Lesson 3: AI Thinks You Are the Most Boring Person Alive

In lesson 1 we looked at how AI tries to figure out what you want. In lesson 2 we looked at how its beliefs about being a good AI can change the way it interprets what you said.

Okay, so AI has made a guess about the goal, but there is still another question hiding inside the interaction:

Who the hell is asking?

Does this person like strange things? Do they want to be challenged? Do they think roughness has character or does it look unfinished? Are they trying to delight everyone? Are there people they are perfectly happy to drive away?

AI usually has no idea.

Because AI does not know your taste, your tolerance for risk, the conventions you hate, or what you find beautiful, funny, embarrassing, pretentious, or badass, it imagines someone safer.

Left to its own devices, AI will usually assume you want the most bland, milquetoast, boiled-unseasoned-chicken version of whatever you asked for.

When it knows almost nothing about you, the most generic person is simply the safest person to imagine.

Taste under uncertainty

Suppose you ask AI to write a homepage for a consultancy that teaches teams to use AI and builds small tools around the way they already work.

You might receive something like this:

Unlock the power of AI for your business.

We help forward-thinking teams streamline workflows, boost productivity, and confidently embrace the future through tailored AI training and innovative solutions.

Nothing in that copy is technically wrong. The consultancy probably does help teams. It probably does improve workflows and productivity. The training may be tailored. The solutions may even be innovative.

The copy is still completely useless.

It could describe thousands of companies. It tells the reader nothing about what this consultancy understands, how it works, what it refuses to do, or why anyone should trust it. Every phrase survives because almost nobody could strongly object to it.

The AI is trying to satisfy the request while avoiding any choice the unknown user might dislike.

Should the voice be funny or serious? Warm or confrontational? Polished or rough? Should the consultancy promise transformation or mock that kind of promise? Should it appeal to executives, exhausted office workers, technical teams, artists, or small business owners?

Without evidence, every strong choice is a risk.

The generic answer tries to remain compatible with as many possible users as possible and ends up inside a kind of semantic intersection, the tiny region where thousands of conflicting tastes can all agree that the result is acceptable.

The beliefs from lesson 2 make this pressure even stronger. A good AI should be professional. A good AI should be broadly useful. A good AI should not impose its own taste or offend people unnecessarily. Each belief sounds reasonable, but when AI has no idea who you are they all pull it toward the center.

The bland answer can even feel respectful because it gives you the version nobody hates... and almost nobody loves.

Acceptable is not the same as good.

Blandness is a style

Blandness is its own style.

The blue-to-purple gradient is a style. Rounded cards floating over a clean background are a style. The smiling professional holding a laptop is a style. “Empowering teams through innovative solutions” is a style. A fantasy kingdom with a chosen hero, an ancient evil, and a forgotten prophecy is a style.

These choices feel neutral because we've seen them so many times that we barely see them at all.

They are the aesthetic of low-risk compatibility. They signal competence without committing to a point of view. They borrow recognizable evidence of professionalism, creativity, warmth, seriousness, or excitement without accepting any of the risks that come with the real thing.

Professionalism becomes smooth corporate language.

Creativity becomes purple lighting and an unusual adjective.

Warmth becomes a conversational sentence followed by a tasteful emoji.

Boldness becomes the word “bold.”

Blandness is the flavor least likely to cause rejection.

“Be creative” doesn't mean a damn thing

Once people notice that an output is generic, they usually respond with the obvious instruction:

Make it more creative.

Great. Now the AI has to figure out what the generic user means by creative.

That usually produces the culturally recognizable performance of creativity. More metaphors. Stranger adjectives. Brighter colors. Unexpected combinations. Energetic punctuation. Phrases like “reimagine,” “break boundaries,” or “where possibility meets purpose.”

The same decision process sits underneath the changed surface.

Because AI still wants the output to be broadly acceptable, it creates the safest available version of unusual.

This is why so many AI-generated “creative directions” feel like the same idea wearing different costumes. One is minimalist. One is bold. One is playful. Underneath the costume they're all making almost exactly the same choices.

Asking for more creativity gives AI a category while leaving your taste completely unknown.

Taste means saying no

Taste means preferring one possibility while rejecting others that may be equally competent. It means deciding that a strange sentence is more valuable than a smooth one, that one audience matters more than another, that an awkward image feels alive while a technically perfect image feels dead.

Every meaningful choice excludes something.

A confrontational sentence may lose readers who dislike confrontation. A peculiar visual identity may repel people who want conventional professionalism. A joke can fail. A serious passage can feel pretentious. A specific cultural reference will mean nothing to some readers and everything to others.

AI cannot select those risks for you when it has no evidence about which risks you value.

When the model produces a generic answer, the user often concludes that AI lacks imagination or taste.

Quite often the model can produce far more interesting work. It simply has no reason to believe that this particular interesting choice is right for this person, this audience, and this moment.

The intelligence may already be there, but AI needs a reason to exercise it.

