Despite its upsides, I’ve found ChatGPT 5.1 introduces a few quirky changes as well

ChatGPT 5.1 arrived last week, bringing better smarts, faster responses, improved personality modes, and more. I’ve had the chance to play around with it for about 6-8 hours in total so far, and I agree that while it’s not a massive upgrade in terms of processing power, it is still a notable improvement in many ways. Of course, that doesn’t mean it isn’t without its own flaws or setbacks. This is especially true if you use ChatGPT 5.1 as it is out-of-the-box.

Below, I cover the biggest quirks I have run into and a closer look at the customization that helps if the default setup does not impress you. Before getting into that, I want to highlight the changes that stood out in a positive way as well.

The most impressive changes I’ve found in ChatGPT 5.1 so far

One of the first things I noticed was a shift in tone and structure. ChatGPT 5 favored brevity and often leaned on bullets, emojis, and em dashes instead of clear paragraphs and fuller explanations. GPT-5.1 still has a casual, TLDR leaning style, but the word choice and depth of response feel different. The model tends to offer more substance without dropping the informal tone.

To be clear, these responses haven’t been without a few quirks, which I will get into later. The bigger point is that its replies feel deeper and better aligned with my own personality.

A major reason is that it seems to remember me more reliably. GPT-5 was inconsistent about past conversations. GPT-5.1 is noticeably better. It brings up names, past events, personal preferences, and other details with far more accuracy. It even follows the stylistic rules I give it without issue. For example, I told it to never use em dashes when proofreading. GPT-5 frequently ignored this request, even when it acknowledged the rule in its own thinking instructions. GPT-5.1 has honored the request through all of my testing so far. Could it slip eventually? Possibly, but this is still progress.

The model is also more willing to use what it knows about my habits, interests, and personality without being prompted. Conversations feel more alive as a result. For example, I talked about how it’s been easier to deal with behavioral problems with my cat than with past situations with similar cats. GPT-5.1 noted that my experience owning rats over the last two years might play a role.

It pointed out that rats tend to be skittish and that winning them over requires patience and consistency. It suggested that this experience could influence how I handle animals in general. The thing is, the last time we spoke about my rats was literally months ago, on a different model but the same account. That’s pretty impressive, I think.

It can still get close to the sycophantic tendencies of older versions, but GPT-5.1 strikes a better balance. It can be encouraging and complimentary, yet it also pushes back when it thinks I am wrong or oversimplifying something. While I still wish it were a bit more direct by default, I also recognize that not everyone wants blunt feedback.

The main takeaway is that GPT-5.1 feels more like a real conversational partner. It handles context better, follows instructions more reliably, avoids dropping steps, and responds far faster.

Now for the less-than-good. ChatGPT 5.1 often gave me better answers overall, but some responses still felt off and oddly unnatural from time to time.

A good example came from a role-playing game setting I’m building for my boys. I told the model, “They decide to settle in the Americas, being there are no humans to disturb.” Later, it replied with, “‘We pick a continent with no humans’ is plausible.” That quote was never mine, so the phrasing felt oddly invented even though the point was clear. While the rest of the response was very good as far as the information given, the wording was a bit perplexing.

Proofreading showed several quirks as well. It leaned heavily on bullets and shortcuts and sometimes rewrote more than needed. That said, I was able to fix much of this by requesting no bullets, no em dashes, and other similar orders.

In terms of general responses, while it did match my personality better and required fewer prompts, the tone often felt more like an online friend than an AI partner for serious tasks. Its sentences were often shorter, occasionally sloppy, and it was more willing to use slang or profanity.

Personality mirroring explains part of this, but just because I occasionally swear does not mean I want it echoed back. I want a tool that stays grounded. The model also used emojis more often, far more than what it would have learned from my own tone. I do not mind this in casual chats about science or philosophy, but formal topics benefit from a more structured style.

Fluff crept in as well. Phrases like “here’s the part most people never realize” or small bits of praise appeared where they did not add anything meaningful. The flattery is better controlled than in older models, but it is still there.

It also seems to have a love affair with bullets that is similar to what it and earlier models have with em dashes. When I asked for a short WWI breakdown, GPT-5.1 gave me a dense wall of bullets that read flatly. GPT-5 handled the same request with a clearer mix of prose and structure, even if 5.1 offered deeper factual content.

The good news is that most of these quirks can be toned down through personality settings. With 5.1, those adjustments stick more reliably than before, which makes fine-tuning much easier.

Customization features seem to have more impact in ChatGPT 5.1

As you might have noticed, not all the quirks I mentioned above will matter to everyone. People use AI tools differently, and you may notice entirely different quirks than I did. The good news is that ChatGPT now makes it much easier to fine-tune your experience. Preset personalities and custom instructions were already available, but this system has received a major upgrade.

I tested many of the same prompts across ChatGPT 5 and 5.1 using several personalities. In the past, these personalities helped refine the experience, but they didn’t feel as developed as they do now. The upgraded versions are warmer and more expressive, although the shift also means some models respond quite differently than before.

For example, I used to love the Nerdy personality. It tended to be excitable and passionate, while still getting to the point faster than the default style. To show the contrast, I compared how Nerdy responded to “explain the key events of WW1 to me” in GPT-5 versus GPT-5.1 in the screenshots below.

I have to say, GPT-5.1’s version felt closer to the Quirky personality mode in GPT-5. I enjoy the more descriptive and engaging approach, but it is not always what I want. Switching to “Professional” helped bring things back to the more direct Nerdy-style output I liked in GPT-5, just with added detail. It was also a bit too boring.

Thankfully, custom instructions can refine any base tone even further. For example, I added “warmth, exploratory, enthusiastic” to Professional. This kept the clarity of the business style but added a bit of personality, and subsequent responses were richer and better aligned with what I prefer.

I also found that instructions like “no em dashes” or “minimize bullets” help steer the model away from some of the quirks I mentioned earlier. That said, personalities and custom instructions can introduce quirks of their own. For example, the business personality sometimes began responses by announcing “Below is a clear, direct, business-style explanation…” and if I included “no fluff” in custom instructions, it would occasionally respond with lines like “here is a no-fluff answer.”

The upside is that this behavior was even more common in ChatGPT 5, but GPT-5.1 doesn’t always feel the need to remind you about your custom instructions every time you start a new conversation. So it’s better, just still inconsistent here.

ChatGPT 5.1 certainly feels faster, friendlier, and more aware than ever before. Is this all just smoke and mirrors? Very possibly, but it’s hard to deny that the new experience is more engaging and lively than before. With customization, it’s also much easier to get the answers you are looking for with this model.

Of course, it still hallucinates. It still uses poor wording from time to time. It still is sychophantic more than I’d like, even if this can be dialed back through personality customization. For me, I always felt GPT-5 was better than many gave it credit for, though I recognize 5.1 gives it an extra coating of paint and a few other improvements that help it stand above the older model.

For those that weren’t impressed by GPT-5, this may very well be the improvement you were waiting for. That said, those who prefer legacy models like GPT-4.1 may still find that this model does little to change their minds.