Variation 23: How I Leveraged the “logit_bias” Parameter in the API to Combat Em Dashes—Suppressing 106 Tokens in the Process. Insights and Code for Your Own “No Dash” Response Test

How to Suppress Em Dashes in AI-Generated Text Using OpenAI API’s Logit Bias Parameter

Dealing with unwanted em dashes in AI-generated responses can be a persistent challenge. Despite employing various strategies like custom instructions and memory tweaks, many users find that models like ChatGPT stubbornly incorporate em dashes into their outputs. However, there’s an effective workaround that you might not have considered: utilizing the logit_bias parameter in the OpenAI API.

Understanding the logit_bias Parameter

The logit_bias feature allows you to assign a bias score between -100 and 100 to specific token IDs, effectively discouraging or encouraging their selection during text generation. My goal was to suppress em dashes (—) entirely, but I quickly realized that simply finding the token ID for an em dash isn’t sufficient. Because tokens can combine with other characters—like spaces or words—multiple variants of the dash might emerge.

A systematic approach was required, leading me to experiment with biasing various tokens that pertain to dash-like characters:

• Initially targeting tokens that directly include the em dash.
• Expanding to tokens that involve neighboring characters, such as spaces.
• Monitoring for substitutions like en dashes (–) or hyphens (-).
• Applying bias of -100 to tokens that contribute to em dash, en dash, and hyphen behaviors.

Through this iterative process, I found that setting biases across 106 different tokens was necessary to effectively prevent em dashes and related dashes from appearing—without notably damaging response quality.

The Process and Results

Here’s an outline of the progression:

• Starting with 3 tokens: '—', ' —', '— '.
• Increasing to 40 tokens that include any variant featuring '—'.
• Expanding to 62 tokens after encountering en dashes (–).
• Final suppression involved 106 tokens, including hyphens that are not connected to adjacent letters.

This brute-force biasing successfully diverted the model’s tendency to insert em dashes, even in models like GPT-4.0-turbo or ChatGPT with memory disabled.

Sample Evaluation

To verify the effect, I used a standard prompt asking for a “hot take” and compared responses:

• Normal Prompt Output: Maintains typical usage, including em dashes.
• Bias-Modified Output: Produces a more straightforward response, with significantly reduced or no em d