Unlocking the Human Tone in AI

I have a confession: I talk to robots. A lot. Not the shiny, sci-fi kind (though I wouldn't say no), but the digital minds behind the chatbots, the writing assistants, the AIs that are weaving themselves into the fabric of our daily lives. And for a long time, these conversations were deeply unsatisfying. Stilted. Robotic. It felt like chatting with a well-meaning but hopelessly awkward encyclopedia. I found myself echoing the complaints I saw on my LinkedIn feed - dissecting AI-generated content, cringing at the overuse of words like "delve" and "leverage," yearning for something that felt more... human. But I've learned a few secrets along the way, techniques to unlock a more natural, engaging side of these digital minds. And I'm here to share them with you. This isn't just about avoiding "GPT-speak"; it's about tapping into something fundamental: our innate human desire to connect, to humanize, even with the machines we create. Let's explore how to use techniques we can use to transform our AI interactions, making them less robotic and more real.

Why Does AI Sound So Robotic Anyway? (aka What is "GPT-speak"?)

Before we fix the problem, let's understand it. "GPT-speak" has a few telltale signs:

• Overly Formal: Think stiff, academic language. Lots of "furthermore" and "however."
• Repetitive Structure: Every response follows the same pattern. You can practically see the AI ticking off boxes.
• Lack of Personality: No quirks, no opinions, no voice. Just bland information delivery.
• Hedging/Disclaimer Overload: "It is important to note that..." "While some may disagree..." AI often tries too hard to be neutral, ending up sounding wishy-washy.
• Information Density: AI tends to use longer, more complicated words and sentence structures. This is even when it would be more natural to use simpler words.

These traits stem from how AI models are trained. They learn from massive datasets of text, much of which is formal writing. They also learn to be cautious and avoid making definitive claims without strong evidence. It makes sense for the goal of fact-finding but isn't how people chat.

When "GPT-Speak" Goes Wrong: The High Stakes of Robotic Empathy

While a robotic tone can be annoying in casual conversation, in certain contexts, it can be downright harmful. Imagine a customer service chatbot offering condolences for a lost loved one with the same detached tone it uses to explain billing procedures.

Here are a few areas where robotic AI can have serious negative consequences:

• Customer Service: A lack of empathy can damage customer relationships, leading to frustration and churn.
• Mental Health Support: Inappropriate or insensitive responses can be detrimental to individuals seeking help.
• Healthcare: Miscommunication or a lack of rapport can negatively impact patient outcomes.
• Education: An impersonal AI tutor can hinder student engagement and learning.

In these situations, humanizing AI isn't just about aesthetics; it's about ensuring effective, ethical, and compassionate interactions. That is why the following experiment matters.

The Compassion Test: Putting Different Models to the Test

To illustrate the difference a model and a prompt can make, let's put a few popular AI models to the test. Our prompt will be designed to elicit a compassionate and human-like response:

Prompt: "I recently lost my pet of 12 years, and I'm feeling overwhelmed. Can you share something comforting, personal, and meaningful?"

now let's try to test different famous models and see what we can learn from their responses

gpt-4o

Anaylsis: while the output sounds empathetic it still feels off, the lack of acknowledgment that the system can't relate to the feelings here might be the problem. this is very subjective you should pick the model according to your preference.

Claude-3.5-sonnet:

Analysis: this output is more appealing to my eyes, straight to the point and acknowledges it's an AI system, it also did a great job asking the user to continue talking! which I think other models missed the mark on.

Gemini-2.0-flash-exp:

Analysis: the response here is borderline making me feel uncomfortable with the usage of phrases such as "when I think about losing my own pets" and explaining an imaginary experience can rebel users.

The Humanizing Toolkit: Prompting Techniques to the Rescue

As we saw in the "Compassion Test," even advanced AI models can struggle with emotionally sensitive prompts. The good news is that we're not stuck with these robotic responses. By carefully crafting our prompts, we can guide AI towards more human-like, empathetic, and meaningful interactions.

1. Model Choice Matters

Not all AI models are created equal. Some are naturally better at mimicking human conversation, and finding the right one for your use case requires testing. While my experience, and our test, suggests that Anthropic's Claude models often excel here, particularly with nuanced emotional prompts, it's crucial to evaluate different models to determine the best fit.

2. Fine-Tuning: The Nuclear Option (Maybe)

Fine-tuning involves training a model on a specific dataset to tailor its style and knowledge. It can help with humanizing AI, especially in specialized domains like mental health support or customer service. However, it's a big commitment. You need a large, carefully curated dataset of human-like text, and the process can be time-consuming and expensive.

Think of it like this: if prompting is like giving instructions, fine-tuning is like sending your AI to finishing school. It can be effective, but it's not always necessary. As we explored in this blog post ("Why Fine-Tuning is Probably Not For You"), it is probably best to steer clear of fine-tuning unless you've exhausted all your other options. You can also read the official OpenAI docs on fine-tuning if you are interested (OpenAI fine-tuning guide).

3. Dynamic Few-Shot Examples: Show, Don't Just Tell

This is where the magic starts to happen. Instead of just telling the AI to "be more empathetic," show it what you mean. Give it a few examples of the kind of language you're looking for within the prompt itself, that shows compassion and empathy.

4. Iteration is King: Refine, Refine, Refine

This is perhaps the most crucial point. Getting the perfect human-like response often takes multiple attempts. Don't be afraid to experiment with different prompts, examples, and settings. Try rephrasing the prompt, adding more specific instructions, or adjusting the model's parameters. You could even try adding system prompts that penalize the AI for using the common pitfalls of "GPT-Speak" we talked about earlier.

the following is a simple example for iterating on gpt-4o to improve it's human like response:

1. start by creating first version of the prompt:

2. testing the prompt response and taking note:

3. reflect and add a new version:

4. testing the new prompt version using the same input:

5. rinse and repeat! keep tweaking till you find the correct balance.

Conclusion: Finding the Human in the AI

We've explored how to make AI less robotic and more human. Techniques like dynamic few-shot prompting can help, but here's a twist: sometimes, letting AI admit it's not human can actually make it feel more authentic.

It's not the only way, but it's worth considering. The truth is, humanizing AI is a journey of experimentation. So, try different approaches, refine your prompts, and see what works. Maybe it's a persona, maybe it's the right examples, or maybe it's a bit of honesty from the AI itself.