New EU Guidelines For AI Labelling

 New EU guidelines, why AI sparkles aren’t enough, when AI labels are required, and what the rules mean for AI-powered features and products.

 

There’s been a lot of confusion and panic this week about “huge fines”, “drastic measures” and “sweeping new AI rules” in the EU. In reality, it’s a lot more narrow — and a lot more sensible. And mostly it’s about making AI more obvious when it actually needs to be obvious — especially for AI-generated content.

Starting from Aug 2, 2026, AI labelling is a legal requirement for any company that serves EU citizens. And similar to European Accessibility Act, it’s not limited to EU companies. It affects any company worldwide with EU operations as long as their AI output is used by people in the EU. Let’s see what exactly it means for us.

The EU’s transparency obligations for AI systems took effect on 2 August 2026. Official statement by European Commission.

What Actually Needs Labelling

The goal of AI labelling is to help everyone exposed to AI content to recognize, in a clear and distinguishable way, that the content has been artificially generated or manipulated.

According to Article 50(4) of the AI Act, AI labelling applies to:

1. Deepfakes. Any image, audio, or video that resembles a real person, object, place, or event and would falsely appear authentic or truthful. Content that is not deceptively realistic generally doesn’t apply.
2. Chatbots and AI agents. Users must be informed if they’re not talking to a human.
3. Fully AI-written text. Specifically on matters of public interest, where there has been no human review or editorial work.
4. Emotion recognition and biometric categorization tools.

Both providers (who build or supply the AI system) and deployers (who use it) carry legal obligations. Similar to GDPR and EAA, a company doesn’t escape Article 50 just because it licensed an external AI tool from a third party.

However, it doesn’t mean that all AI-generated content must be explicitly labelled.

Sparkles everywhere in AI products: but they don’t always communicate what exactly is AI-generated, and what isn’t. Source: Chris Joyce, LinkedIn.

Not All AI-Generated Content Must Be Labelled

Beyond the use cases above, pretty much everything else — the vast majority of AI-assisted work — simply isn’t covered by new transparency rules. Most notably, the disclosure obligation does not apply where the AI-generated text has been reviewed and edited by a human, with a named person or entity taking editorial responsibility for it.

Some confusion circles around what exactly “public interest” means, where it starts and where it ends. On its own, it refers to health, safety, environment, economy, finances, politics, science, or culture. If AI-generated product claims touch upon them, the disclosure rule applies.

Some law firms recommend labelling realistic AI-generated illustrations or photos as a precaution for advertising, marketing and other commercial content. AI-generated product illustrations, photos, or posters do need a disclosure, as long as they resemble a real person, place, object, or event.

Carbon’s AI label in context within a complex data dashboard. (Image source: Carbon Design System) (Large preview)

AI label placement examples across different interface components. (Image source: Carbon Design System)

The Fine Line Between “Edited” And “AI-Generated”

But at which point does edited AI content stop being AI content? When a form is pre-filled with AI, but then a user edits it, is it still AI? EU Commission’s guidance is a little fuzzy. Small assistive edits — spellcheck, grammar, formatting, cropping, colour correction, and AI-generated translation — don’t count as AI generation.

AI-generated summaries, composite imagery, substantive rewrites, or adding and removing elements from a photo are considered AI generation. In practice, fine-tuning a sentence a person wrote is fine, but generating the sentence on its own requires a disclosure.

“A human skimmed it before publishing” doesn’t qualify as editorial review. The Commission is explicit that it needs to be substantive, with a named person responsible for the editorial control.

In other words, the fine line lies between intentional manual intervention and automated generation. The latter always has to be disclosed (exception: closed B2B environments).

Carbon AI Label usage variants: inline, icon-only, and explainability panels. (Image source: Carbon Design System) 

AI Sparkles Probably Not Enough

As part of the Code of Practice, the European Commission has published an EU AI icon set. It’s a specific “AI” mark (similar to the AI label in Carbon Design System) — not the generic ✨ sparkle that many products use to signal AI. The signal must be “clear and distinguishable”.

The sparkle might be too ambiguous to signal AI clearly. Mostly because it’s often used to mean “AI-powered feature”, rather than “this specific content was generated by AI”. That’s the kind of signal EU guidelines are trying to rule out.

The EU’s official AI icon set: three variants covering basic AI, fully generated, and partially modified content. (Image source: European Commission)

The Commission is explicit: using an icon “does not establish legal compliance by itself.” A barely visible icon, a note buried in the footer, or a label that flashes for a second are all not compliant.

