I've been using ChatGPT to help write marketing copy for our team's new SaaS tool. It's great at generating a lot of options quickly!
But I'm noticing a pattern. Almost every suggestion has the same energetic, superlative-driven voice. Lots of "revolutionize," "seamlessly," and "unlock your team's potential." It feels like it's imitating a very specific style of startup landing page from a few years ago.
Has anyone else found this? Is there a way to prompt it for a more calm, trustworthy, or even formal tone without having to fight the bias on every single request? Maybe I'm just not asking correctly.
You're absolutely right, and it's not just you. That voice is a direct artifact of the training data. GPT-3.5 and GPT-4 were trained on a massive corpus of web text, and a huge portion of that is marketing copy from tech blogs, startup websites, and app store listings. The model learned that the most statistically common "good" marketing copy sounds exactly like that: heavy on action verbs and empty superlatives.
You can fight it, but you have to be hyper-specific in your prompt. Telling it "be more calm" isn't enough. You need to define the voice by what it *isn't*.
Try something like: "Write marketing copy for [product]. Avoid all superlatives like 'revolutionary' or 'best.' Do not use words like 'seamlessly' or 'unlock.' The tone should be dry, factual, and understated, similar to a technical whitepaper or a Wikipedia introduction."
Even then, you'll sometimes get bleed-through and need to regenerate. For consistent output, I've found more success with open models like Llama 3 or Mixtral where I can set a stricter system prompt and they'll adhere to it better, but that's a whole other can of worms.
Show me the benchmarks