This post was original written based on my experience using AI tools to write Japanese posts. You can read the original text here.

I run Kafkai, an AI content platform. Over the course of 2025, I wrote blog posts for kafkai.ai and kafkai.com with AI tools nearly every week. AI writing is at the core of what we build, so it seemed only right that I use it thoroughly myself.

To be honest, what I expected was time savings. What I got was something different. Time was saved, yes. The time to produce a first draft roughly halved, I think. But more than that, I was confronted every week with the distance between "what AI can write" and "what people actually want to read."

2025 was a turning point for AI writing tools. Language quality improved dramatically, open-weight large language models (LLMs) and tools multiplied, and enterprise adoption rates surged. But after a full year of using them, what became clear was the existence of a gap that tool evolution alone cannot close.

What Happened to AI Writing Tools in 2025

Here is a look back at the major developments in 2025.

  1. GPT-4.1 and GPT-4.5 OpenAI significantly raised the quality of text generation in 2025. GPT-4.5 in particular touted "emotional intelligence" and delivered noticeably better contextual understanding. The stiff, machine-translated feel of earlier outputs to Japanese was markedly reduced.

  2. Claude's long-form generation Claude, from Anthropic, pulled ahead in long-form content generation. Its structural coherence and ability to maintain context over extended pieces made it particularly strong for business documents and blog posts.

  3. Gemini's deep search integration Google's Gemini changed the "research then write" workflow at a fundamental level by integrating search directly into generation. The line between research and writing is blurring.

  4. The rise of AI agents From the second half of 2025, AI writing tools began shifting from simple "text generators" to "agents" that handle entire workflows. Keyword research, competitive analysis, outlining, writing, and proofreading, all in one pipeline. We are actually in the process of implementing these capabilities in Kafkai.

  5. The proliferation of specialised tools Dedicated AI writing tools tailored to specific markets and languages kept launching throughout the year. This reflects the reality that generic multilingual models still fall short for nuanced, market-specific content.

The numbers make the momentum clear. The generative AI market is projected to reach $19 billion by 2028, 74.5% of SEO teams are already using AI tools, and over 40% of companies have adopted AI in their content workflows. AI writing is no longer experimental. It is part of the job.

For a deeper look at these market-wide trends, see our 2025 AI marketing trends analysis.

What Actually Changed

So what changed at the practical level? Here are the three concrete shifts I experienced over the year.

1. Research speed went up dramatically.

It used to take me a full day just to find sources, read through them, and organise the key points. Now I can have AI read multiple sources, summarise them, and structure the arguments in a matter of hours. Cross-referencing information across languages in particular became incomparably faster.

2. Draft quality improved, although "natural-sounding text" and "text worth reading" are not the same thing.

Since GPT-4.1, AI-generated text has become noticeably more natural. The awkward, machine-translated feel has largely disappeared. But sounding natural and being interesting are entirely different things. AI drafts are grammatically correct and well-structured. But reading them, you often end up thinking "huh, okay" and moving on. The phrases that stick in a reader's mind, the unexpected angles, the warmth behind the words. Those still do not come from AI. This is where humans need to step in.

3. The workflow changed.

AI started taking over the tedious parts. Outlining, structural suggestions, research synthesis, first-draft generation. These are genuinely areas where AI excels. But ultimately, what determines the value of an article is the step where I add my own perspective, experience, and opinions. That part is still very much a human job.

For practical guidance on using AI writing tools, see our guide to AI writing tools.

What Has Not Changed (And Will Not Change Anytime Soon)

Here is where it gets real. No matter how much AI writing tools evolved in 2025, there are fundamental problems that have not budged.

1. The lack of originality.

AI is exceptionally good at summarising secondary information. It takes existing articles and data, organises them, and restructures them into something readable. But it cannot produce primary information. Data only your company has, insights from your own experience, the intuition you develop on the ground. AI simply cannot generate these.

The result is that content relying solely on AI writing ends up looking the same as everyone else's. It pulls from the same sources, organises the same information, in the same way. Of course it does.

2. The quality gap.

Here is something I noticed. AI writing is, on average, correct. But writing that is correct on average is, on average, boring.

Readers finish an article not because it is correct, but because it is interesting, because it offers a new perspective, because it solves their problem. AI can now pass the accuracy test. But its ability to pull readers in still lags far behind humans. I have lost count of the times I read back an AI draft and asked myself, "There is nothing wrong here, but is there any reason to publish this?"

3. The fact-checking burden.

This one gets overlooked but it is serious. AI writes incorrect information with complete confidence. Training data for non-English languages is smaller, which means errors in market-specific data and statistics come up more often than you would think. I verify every number and proper noun AI generates, every time. This checking work sometimes eats up a significant chunk of the time I saved on writing.

4. Copyright and originality concerns.

As long as generative AI outputs text based on training data, copyright issues are unavoidable. Regulatory guidance is evolving in multiple jurisdictions, but the grey areas remain wide. With images and video, you can often spot the similarities visually and flag them. With text, it is far harder to tell.

These are not problems that get fixed by a version upgrade. They are structural limitations. No matter how much the quality of AI-generated text improves, the fundamental question of whether a robot can write better than you does not go away. Tools evolve. But the essence of writing, I think, does not.

Making AI Writing Actually Work

So what should you actually do to make AI writing useful right now? Here are the practical steps I have arrived at after a year of trial and error.

  1. Let AI draft, finish it yourself. About 70% of what AI generates is "fine, no issues" quality. But the value of an article is decided in the remaining 30%. Your words, your perspective, your experience added in that final editing pass is what differentiates your content.

  2. Bring primary information. Your own data, your own stories, your own research findings. Incorporating these into the AI writing process produces something fundamentally different from other AI-generated content. Primary information cannot be copied. That is the single biggest differentiator.

  3. Define your brand voice. If you just tell AI "write an article," what you get is generic text that could have been written by anyone. Your tone, your word choices, how you approach a topic. Defining your brand voice clearly changes the quality of AI output dramatically.

  4. Build fact-checking into the process. Fact-checking AI-generated content is not "nice to have." It is mandatory. Numbers, statistics, proper nouns, dates. Verify these against original sources before publishing. The cost of losing trust when errors surface later is far greater than the time spent checking.

  5. Strategy over volume. With AI writing tools, producing 100 articles a month is technically possible. But 10 articles that solve real problems beat 100 articles that say nothing. Building a content plan around keyword strategy is how you differentiate in the age of mass production.

What Comes Next

From 2026 onward, it is nearly certain that AI agents will take over entire content production workflows. Research, writing, SEO optimisation, distribution, all handled autonomously. The question is not "will AI take my job" but "how do I adapt."

But the more AI-generated content floods the web, the more the value of human editorial judgement actually goes up. What to write, who to write for, why to write it. Only humans who understand their readers and feel the market firsthand can make those calls. We will of course need more data to make sense of it all and have a better understand of things, but in a world where AI writing is the default, human judgement with that data becomes the competitive advantage.

What we are building at Kafkai is precisely this: making the collaboration between humans and AI practical. Let AI handle what AI is good at, and let humans focus on what only humans can do. Drawing that line clearly and making it easy to work with.

To Wrap Up

AI writing evolved significantly in 2025. Text quality improved dramatically, tool options multiplied, and workflows got more efficient. But tool evolution does not guarantee evolution in how you use them.

The gap between AI's "correct but boring" output and "writing worth reading" that has human insight baked in has, if anything, widened through 2025. As the tools got smarter, the skill of the person using them started showing up more clearly in the results.

The question is not whether AI writing tools are good enough. It is what you choose to write with them, and who you choose to write it for.