The icon should be clearly visible, with a plain language label and accessible to assistive technologies. A safe bet is to pair any icon with plain text (“AI-generated”) — and it needs to persist when being reshared or downloaded.

In fact, the EU Commission also published Code of Practice on marking and labelling of AI content.

It Isn’t Just EU

It might feel like a yet another regulation coming from the EU, but in reality there are plenty of other similar regulations that emerged recently worldwide:

1. China has mandatory AI labelling since 1 September 2025. With visible tags and watermarked metadata.
2. California has SB 942, as amended by AB 853, which became mandatory on the exact same day as the EU rules (2 August 2026), deliberately timed to align.
3. South Korea has the AI Basic Act that took effect on 22 January 2026, widely cited as the first comprehensive national-level AI law to mandate deepfake labels. Fines are modest by EU standards (roughly $20K per violation), with a one-year grace period before enforcement bites.
4. India has an IT Rules amendment, in force since 20 February 2026. Platforms must label “synthetically generated information”, and takedown timing for most harmful deepfakes was cut to 3 hours.

Why the sparkle ✨ alone isn’t enough to signal AI-generated content — users need clearer disclosure. (Image source: NNGroup) 

All of these are signs of upcoming AI regulation that looks more like a pattern, rather than a coincidence. So if you’re shipping anything AI this year, it’s probably a good idea to have a conversation about what exactly is going to be AI-labelled, and what not.

Wrapping Up

One final note is that new EU AI transparency rules are much broader than US laws on AI disclosure, where certain state laws require disclosures for synthetic human performers, political advertising or specific AI applications.

None of this really deserves panic or confusion. It’s about a fairly simple idea that has been emerging worldwide at almost the same time:

When AI content could easily be mistaken for human content, creators must say so — in a way that is clear, obvious, and unambiguous. And parts of the UI that are AI-generated must be disclosed as such.

If anything, it will help people distinguish between AI slop and not AI — and everybody can only benefit from that.

 

 

Useful Resources

• Safer and more transparent AI, European Commission’s official announcement
• EU Icons for labelling AI-generated content, the actual icon set and placement rules
• Guidelines on transparency obligations (Article 50), the detailed compliance guidance
• Code of Practice on marking and labelling of AI-generated content
• FAQ: Transparency obligations under Article 50, plain-language Q&A
• The Problem With AI Sparkle Icons, why ✨ is too ambiguous as a disclosure signal (Nielsen Norman Group)
• Carbon Design System: AI Label, a production-ready pattern for clear, accessible AI disclosure

Building Tactile UX: Honoring Intentional Design With Lottie

 When tasked with building a highly interactive, tactile web experience, the architecture must serve the art direction. In this article, it explains their architectural rationale for building a digital stress-relief squeeze toy game using Lottie animations, DOM events, and distance-based math to maintain absolute control over their designers’ intentional motion.

 

When front-end developers and UX engineers are tasked with building a web interface that feels tactile, bouncy, or destructive, the industry instinct is almost always the same: reach for a physics engine. Frameworks like Matter.js, Cannon.js, or custom WebGL solutions have become the gold standard for creating immersive, gamified websites.

When our team at Isadora Agency set out to build Stress Release, a digital stress-relief squeeze toy designed to let burnt-out creatives smash, stretch, and distort animated UI characters, we initially explored that route. The goal was to build a highly tactile experience where every click yielded a satisfying, squishy reaction.

But as we began prototyping, we realized something crucial: Physics engines produce plausible motion, but in our case, the animators produced intentional motion.

We didn’t need our characters to act like realistic rubber balls bouncing uncontrollably around a canvas. We needed them to react in very specific, highly designed ways. So, we scrapped the physics engine entirely.

In this article, we’ll break down how we built a real-time stress-relief squeeze toy without a single line of WebGL or Matter.js, relying entirely on programmatic Lottie state controls, DOM manipulation, and distance-based math.

The Stress Release character shelf introduces players to a collection of animated stress-relief toys, each powered by bespoke Lottie animation states.

The Design Requirements: Intentional Motion

Our core requirement for Stress Release was absolute deterministic control. Our animators had crafted bespoke .json Lottie files that required exact, frame-by-frame sequencing.

