ExpertsFMHow To Avoid AI Design Slop
Y Combinator Startup Podcast · March 6, 2026
5 Selected Lessons:
heuristic · 3 reasoning_pattern · 1 priority · 1
What AI frees you to focus on vs. what it shouldn't replace
Source
Y Combinator Startup Podcast — How To Avoid AI Design Slop (March 6, 2026) · at 25:39 ▸ · at 33:26 ▸
When this applies
when a new tool removes a previously time-consuming step in your workflow
The takeaway
When a tool removes friction from execution, redirect the freed capacity toward the judgment-heavy questions the tool can't answer, not toward more unexamined output.
Why it works
AI removes the cost of implementation, but the core strategic questions (what are we making, for whom, why is it valuable) still require human judgment; if the freed time is spent accepting more AI flourishes instead of answering those questions, the real opportunity is wasted.
One level up
The value of automation is realized only if the saved effort is reinvested in higher-order thinking, not consumed by lower-order busywork.
Where it stops holding
This stops working when there genuinely is no higher-order question left to think about (e.g., a purely technical, back-office tool where visual polish is the entire value proposition).
Detail
hierarchy:
- 1. Use AI to handle technical/implementation details
- 2. Spend the freed time on 'what are we making, for whom, why is it valuable' — the hard strategic questions
- 3. Only then use AI to polish/iterate on visuals, and evaluate output critically rather than accepting it wholesale
rationale: The cool aspect of AI tools is that it frees you from fiddling with technical details so you can work on the hard questions of your offering — but if you just use it to create AI slop, that power gets wasted.
In their words
Look, I think the really cool thing about AI tools is that it kind of frees you from having to kind of like, you know, fiddle with like the technical details. So it can really work on the hard, hard kind of questions of your offering, of your product, of your company, such as like, what are we making? For whom? Why is this valuable to them?
LLM-driven trend homogenization
Source
Y Combinator Startup Podcast — How To Avoid AI Design Slop (March 6, 2026) · at 0:00 ▸
When this applies
when a tool trained on 'what's already popular' is used by many competitors simultaneously, causing rapid convergence on the same patterns
The takeaway
Any tool that learns from and redistributes 'what's currently successful' will accelerate convergence and erode the differentiation that early adopters once had.
Why it works
LLMs are trained on examples that got linked/cited a lot; once one good site establishes a trend, it gets absorbed into the model and redistributed to every user almost instantly, collapsing the normal lag time that let trends stay rare.
One level up
Shared infrastructure that encodes 'best practice' tends to commoditize the very advantage that created the practice.
Where it stops holding
This stops working as an explanation when a trend is driven by genuine functional necessity rather than aesthetic fashion — not every repeated pattern is slop.
Detail
approach: Notice a repeated aesthetic across many unrelated products built with the same AI tools; trace it back to the tool's training data rather than independent convergent taste.
signals:
- sudden simultaneous appearance of an effect across many sites
- effect historically too costly/rare to implement by hand
rationale: Trend propagation that used to take years via human copying now happens in a week because the model itself encodes and redistributes the winning examples.
In their words
But now with LLMs, if there's a good website with a purple gradient, it makes it into the LLM because the LLM gets trained on the good examples that get linked to a lot. And then all of a sudden, the next week, all the startup websites look the same.
Hover effects should invite clicking, not hide it
Source
When this applies
when adding any interactive affordance to a UI — the affordance should amplify the existing 'this is actionable' signal, never weaken or obscure it
The takeaway
Any added embellishment to an interactive element should reinforce its core affordance (make it feel more actionable), never contradict or obscure the baseline signal users already rely on.
Why it works
Browsers already signal clickability via cursor change; an effect that fades or dims the element contradicts that built-in signal and creates cognitive dissonance about whether the item is actionable.
One level up
New capabilities should be layered on top of proven defaults, not used to override them.
Where it stops holding
This stops working when the element is deliberately meant to be de-emphasized or non-interactive, where fading is the intended, honest signal.
Detail
rule: Make hover effects make things pop/feel more clickable (lighter shade, slight glow), never make them fade or disappear.
threshold: If the effect contradicts the default browser cursor-hand signal, it's wrong.
rationale: The browser already tells users 'this is clickable'; added CSS should amplify, not negate, that signal.
In their words
So if you want to add a CSS effect, then make it maybe like a little pop. Maybe make it one shade lighter or more active or add like a little, you know, a little glow around it, but don't make it go away.
'Possible' is not 'worth doing' — the AI-ease test
Source
Y Combinator Startup Podcast — How To Avoid AI Design Slop (March 6, 2026) · at 9:19 ▸ · at 13:28 ▸ · at 25:39 ▸ · at 33:26 ▸
When this applies
when a new capability makes something previously expensive now free — deciding whether to actually use it
The takeaway
When a cost barrier that used to filter out low-value additions disappears, you must replace it with a deliberate value-based filter, or you will accumulate unvetted additions.
Why it works
Because AI tools make elaborate effects nearly free to generate, builders accept them by default rather than asking whether the effect serves the user's goal, since the old cost-based filter ('is this worth a week of engineering?') no longer applies.
One level up
Removing friction from production doesn't remove the need for editorial judgment — it increases it.
Where it stops holding
This stops working when the free/easy addition genuinely reinforces the core message or brand (e.g., the game controller hover animation that supported the product's narrative).
Detail
rule: Ask 'would I build this if it cost a week of engineering time?' If no, don't ship it just because AI made it free.
threshold: Distracting, attention-stealing, or value-free effects should be cut regardless of implementation cost.
rationale: There's a lot of things now at your fingertips with AI that, anything imaginable is possible, but just because something is possible doesn't mean you should say yes to it.
In their words
There's a lot of things that are now at your fingertips with AI that, anything imaginable is possible, just because something is possible doesn't mean you should say yes to it.
Scroll-jacking and timer-based fade-ins break usability
Source
Y Combinator Startup Podcast — How To Avoid AI Design Slop (March 6, 2026) · at 13:28 ▸ · at 16:43 ▸ · at 22:26 ▸ · at 30:27 ▸
When this applies
when a product overrides a platform's native, well-understood interaction (scrolling, navigation) with custom logic to show off a feature
The takeaway
Don't override a platform's native, trusted interaction model to showcase a custom animation; the showcase cost (confusion, lost orientation) usually outweighs the visual payoff.
Why it works
Users already get motion from scrolling itself; adding scroll-jacking or timed fade-ins fights the user's own input, causes missed content when scrolling fast, and removes the scroll bar's function as a 'how far through this am I' indicator.
One level up
Respecting existing user mental models is usually more valuable than novel flourishes.
Where it stops holding
This stops working when the custom scroll behavior is used sparingly and clearly communicates something the content itself can't (e.g., Sphinx's orange line arguably mirroring the product's workflow UI) — but even then it risks overshadowing the message.
Detail
rule: Avoid scroll-jacking; avoid timer-based fade-ins that can be scrolled past before fully rendering.
threshold: If a user can lose track of 'how far through the page they are,' the scroll indicator has failed and the pattern should be cut.
rationale: It still feels like going through molasses because it's all scroll jacking — hijacking native browser scrolling to do some fancy thing with JavaScript.
In their words
Here, it's kind of like, I don't know, maybe it feels like we're halfway through, but maybe you know it's gonna keep me for another 15 seconds on some like really rich build out of a landing page. So the scroll indicator is a really important tool, I think, to get a grasp on the landing page. And here it's completely failing me.