AI and Technical Writing (#3)
What's Still a Human Job?
So GitHub Copilot cranked out 1,000 lines of code for your engineers yesterday.
Amazing.
The engineering team's sprint velocity is up 20%. Engineering resources are being "optimized." Executives are thrilled.
And yet your users still can't figure out how to actually use the thing the team built.
That's because AI doesn't eliminate complexity. It just moves it. Products still need to be explained, understood, and trusted. The need for clear communication hasn’t changed, but how we’re involved has.
AI-Written Documentation is not Enough
AI can generate passable documentation – the same way it can generate passable code, marketing copy, and dating profiles. But passable work doesn't build trust or create real understanding.
Consider the following things that AI-generated content consistently gets wrong.
Fabricated Information
Ever had ChatGPT invent a citation? Or explain a feature that doesn't exist? AI systems hallucinate with remarkable confidence. A Stanford study found legal-focused models fabricate information in one out of six queries.
This isn't theoretical. In 2023, a lawyer was sanctioned after submitting a legal brief containing six completely fictional cases that ChatGPT had confidently invented. When questioned by the judge, the lawyer admitted he had no idea the tool could make up cases that sounded real but didn't exist.
The difference between sounds right and is right is the difference between a system that helps users and one that actively misleads them.
Human Frustration
AI has never stood in a kitchen staring at a pot of burned rice, wondering what went wrong. It has never muttered, "This should have worked!" while following vague instructions. It doesn’t understand the emotional journey of users.
The most useful documentation anticipates where users will get stuck.
# Typical AI-generated instruction:
1. Add rice and water to the pot.
2. Bring to a boil.
3. Reduce heat and simmer until done.# Human documentation:
1. Rinse the rice thoroughly until the water runs clear to remove excess starch.
2. In a pot add 1.5 cups of water for each 1 cup of rice.
3. With the pot uncovered, bring the water to a boil.
4. Reduce the heat to Low and cover with a tight-fitting lid.
IMPORTANT: DO NOT STIR during cooking. This makes the rice gummy.
5. After 15 minutes, turn off the heat. Leave the lid on for another 10 minutes to steam.
6. Remove the lid.
7. Fluff the rice with a fork before serving.
Additional Tips:
- If your rice burns or sticks, check your heat settings. If the low temperature is too high, the rice will be scorched.
- If your rice turns out mushy, you may have used too much water for the type of rice you are cooking.AI excels at listing steps that work under ideal conditions. It struggles to address what humans actually need: help when things inevitably go wrong.
Personalized Echo Chambers
The best documentation meets users where they are — guiding beginners, offering depth for experts, and providing clarity for decision-makers. But AI-generated content often fails at this because it doesn’t challenge users — it panders to them.
Recent research highlights a growing issue: modern LLMs tend to mirror user preferences and biases instead of delivering truly objective information. This leads to AI-generated echo chambers, where interactive documentation adapts to what the user expects rather than what they actually need.
The result?
Beginners get oversimplified explanations that hide critical nuance.
Experts see jargon-heavy content that lacks real insight.
Decision-makers receive vague, noncommittal answers.
AI tailors content to reinforce each user’s preconceptions — which means every audience gets something slightly wrong.
This isn't documentation that serves everyone. It’s documentation that subtly misguides each audience in different ways.
Human Value in an AI World
If you create products or explain complex ideas, you're already a technical writer – whether that's in your job title or not. And if you want to remain valuable, focus on the areas AI stumbles.
Strong Fact-Checking Skills
Your most valuable skill is the ability to say "that doesn't sound right" and then verify:
Test instructions yourself, step by step
Question anything that sounds vague
Verify that described features actually exist
Check edge cases the AI might have overlooked
A February 2025 study revealed that LLMs often display high confidence in incorrect answers even when they have access to the correct information. This "high-certainty hallucination" means the AI won't warn you when it's wrong. You have to catch the errors yourself.
This goes beyond proofreading. This is intellectual quality control, the kind of judgment AI simply cannot replicate.
Focus on the Recovery Paths
Users don't need help when everything goes perfectly. They need help when they're stuck.
The most valuable documentation focuses on recovery paths:
What specific error messages actually mean in plain language
Common failure points and their solutions
How to troubleshoot complex problems
What to check when things break
Consider this real example. When a system throws a generic "connection timeout" error, what does that actually mean?
An AI might say: "This error indicates the connection timed out. Try again later."
A good technical writer would say: "This usually happens for one of three reasons: the server is down, your credentials expired, or your network is blocking the connection. Here's how to check each one..."
The first explanation is correct but useless. The second one actually solves the problem.
Audience Decisions
Unlike AI, you can make actual judgment calls about your audience:
Create separate content paths for different experience levels
Decide what to explain vs. what to assume
Determine when precision matters more than simplicity (and vice versa)
Prioritize based on real user needs, not generic AI optimization
These are design decisions that require both subject knowledge and human empathy.
Shape the Training
Want to be really valuable? Get involved before the documentation even exists.
AI doesn’t just read documentation. It processes structured knowledge. The person who defines what information gets captured, how it’s structured, and how AI retrieves it is far more valuable than someone who just edits AI outputs.
A recent survey on knowledge management and knowledge graphs emphasizes that AI systems rely on structured data:
Knowledge graphs: Mapping relationships between concepts
Taxonomies: Defining classifications and categories
Ontologies: Structuring domain-specific knowledge for machine understanding
If you help architect the knowledge itself — connecting concepts, structuring relationships, and ensuring retrieval is accurate — you’re shaping the AI’s intelligence.
This isn’t just documentation. It’s knowledge design. And it’s one of the most valuable roles you can play.
Empathy Matters
There's a predictable emotional journey users take with technical products:
Enthusiasm: "This looks awesome!"
Confusion: "Wait, how do I..."
Frustration: "Why isn't this working?"
Self-doubt: "Maybe I'm too stupid to understand this."
Anger: "This product is garbage."
AI can help with stage 1. It makes stage 3 worse. And it has absolutely no concept of stages 4 and 5.
Your value lies in understanding the emotional journey and addressing it directly:
Validate the frustration: "Yes, this is genuinely confusing"
Provide reassurance: "No, you haven't missed something obvious"
Create confidence: "Here's exactly what's happening"
Deliver relief: "And here's how to fix it"
As mentioned in our "Jargon is Killing Your Message" article, research shows that when people encounter frustration with technical (science) content, they often internalize it as personal inadequacy ("I must not be a science person"). The danger? This creates psychological distance that makes them disengage completely. Good technical writing bridges this gap.
It doesn’t just explain what to do — it reassures the user that they’re not the problem.
Final Thoughts
AI can generate endless documentation. However, it's that last mile – the bridge between information and understanding – where human communication becomes irreplaceable.
Anyone can document what something does. The real value is in explaining why it matters, when it breaks, and how to think about it.
AI draws the map, but maps aren’t enough. A real guide knows the following:
Where the roads are washed out
When a trail is too steep
When the “official route” is a dead end
As automation accelerates, the ability to clearly explain complex concepts to humans is becoming the ultimate career advantage.
The people who make things make sense will always be more valuable than those who can’t — no matter how much content AI generates.
Because in the end, the difference between technically correct and actually helpful is the difference only humans can provide.