For instance, our ‘mega squeeze’ reaction required a precise 181-frame build-up followed by a specific release sequence. To honor this design, we needed an architecture that wouldn’t overwrite the animators’ crafted keyframes with algorithmic approximations.

The tighter the click-feedback loop (click → squish → score), the more you need deterministic frame control. By choosing programmatic state control using Lottie’s native API, we ensured that the interaction layer acted as a flawless trigger for the animation layer.

Character cards showcase the intentionally designed personalities and visual identities that informed each animation sequence and interaction state.

Creating Tactile Feedback: Mapping DOM Elements To Lottie States 

Because our architecture relied on Lottie and the standard DOM, rendering is handled directly by the Lottie runtime, which plays the JSON-based vector animations as SVGs internally. We selected elements directly by ID and CSS class, driving their behavior using a combination of Lottie animation segments, CSS transforms, and click-event math.

To achieve a deeply satisfying “tactile feel” upon hitting a character, we used radial input mapping. The first step was converting the click from page coordinates into the character’s local coordinate space.

Every click was measured against the character’s center point, then translated into score, feedback intensity, and explosion placement:

// Character's center point in its own coordinate space
var x_center = parseFloat($("#playChar").width()  / 2);
var y_center = parseFloat($("#playChar").height() / 2);

// Click position relative to the character's top-left corner
var offset = $("#playChar").offset(); // document-relative position
var X = parseFloat(e.pageX - offset.left);
var Y = parseFloat(e.pageY - offset.top);

// Vector from center to click point
var a = parseFloat(X - x_center);
var b = parseFloat(Y - y_center);

Then we calculate the straight-line distance from the center of the click using the Pythagorean theorem:

var distance = Math.hypot(a, b);

That single number drives everything: the score, the feedback intensity, and where the explosion animation appears:

// Distance zones map to point rewards
if      (distance < 10)  givePts = 100; // bullseye
else if (distance < 40)  givePts = getRndInteger(70, 90);
else if (distance < 70)  givePts = getRndInteger(40, 70);
else if (distance < 100) givePts = getRndInteger(20, 40);
else if (distance < 120) givePts = getRndInteger(10, 20);
else if (distance < 145) givePts = getRndInteger(1,  10);
else givePts = 0; // miss

// Explosion Lottie repositioned to the exact click point
var shiftPosition = window.innerWidth < 1023 ? -20 : 200;
$("#explosionChar").css({
  "margin-left": a + shiftPosition + "px",
  "margin-top":  b + shiftPosition + "px",
});

// Fire the squish animation instantly
explosion.goToAndPlay(0);

The result is a concentric zone system — a perfect circle of scoring rings around the character’s center, similar to a dartboard. The visual complexity of the Lottie SVG is completely irrelevant to hit detection; the hitbox is always a clean circle. Critically, the explosion Lottie animation is repositioned to (a, b) — the same vector used for scoring, so it always appears exactly where the player clicked. This spatial accuracy creates the tactile “I hit that” sensation entirely through math and DOM positioning.

Distance-based click detection and synchronized Lottie reactions create the tactile sensation of physically hitting the character. 

Interaction Handling: Controlling The Narrative

Because the experience used DOM-managed SVG elements, desktop clicks and mobile taps could be handled directly through native event listeners. This avoided extra raycasting or coordinate remapping layers, while keeping the interaction model aligned with how the animations were rendered.

Since the game requires a visual reaction at a specific point, Lottie handles all the squish and bounce feelings internally through its animation curves. Each character has a defined set of animation sections (idle loops, reaction frames, and end states) stored as frame ranges. When a click lands, we jump directly to the exact segment that matches the current game state:

// Animation sections defined as frame ranges per character
const play_segments = [{
  charId: 0,
  sections: {
    idle:     [0,  40],   // looping idle state
    squeeze1: [41, 80],   // light reaction
    squeeze2: [81, 120],  // medium reaction
    squeeze3: [121, 160], // heavy reaction
  },
  playOrder: ["squeeze1", "squeeze2", "squeeze3"],
  endAnimation: [161, 200]
}];

On every click, we advance through the play order and fire the next segment:

function stepAnim() {
  let p         = play_segments[0];
  let i         = p["playOrder"][curr_order_play];
  let playNow   = p["sections"][i];

  playChar.stop();                    // halt current segment immediately
  playChar.loop = false;              // no looping - play once and stop
  playChar.playSegments(playNow, true); // jump to exact frames, force immediately

  curr_order_play++;
  canPlayAnim = 0;                    // lock out further clicks mid-animation

  if (curr_order_play > p["playOrder"].length - 1) {
    curr_order_play = 0;              // cycle back to start of sequence
  }
}

When the segment completes, control returns to the idle loop:

playChar.onComplete = function() {
  canPlayAnim = 1;          // unlock clicks again
  if (!playEnd) playIdleState();
};

function playIdleState() {
  playChar.playSegments([0, 40], true); // return to idle loop
  playChar.loop = true;
}

By triggering precise Lottie animation segments programmatically, the interaction layer maintains deterministic control over every squash and distortion state.

