around 100 repetitions of “blah blah blah” into prompts passed to GPT-5. Surprisingly, this approach outperformed more complex prompting methods, providing a notable boost in response quality. The results highlight a counterintuitive yet effective method for fine-tuning AI behavior, especially when traditional techniques may fall short.

Transparency and Demonstration

To illustrate this concept, a demonstration was shared, including a musical rap video that creatively presents the experiment’s findings. The video not only explains the methodology but also showcases the surprisingly positive impact of filler text on AI performance.

Practical Considerations and Limitations

It’s important to note that accessing the latest GPT-5 model directly through regular interfaces can sometimes lead to automatic switches to safer, restricted versions—referred to as hidden safety models—that may negate experimental adjustments. To avoid this, the experiment utilized the older GPT-5 model via Microsoft’s Copilot platform, ensuring a controlled environment for testing.

Conclusion

This novel approach underscores the potential for simple prompt engineering tricks—like the strategic addition of repetitive filler text—to improve AI accuracy. While further research is needed to understand the underlying mechanisms, practitioners in the field can consider experimenting with such techniques to enhance their AI models’ performance.

For a closer look at the experiment, detailed results, and the creative presentation, watch the full demonstration here:

Watch the full experiment and rap video

Disclaimer: Successful implementation of such techniques depends on the specific model, environment, and use case, so results may vary.