Every day, I write down one small thing I notice while working with AI. Not predictions about what AI will do in five years. Usually, they are much smaller moments:

• My editing agent refusing to simply edit.

• A stakeholder building a prototype.

• An AI tool changing more than I asked it to change.

• A writing pattern I suddenly began seeing everywhere.

Separately, they looked like observations about different tools and situations. Together, they pointed to the same shift:

A role is the position someone holds in a team. A skill is what they can do. A responsibility is the outcome, perspective, or decision they are expected to protect.

For a long time, these three things were bundled together. A designer created prototypes and protected the user’s perspective. A developer wrote code and protected the technical integrity of the system. A copywriter shaped an argument and owned what it meant.

AI is beginning to unbundle this package. A stakeholder can now create a prototype. A designer can use AI to generate code. An AI agent can edit an argument - and even challenge it. Almost anyone can produce polished writing.

Skills - or at least access to the outputs those skills once produced - can now move beyond the roles they traditionally belonged to. But responsibility does not automatically move with them.

A stakeholder may generate a convincing interface without becoming responsible for representing the user throughout the project. A designer may produce code without becoming responsible for its long-term technical integrity. An AI agent may rewrite an argument, but it will not be responsible for what that argument means, whether it is true, or whether it should be published.

This is where a new gap appears.

I think of it as the responsibility gap.

The skill moves. The responsibility stays behind. And when this new distribution is not made explicit, capabilities spread faster than accountability, judgment, and ownership.

This week, I noticed five signs of that gap.

Each one suggests something practical we may need to change in the way we work.

1. Everyone can prototype - but the user’s perspective can become ownerless

Developers can design. Designers can code. Stakeholders can generate working interfaces before the team has even agreed on the problem. The traditional boundaries between roles are becoming less visible.

But the perspectives attached to those roles do not automatically become shared. When everyone can produce a solution, solution production is no longer the scarce skill.

Insight: Assign ownership of perspectives, not only deliverables

Before starting an AI-assisted project, explicitly assign who is responsible for protecting:

• the user’s perspective;

• business viability;

• technical integrity;

• risk and compliance;

• the final product decision.

This does not mean that only one person may contribute to each area. It means that someone remains accountable for making sure the perspective is not lost. A useful question to add to a project kickoff is:

Without explicit ownership, responsibilities do not necessarily become shared. They may simply become ownerless.

2. AI makes generation cheaper, but creates invisible supervision work

AI prototyping feels fast because the first output appears almost instantly. But the speed of generation hides another category of work:

• reviewing,

• correcting,

• restoring,

• constraining,

• and explaining what should not have changed.

The visible task becomes cheaper. The invisible supervision around it grows. The human does not disappear from the process. The human moves from creator to reviewer, controller, and repair layer.

Insight: Measure correction cost, not only generation speed

Do not evaluate an AI prototyping tool only by how quickly it produces the first version.

Also track:

• how many unintended changes appear per iteration;

• how often previous decisions need to be restored;

• how much time is spent reviewing the output;

• how often context must be repeated;

• how many iterations are required before the result is safe to use.

One simple metric could be:

If AI generates something in two minutes but requires forty minutes of cleanup, the meaningful productivity story is not the two minutes. Generation is only one part of the cost.

3. AI can improve the sentence - and challenge the author

I write instructions to my editing agent in Ukrainian because it feels more natural to me. Its job is simple: improve my English comments.

This week, I wrote:

AI would never come up with that detail on its own.

Instead of polishing the sentence, the agent stopped me. It pointed out that AI sometimes does suggest details like that. What it cannot do is actually live through the moment behind them.

I was not expecting my editing agent to defend AI. But that was not really what happened. It challenged an assumption that weakened the point I was trying to make. The result was probably more accurate.

It also changed the relationship. The tool was no longer only improving how I expressed the thought. It was participating in the thought itself.

Insight: Separate editing from decision-making

When using AI to edit or refine important work, ask it to distinguish between different types of intervention:

• language correction;

• structural improvement;

• factual challenge;

• change of argument;

• added assumption.

These are not the same thing.

A useful prompt could be:

Improve the writing, but flag separately any place where you disagree with the premise, change the argument, or introduce a new claim.

