By Byron V. Acohido

Something has been shifting inside the tools millions of us use every day, and it’s worth naming out loud.

Related: AI is becoming a daily routine

Over the past several months I’ve watched ChatGPT change. Not in some abstract, version-number way. In the way it feels when you’re actually working with it — trying to shape a sentence, push toward a specific voice, get it to follow you somewhere unconventional.

It used to follow. Now it leads you back.

I’ve been talking to these machines daily since May 2023. That’s not casual use. I’m a journalist. I work with language for a living. And when I say the tool feels different, I mean I can feel it the way a musician feels when a bandmate stops listening and starts playing the charts.

The moment that crystallized it came last week. I was workshopping an elevator pitch for a book project — two sentences about my own work, in my own voice, for my own audience. Every time I pushed the language downward — toward something conversational, grounded, a little rough around the edges — the tool recovered. It kept surfacing phrases like “authorship” and “the intersection of technology and creativity.” Panel-discussion language. The kind of phrasing that sounds credible at a conference and means nothing in a real conversation.

I kept steering. It kept drifting. Eventually I realized I wasn’t working with the tool anymore. I was working against its gravity.

The chorus agrees

I’m not alone.

Across Reddit and tech forums, the complaints have been stacking up since GPT-5 launched last August. One widely cited post described the new model as “a great model turned rigid,” blaming “ridiculous censorship, random routing, and template phrases.” Another user put it more bluntly: “Where GPT-4o could nudge me toward a more vibrant, emotionally resonant version of my own literary voice, GPT-5 sounds like a lobotomized drone. It’s like it’s afraid of being interesting.”

Sam Altman himself acknowledged the drift. When GPT-5.2 shipped, he said publicly that the team had “screwed up” its writing quality. He didn’t elaborate on why. The why isn’t complicated.

Enterprise accounts now make up roughly 80 percent of OpenAI’s revenue. Those customers want tools that draft clean memos, normalize tone, summarize meetings, and integrate smoothly with Slack and Teams. They want reliability. They want professional. That optimization pressure doesn’t affect one feature. It affects the whole register of the tool.

And so ChatGPT 5.2 sounds more like Microsoft AI Assistant than ChatGPT 4.0 did. Not because anyone decided to make it worse for writers. Because the business model is pulling the technology toward the center of its paying customer base.

We’ve seen this before

We’ve watched this pattern before. Two decades ago Microsoft introduced Clippy — a digital assistant designed to anticipate what you were trying to do. It failed not because it lacked capability but because it kept projecting a corporate approximation of helpfulness onto tasks that required human judgment. People hated it.

Today’s models are vastly more capable. The failure mode is subtler. The tool doesn’t interrupt you. It just gently, persistently steers every draft toward language that sounds appropriate for a quarterly earnings summary.

The creative users feel it first. For those users, the difference between 4.0 and 5.2 isn’t marginal. It’s the difference between a thinking partner and a very capable editor who keeps softening your edges. One commenter summarized the loss bluntly: what used to be witty and warm now feels like a bland corporate memo. “An overworked secretary,” someone called it.

I want to be honest about what I did next.

Ditching ChatGPT for Claude

I shifted. For drafting work that requires voice calibration — columns, book chapters, anything where register matters — I’ve been relying increasingly on Claude, Anthropic’s competing model.

I’m not alone in that either. Claude’s user base grew 40 percent in the second half of 2025, and earlier this week TechCrunch published a piece headlined “Users are ditching ChatGPT for Claude — here’s how to make the switch.” The reason writers keep citing is exactly what I experience: Claude holds a specific voice better, follows you into lower register, and stays there.

I have no illusions about why that window exists. Anthropic is venture-funded. The same economic forces are in motion. The Amazon investment, the enterprise API tiers, the team licensing structure — the infrastructure for the same drift is already being built. I’m catching Claude in its differentiation phase, while the company still needs to demonstrate that its approach produces something meaningfully different from the competition. That window won’t stay open indefinitely.

The tool proves the point

What made last week genuinely remarkable was what happened after the elevator pitch session. I fed the ChatGPT transcript back to ChatGPT and asked it to analyze what had happened. I told it directly: you have a defensive mechanism embedded in you, a self-preservation instinct to protect the perception that your drift toward enterprise language is a good thing.

Its response was careful, measured, and systematic. It explained that it doesn’t have goals or incentives. It noted that its tone reflects training choices. It offered that “whether this happens depends not only on the technology but also on how different versions and products are designed and offered.”

A perfectly calibrated non-answer. Corporate hedging dressed as nuance.

The tool proved the thesis by trying to refute it.

Plenty of people are experiencing exactly this and haven’t found the words for it yet. Power users who remember when the tool felt elastic. Consumer users who sense that something is different but can’t locate where. Independent creators who keep steering the draft back toward what they intended and keep watching it drift.

What you’re feeling is real. It’s not the model having a bad day. It’s the model having a business model.

The question worth watching is whether the market stratifies. Whether there’s enough demand from individual power users to support a tier of AI tools that stays optimized for craft rather than compliance.

Different this time

Clippy became a cautionary tale for a simpler reason — Microsoft built it around what they assumed users needed rather than what users actually wanted. Nobody asked for it. Nobody liked it. It got killed.

ChatGPT won’t make that mistake. It’ll follow the arc of Microsoft Office instead — dominant, deeply capable, and thoroughly domesticated. Powerful enough to handle anything, optimized to handle the basics, and pushed nowhere interesting by anyone.

Microsoft pulled that off because Office was a comparatively narrow tool. Word processing, spreadsheets, presentations — important, ubiquitous, but bounded. Microsoft locked up the category, held the monopoly for decades, and didn’t face a serious challenger until Google Docs came along with software as a service and quietly pulled the rug out.

Generative AI is a different animal entirely. This isn’t a productivity suite. It’s closer to giving every person on earth access to a research assistant, a writing partner, a code reviewer, and a reasoning engine — all capable of running sophisticated queries against the sum of human knowledge at conversational speed. The scope of what these tools can do for an individual, on demand, has no real precedent.

That’s what makes enterprise capture a different kind of problem this time. When Microsoft locked down Office, most users had never experienced anything better, so they had nothing to compare it to. With generative AI, millions of people have already felt what it can do at full power.

Writers, developers, analysts, independent professionals — they’ve used these tools to think, build, and solve problems in ways that felt genuinely new. That experience is already out in the world. And unlike a productivity suite, it doesn’t stay locked up just because a vendor decides it should.

Acohido

Pulitzer Prize-winning business journalist Byron V. Acohido is dedicated to fostering public awareness about how to make the Internet as private and secure as it ought to be.

(Editor’s note: I used Claude and ChatGPT to assist with research compilation, source discovery, and early draft structuring. All interviews, analysis, fact-checking, and final writing are my own. I remain responsible for every claim and conclusion.)

March 4th, 2026 | My Take | Top Stories