And for the mega squeeze build-up, the bar loops on a specific frame range until triggered:

// Loop the "ready to release" frames until player activates
indikL.loop = true;
indikL.playSegments([181, 302], true);

// On activation - play the release sequence once
indikL.loop = false;
indikL.playSegments([96, 396], true);
indikL.goToAndStop(0, true); // hard reset after completion

The Responsive Benefit Of DOM Elements

Another major factor in our architectural decision was responsive behavior. Because we built Stress Release in the DOM, we bypassed the complexities of scaling bounding boxes and collision vectors across different devices.

We handled responsive resizing entirely through CSS variables. By recalculating CSS custom properties on every resize, the layout simply reacts to the updated variables, and the Lottie SVGs scale naturally inside their containers without losing their state:

const appHeight = () => {
  const doc = document.documentElement;
  doc.style.setProperty("--doc-height", `${window.innerHeight}px`);
  doc.style.setProperty("--doc-width", `${doc.clientWidth}px`);
};
window.addEventListener("resize", appHeight);
appHeight(); // run immediately on init

Mobile Performance Optimization: The Cost Of Lottie

While this architecture gave us total control over the art direction, it introduced a different challenge: file size.

Lottie JSON files can be heavy. We had 21 different character animations, plus multiple explosion variants that all needed to load. To ensure the experience remained fluid — especially on mobile devices — we implemented a few aggressive optimization strategies:

• Connection monitoring
  We tracked initial asset load time using performance.now() to detect slow connections and flag when load times exceeded 5 seconds.
• Sequential asset loading
  Rather than initialising all 21 character animations simultaneously, we load them in pairs using await, advancing only when each pair completes. This prevents a burst of simultaneous network requests and render work from blocking the browser on low-end devices.
• Aggressive memory management
  Instead of keeping our heavy explosion animations in memory, we destroy and recreate them on the fly. This trades a tiny instantiation cost for a much lower idle memory footprint.
• Dynamic quality reduction
  Quality reduction is a single API call applied immediately after each shelf character loads. The key is applying different quality levels depending on the character’s role in the scene:

// Shelf screen - 21 animations playing simultaneously
shelf = lottie.loadAnimation({
  container: document.getElementById("charShelf" + i),
  renderer: "svg",
  loop: true,
  autoplay: true,
  path: "assets/shelf/" + shelfFolders[i] + "/" + shelfFolders[i] + ".json",
});
lottie.setQuality(0.5); // 50% quality - reduces interpolation calculations
shelf.setSpeed(0.6);    // 60% speed - fewer frame calculations per second

// Play screen - single focused character
playChar = lottie.loadAnimation({
  container: document.getElementById("playChar"),
  renderer: "svg",
  loop: true,
  autoplay: true,
  path: chosenChar.url,
});
lottie.setQuality(1); // full quality - only one animation at a time

Conclusion: Choosing The Right Tech For The Design 

When determining the stack for a gamified web experience, it is critical to let the design requirements dictate the technology.

The “Mega Squeeze” state combines layered animation sequences, background transitions, and timed interaction feedback to heighten the sense of impact.

Because our interactions required bespoke, highly controlled visual reactions, we opted for programmatic state control over emergent simulation. This decision empowered the animators to dictate the exact feel of the experience, leaving the code to do what it does best: listen, calculate, and trigger.

By mapping Lottie’s native timeline capabilities to the DOM, you can deliver incredibly rich, tactile user experiences while maintaining absolute control over the art direction.

“

Further Resources

Want to try implementing this yourself, or see exactly how it feels in the browser? Check out these resources:

• Play with the code.
  We have prepared a simplified demo example on CodePen demonstrating a character reacting to a click using playSegments().
• See the final product.
  Check out the live Stress Release site to see all 21 characters and the optimization strategies in action.
• Read the docs.
  Explore the official Lottie Web documentation to learn more about the player controls we utilized. Specifically, explore loadAnimation(), playSegments(), setSpeed(), and setQuality() — the four methods that power the entire interaction layer described in this article.