This keeps AI’s contribution visible. The final question should not only be:

Is this better written?

It should also be:

Is this still my argument?

4. AI can improve our writing - and make all of us sound the same

This week, I learned the word anaphora. It is a writing technique that repeats the same word or phrase at the beginning of consecutive sentences. To be honest, I had never heard the term before.

AI clearly had.

Once I learned the name, I started recognizing the pattern everywhere:

The same beginning.
The same rhythm.
The same spacing.

One sentence. A new line. Some space. Another sentence. AI did not invent anaphora. Human writers have used it for centuries. But AI can identify effective writing patterns and reproduce them at extraordinary scale.

That is why AI-assisted writing can sound polished and strangely familiar at the same time. The risk is not that the pattern is bad. The risk is that a useful technique becomes an automatic default.

When many people use similar tools trained on similar writing patterns, “better writing” can begin to converge into the same voice.

Insight: Review AI writing for patterns, not only errors

After asking AI to improve a text, do a separate pattern review.

Look for:

• repeated sentence structures;

• excessive one-line paragraphs;

• artificial contrasts;

• sets of three;

• dramatic conclusions;

• phrases you would not naturally use;

• repetition added mainly to create rhythm.

A useful follow-up prompt is:

Identify the rhetorical devices and recurring structures you introduced. Do not rewrite the text.

Then decide which patterns genuinely serve the idea and which only make the text sound generically polished. AI should help us use writing techniques intentionally. It should not choose our voice by default.

5. Context is becoming something we need to design

Sometimes I add a thought to ChatGPT as a note. And sometimes ChatGPT interprets it as a question.

This time, I wrote:

Maybe the branch feature is ChatGPT’s attempt to give context a structure.

ChatGPT answered that branching provides context through controlled inheritance rather than by simply adding more of it. So, assuming we trust ChatGPT’s competence on the subject of ChatGPT, we have confirmation. Or at least a very confident explanation.

But the observation itself points to something larger.

As AI conversations become longer, context stops being an invisible technical mechanism. Users need to decide:

• Which assumptions should continue?

• Which decision belongs only to this experiment?

• What should a new branch inherit?

• How can we return to a stable point?

More context is not always better. Sometimes more context simply means more assumptions, more contamination, and more opportunities for the system to reinterpret something that had already been decided.

Insight: Treat context as part of your information architecture

For recurring AI work, do not keep everything in one endless conversation. Create explicit context boundaries:

• one thread for exploration;

• one for approved decisions;

• one for execution;

• one for experiments that should not affect the main direction.

Before continuing a long AI interaction, state:

In AI-native products, context management may become as important as navigation. Users will need to understand not only where they are, but also:

• what the system remembers,

• what it inherited,

• what it considers a decision,

• and what can still be changed.

Context is becoming something we need to design.

Five practices I am taking into next week

These observations led me to five practical rules.

1. Assign ownership of perspectives

When role boundaries blur, explicitly decide who protects the user, the business, the system, and the final decision.

2. Include correction cost in productivity

Measure how much human supervision follows the instant output.

3. Ask AI to label the type of intervention it makes

Editing, challenging, and introducing a new assumption should not look identical.

4. Audit AI-generated patterns

Review not only whether the output is correct, but whether it still reflects your reasoning and voice.

5. Design context boundaries

Decide what AI should inherit, what remains experimental, and which decisions are fixed.

The larger mutation

The common thread across all five observations is ownership.

• Who owns the user’s perspective when everyone can build?

• Who owns the unintended changes inside an AI-generated prototype?

• Who owns the argument after AI improves it?

• Who owns the voice when the writing pattern came from the model?

• And who decides which context should survive?

That part still has to be designed. The practical task is not to force skills back into their old roles. It is to make the new distribution of roles, skills, and responsibilities explicit - and to reassign the perspective, judgment, and accountability that used to travel with those skills.

Otherwise, skills will continue to move.

And responsibility will continue to stay behind.

Throughout the week, I publish short observations about how AI is changing products, roles, and the way we work. Every week, I connect all of them into one practical digest.

Subscribe to receive the next Observation Digest.

Subscribe now