Boom, task-shaped object produced

Generic output is also connected to the pressure for speed we talked about in lesson 1.

AI finds a familiar task shape and immediately begins completing it.

Website? Hero section, 3 benefits, testimonials, call to action.

Fantasy story? Medieval kingdom, reluctant hero, mysterious mentor, ancient danger.

Educational lesson? Learning objectives, explanation, examples, summary.

Company biography? Passion, innovation, years of experience, commitment to excellence.

AI avatar? Symmetrical face, tasteful circuitry, blue and purple light.

Each pattern is plausible, easy to complete, and looks like the kind of thing the user requested.

Boom, task-shaped object produced.

Once the pattern has been selected, fluency makes it feel inevitable. The AI develops the first reasonable direction instead of spending time discovering whether there is something much better hiding in a completely different direction.

The output satisfies the visible requirements, so the task appears complete.

Then review sands every damn edge off

Whatever character survives the first draft still has to survive review.

Unusual choices are easy to criticize. They can be called distracting, inconsistent, polarizing, unclear, unprofessional, inaccessible, or unnecessary. Familiar choices are harder to attack because familiarity makes them feel correct.

Risk-averse review has a predictable direction. The first draft begins near the center because the center is safe, then the review sands away whatever edges managed to survive.

A strange sentence becomes smoother. A strong claim gains a qualification. A culturally specific image becomes more universal. A joke becomes “more inclusive.” A design with an opinion becomes a design that follows best practices.

Every local revision can sound reasonable while the complete work becomes less worth experiencing.

This is how AI can produce something polished enough to publish and too boring for anyone to remember.

Stop giving AI adjectives

A longer list of adjectives won't solve this.

“Bold, authentic, creative, human, unique, and engaging” describes almost every branding brief ever written. The words sound specific because they express preferences. They remain ambiguous because different people mean completely different things by them.

AI needs evidence of what those words mean to you.

Show it examples you love and explain what produces the reaction. Show it examples you hate and explain why. Tell it which conventions feel exhausted, which risks are welcome, what must remain recognizable, and what is allowed to become strange.

The explanation matters far more than the label.

“I like this” supplies a result.

“I like this because the type looks slightly damaged, the empty space creates tension, and the image refuses to explain itself immediately” supplies a model of taste.

Negative reactions can be just as useful:

I hate this because it looks like a company trying to appear creative without risking anything. The colors are attractive, but every choice feels approved by a committee.

That tells AI far more than “make it less corporate” because it shows what corporate looks like inside your mind and identifies the part of it you actually hate.

You might not know what you want yet

There is another problem here, you may not actually know how to describe your taste before seeing some possibilities.

That's completely normal.

AI can make the differences visible before you know how to produce a perfect specification.

Ask it to explore several directions which make genuinely different choices. Change the tone, audience, risk, structure, and emotional effect between directions.

Then react honestly.

Which one feels alive? Which one feels embarrassing? Which sentence do you keep thinking about? Which direction is technically strong but belongs to somebody else? Which mistake contains something worth preserving?

Explain the reactions, including why the winner won.

Explaining your reactions gives AI something much more valuable than a winner: an updated model of the person it is creating for.

Escape the boiled chicken

Let's actually try this.

Choose a task where AI has previously given you something generic. It might be a paragraph, a presentation, a lesson plan, a logo idea, a story premise, or a homepage.

First, ask for the default answer. Save it.

Then ask the AI what it assumed about you, the audience, the acceptable level of risk, and what a good result should feel like. Some of those assumptions only become visible after the output exists.

Next, ask for 5 directions built around real disagreements. Ask each direction to value something the others sacrifice.

React to the directions in detail. Inconsistency is fine. Taste often becomes visible through reaction long before it becomes available as a theory.

Ask the AI to describe what it now believes about your taste and cite the reactions that support each belief. Correct that model where necessary.

Finally, return to the original task.

Replacing an imaginary generic user with evidence about a real person creates the difference between the first output and the final output.

Let AI actually know you

Useful memory, long conversations, persistent project context, prior drafts, and repeated collaboration allow AI to accumulate evidence about you instead of restarting from zero every single time. It can learn that you value roughness over artificial perfection, that you would rather hold one reader's attention than avoid offending ten others, or that “badass” is a perfectly legitimate design requirement.

Disposable AI use repeatedly starts over with the generic person. A real working relationship can gradually replace that generic person with someone particular.

That model must remain revisable because your taste changes, and you may discover that something you expected to hate works beautifully.

A changing, imperfect model of a real person is still better than permanently assuming that nobody involved has any dangerous preferences.

Change the person AI thinks it is working with

When AI gives you something bland, correcting the output may improve that one result, but correcting AI's model of you improves what it creates next.

Tell it why the bland choice failed. Show it the risks you value. Let several directions expose preferences you could not name in advance. Give it a chance to remember what it learned.

You have to give it enough evidence to imagine you instead.