How Baseline Can Help You Ship Less JavaScript

 

The gap between “you need a library for this” and “the browser does this” keeps closing. A practical guide to auditing your dependencies and finding what the web platform can now handle for you.

Most of us install a dependency once and never look at it again. It does its job, the tests pass, and we move on. But the web platform keeps moving too, and a surprising number of the libraries sitting in your package.json today are now built into the browser.

In a typical mid-sized JavaScript app, you can often find somewhere between 60KB and 90KB (minified and gzipped) of dependencies that the platform can now handle on its own. Date and number formatting, HTTP requests, modals, tooltips, deep cloning, grouping arrays: these were all real gaps a few years ago. A lot of them aren’t gaps anymore.

The reason those libraries stick around isn’t laziness. It’s that most teams don’t re-audit their dependencies on a Baseline cadence, or are simply not aware of how fast browsers are shipping these days. You check npm audit for security, but is this library still doing something the browser can’t? is a question that rarely gets asked. So the libraries stay.

In this article, we’ll run that audit together. Instead of going through dependencies one by one, we’ll work in clusters, because the wins tend to come in groups. We’ll do the bundle math, build a small decision framework you can reuse, and stay honest about the cases where the platform still falls short. By the end, you’ll have a repeatable process you can run on your own package.json.

What “Baseline” Actually Means 

Before we start deleting things, let’s quickly recap what Baseline is. Feel free to skip this section if you’re already familiar.

Baseline is a project from the WebDX Community Group that tells you, in plain terms, how safe a web feature is to use across the major browsers (Chrome, Edge, Firefox, and Safari). A feature can be in one of three states:

• Limited availability
  The feature hasn’t shipped in all the major engines yet. Not safe to rely on without a fallback.
• Baseline Newly available
  The feature has just landed in all the major engines. It works for users on up-to-date browsers, but older devices in the wild may not have it yet.
• Baseline Widely available
  The feature has been in all the major engines for 30 months. At this point, you can reach for it without much thought.

That 30-month gap between “Newly” and “Widely” matters a lot for this audit. A feature that’s Widely available is something you can usually drop a library for today. A feature that’s only Newly available is something you can drop a library for if you check your audience first, or if you’re comfortable with a small feature check. We’ll treat those two cases differently throughout.

You can look any feature up on webstatus.dev, on MDN (every reference page shows a Baseline badge near the top), or programmatically with the web-features npm package. We’ll use all three later when we run the audit on a real project.

A Decision Framework Before You Delete Anything

It’s tempting to read “the browser does this now” and start ripping libraries out. Let’s not do that. A swap that looks free on paper can quietly break things for a chunk of your users, or cost you a feature you were relying on without realizing it.

So before dropping any library, ask three questions. We’ll reuse these in every cluster below.

1. Is the replacement Baseline-safe for my audience?

Not “is it Baseline” in the abstract, but “is it safe for the people who actually use my app.” If the native feature is Widely available, this is usually a yes. If it’s only Newly available, check your analytics or your browserslist config and see what share of your users would miss out. A B2B dashboard where everyone’s on the latest browser is a very different situation from a public-facing site with a long tail of old Android devices.

2. What does the swap actually cost?

Dropping a library isn’t always free. Sometimes the native feature isn’t supported widely enough yet, so you’d reach for a polyfill. If that polyfill is heavier than the library you’re removing, you’ve made your bundle bigger, unless you load it conditionally. We’ll see exactly this with Temporal later.

3. Does the platform feature cover my real use case?

Libraries often do more than the platform feature they resemble. axios isn’t just fetch with automatic JSON parsing; it has interceptors, request cancellation, and retries. If you’re using those, a straight swap to fetch will leave you reimplementing them. Check what you actually use before assuming it’s a drop-in replacement.

Keep these three in mind. Every cluster below is really just these questions applied to a different corner of your dependencies.

Cluster 1: Internationalization (The Biggest Drop Today Win) 

This is the cluster where you’ll usually find the most KBs sitting on top of features that are already Widely available. The browser ships a whole family of formatting tools under the Intl namespace, and a lot of small, popular libraries became unnecessary.

