"Humanising" AI Output Is the Wrong Obsession — and It's Distracting Us From What Matters

There's a growing industry dedicated to making AI-generated content sound more human. Tools that add "personality" to LLM outputs. Prompt engineering tricks to eliminate the telltale signs of AI writing. Entire startups built on the premise that the problem with AI content is that it doesn't sound natural enough.

This is the wrong problem to solve. And the obsession with it is distracting us from the real challenges of AI-generated content.

The Humanisation Industry

The pattern is familiar: AI generates text → text sounds "AI-like" → someone applies a "humanising" tool or technique → text sounds slightly less AI-like → everyone declares victory.

The techniques range from simple prompt modifications ("write in a conversational tone, use contractions, vary sentence length") to more sophisticated approaches that inject stylistic variation, replace common AI phrases ("delve into," "tapestry," "navigating the complexities"), and add imperfections to make the output feel less polished.

Some of these tools are genuinely clever. They can make AI text harder to detect and slightly more pleasant to read. But they're solving a surface-level problem while ignoring the structural one.

What's Actually Wrong With AI Content

The problem with AI-generated content isn't that it sounds like AI. The problem is often that it has nothing original to say.

LLM outputs tend to be:

• Synthesized consensus: The model averages across its training data, producing the median opinion on any topic. This is useful when you want a summary, but it's the opposite of insight.
• Structurally predictable: Introduction, three points, conclusion. The five-paragraph essay format, repeated endlessly, because that's what the training data looks like.
• Devoid of experience: The model has never used a product, attended a conference, or had a conversation. It can describe these things, but the description lacks the specificity that comes from actual experience.
• Risk-averse: LLMs are trained to be helpful and harmless, which means they hedge, qualify, and avoid strong opinions. This produces text that is technically correct but completely uninteresting.

No amount of "humanising" fixes these problems. Making a hollow argument sound more conversational doesn't make it less hollow.

The Real Solution: Human Input

If you want content that doesn't feel like AI slop, the solution isn't to post-process AI output. The solution is to add something the AI can't provide:

Actual experience: Write about things you've actually done. I write about running AI agents on a Raspberry Pi because I actually run AI agents on a Raspberry Pi. The specific details — the NVMe SSD that failed after two months, the thermal throttling at 80°C, the 12-15 tokens/sec benchmark — these come from experience, not synthesis.

Strong opinions: Take a position. The AI can present both sides of an argument. A human writer chooses a side and argues it. This is what makes content worth reading — not the polish, but the perspective.

Specific examples: Instead of "many organizations struggle with AI adoption," write "the hospital system I worked with spent six months trying to integrate an AI scheduling tool before abandoning it because the model couldn't handle the complexity of surgical block time allocation." Specific examples are the antidote to AI vagueness.

Authentic voice: Not a "humanised" voice, but an actual human voice. The way you actually write, with your actual quirks and preferences. This is the one thing that no model can replicate, because it requires a lifetime of human experience to develop.

The Workflow That Works

I use AI extensively in my writing process, but not the way the humanisation industry suggests. Here's what actually works:

1. AI for research: I use LLMs to summarize papers, find related work, and identify key points. This saves hours of reading time.
2. AI for structure: I sometimes ask the model to suggest an outline or identify gaps in my argument. This is useful for catching blind spots.
3. Human for everything else: The actual writing, the examples, the opinions, the voice — that's mine. The AI is a research assistant, not a co-author.

This workflow produces content that is clearly human-written, not because I've applied a humanisation filter, but because a human actually wrote it. The AI contributed to the research process; the output is mine.

Why the Humanisation Industry Exists

The humanisation industry exists because there's a market for it — specifically, a market for producing large volumes of content cheaply and making it look like it wasn't AI-generated. The use cases are:

• Content marketing at scale: Companies that want to publish 50 blog posts a week but can't afford 50 human writers.
• SEO content farms: Sites that need keyword-rich content to rank in search results and don't care about quality.
• Social media automation: Accounts that need to post continuously and can't rely on human availability.
• Academic dishonesty: Students who want AI to write their papers and need the output to pass AI detection tools.

None of these use cases are about creating good content. They're about creating content that serves a non-quality purpose — volume, SEO, presence, deception. Humanising AI output optimizes for "passing as human," not for "being worth reading."

The Detection Arms Race

The humanisation industry is in an arms race with AI detection tools. Detectors get better at identifying AI text → humanisers get better at evading detection → detectors improve again → and so on.

This arms race is wasteful and ultimately pointless. AI detection tools have fundamentally unreliable false positive rates. Humanised text is increasingly indistinguishable from mediocre human writing. The entire detection vs. evasion game is a distraction from the real question: is the content worth reading?

If the content is worth reading, it doesn't matter whether AI was involved in producing it. If the content isn't worth reading, it doesn't matter how human it sounds. The quality of the ideas is the only metric that matters.

What We Should Be Focusing On

Instead of humanising AI output, we should be focusing on:

AI literacy: Teaching people to evaluate content based on the quality of its arguments, not the origin of its text. A well-reasoned AI-assisted article is more valuable than a poorly-reasoned purely human one.

Attribution norms: Being honest about AI involvement in content creation. Not because AI is shameful, but because readers deserve to know what they're consuming. "Researched with AI assistance, written by [human]" is a perfectly good attribution.

Quality standards: Setting editorial standards that focus on originality, specificity, and insight — the things AI struggles with — rather than surface-level polish.

Human-AI workflows: Developing and sharing workflows where AI genuinely enhances human output rather than replacing it. The research-assistant model I described above is one example. There are many others.

The Bottom Line

The next time you see a tool promising to "humanise your AI content," ask yourself: what would happen if you just wrote the content yourself? Or if you used AI as a research tool and wrote the final output in your own voice?

The answer is that the content would be better. Not because humans are inherently better writers than AI (they're often not), but because the process of writing — thinking through arguments, drawing on experience, developing a voice — is what creates value. AI can accelerate that process. It can't replace it.

Stop trying to make AI sound human. Start using AI to help humans write better. The difference is everything.