ChatGPT-5 Test Series - Part 4: Authenticity in Text Generation
🕓 Read Time: ~5 minutes
What does “authentic” mean in AI text?
When we talk about authenticity, we’re asking: does the writing feel human and natural, or does it read like AI-polished copy?
GPT-5 introduces new “personas” you can select to shape tone. For me, that’s not useful. I work across so many different contexts that I’d constantly have to switch settings. Instead, I rely on custom GPTs, where I’ve already built in the tone, structures, and do’s and don’ts of my writing voice.
So the real question was: with the same setup, does GPT-5 do a better job than GPT-4o?
The experiment: same inputs, different models
I ran a simple test:
-
Same Assistant (aka CustomGPT)
-
Same inputs from me (topic, insights, angle, CTA).
-
The only difference: the model (4o vs 5).
I tested this across two use cases:
-
LinkedIn posts
-
This newsletter series itself
Then I evaluated the outputs on three criteria:
-
How good was the initial draft?
-
How well did it follow my tone instructions (+ the do’s and don’ts)?
-
How much rework did I need to do before publishing?
🔍 My take
GPT-5 was much closer to my voice. It didn’t overcorrect, and it stayed much closer to my tone instructions, especially around the use of emojis and avoiding overhyped “AI-speak.” With GPT-4o, I often had to strip out fluff and correct the tone myself.
Of course, I still make my own edits as the content evolves, but with GPT-5, the first draft is far closer to what I intended, provided that I give it enough context, insights, and angles to work with.
Interestingly, the newsletter outputs were stronger than the LinkedIn posts. At first, I thought this was a model issue, but on review, I realized the difference was in my setup: my Newsletter Assistant instructions are far more detailed than the ones for my LinkedIn Assistant. In other words, it wasn’t GPT-5’s fault, it was my instructions. Time to give my LinkedIn Assistant a good overhaul!
Key Takeaway
GPT-5 does a noticeably better job at authenticity than 4o. As long as it is well instructed on the tone, the do’s and don’ts, and you provide it with sufficient context, insights, and angles, the output will feel much closer to your true voice. Garbage in, garbage out continues to hold. And hey... forget about the persona feature, it's not really helping.
Til next time,
Elena
P.S. Curious how one of these newsletters comes together? I can record a quick Loom to walk you through the process. Just hit reply if you’d like to see behind the scenes and I'll make it a follow-up edition.