Here are the usual suspects and what replaces them:

• timeago.js (1 KB gz) → Intl.RelativeTimeFormat
• pluralize (2.3 KB gz) → Intl.PluralRules
• numeral (3.9 KB gz) → Intl.NumberFormat
• humanize-duration (6.6 KB gz) → Intl.DurationFormat
• list-joining helpers → Intl.ListFormat

Let’s walk through some of them.

Relative Time

timeago.js exists to turn a timestamp into “3 hours ago”. Intl.RelativeTimeFormat does the same thing, and it’s Baseline Widely available.

const rtf = new Intl.RelativeTimeFormat("en", { numeric: "auto" });

rtf.format(-1, "day"); // "yesterday"
rtf.format(3, "hour"); // "in 3 hours"
rtf.format(-2, "week"); // "2 weeks ago"

The numeric: "auto" option is the nice touch here: it gives you “yesterday” instead of “1 day ago” where the language has a word for it. You pass a number and a unit, and you get a localized string back.

You may be wondering about the one thing timeago.js does that this snippet doesn’t: it picks the unit for you. Given a date, timeago.js decides whether to say “seconds” or “days.” Intl.RelativeTimeFormat expects you to do that part. It’s a few lines of arithmetic (work out the difference, find the largest unit that fits), and once you’ve written that helper, you don’t need the library anymore.

Numbers, Currency, And Lists

Intl.NumberFormat covers most of what number-formatting libraries do: thousands separators, currency, percentages, and compact notation.

new Intl.NumberFormat("en-US").format(1234567.89);
// "1,234,567.89"

new Intl.NumberFormat("en-US", { style: "currency", currency: "USD" }).format(
  1234.5,
);
// "$1,234.50"

new Intl.NumberFormat("en", { notation: "compact" }).format(1200000);
// "1.2M"

And Intl.ListFormat, Widely available, handles the “join an array into a sentence” problem, including the Oxford comma, which is the kind of thing people write fiddly helper functions for:

const lf = new Intl.ListFormat("en", { style: "long", type: "conjunction" });

lf.format(["Alice", "Bob", "Carol"]);
// "Alice, Bob, and Carol"

The One Caveat: Durations

humanize-duration turns a number of milliseconds into “1 hour, 30 minutes”. The platform equivalent is Intl.DurationFormat:

const df = new Intl.DurationFormat("en", { style: "long" });

df.format({ hours: 1, minutes: 30 });
// "1 hour, 30 minutes"

One thing to keep in mind is that Intl.DurationFormat is Baseline Newly available at the time of writing, not Widely available. It landed in all the major engines in March 2025, and it’s on track to become Widely available in 2027. So this one fails question 1 for broad-audience apps unless you check your traffic first or add a fallback. For an internal tool on modern browsers, it’s fine today. For a public site with old devices, give it another year or guard it with a feature check.

The Math On This Cluster

If your app uses the full set (humanize-duration, timeago.js, pluralize, numeral), that’s roughly 14 KB gzipped of dependencies, most of it replaceable right now with Widely available APIs. The internationalization cluster is usually the easiest win in the whole audit.

 

Cluster 2: HTTP Clients

This cluster is more nuanced, so it’s a good one to slow down on.

The browser HTTP libraries people reach for are axios (17 KB gz) and superagent (19 KB gz). For most requests, fetch plus AbortController covers what you need, and both are Widely available.

A basic GET looks like this:

// axios
const { data } = await axios.get("/api/users");

// fetch
const res = await fetch("/api/users");
const data = await res.json();

The one extra line (res.json()) is fetch being explicit where axios was implicit. That’s the pattern across this whole cluster: fetch does less for you by default, and you decide whether you want the things it leaves out.

Timeouts #

axios has a timeout option. fetch has AbortSignal.timeout():

const res = await fetch("/api/users", {
  signal: AbortSignal.timeout(5000), // abort after 5 seconds
});

Where fetch Doesn’t Replace axios

This is where question 3 does most of the work, so let’s be specific about the gaps:

• fetch doesn’t reject on HTTP errors.
  A 404 or 500 is a resolved promise, not a rejection. You have to check res.ok yourself. axios rejects on any non-2xx status.
• No interceptors.
  If you rely on axios interceptors to attach auth tokens or handle 401s in one place, fetch has no equivalent. You’d wrap fetch in your own function or class to get the same behavior.
• No automatic retries.
  axios (with a plugin) can retry failed requests. With fetch, that’s your code to write.
• No upload progress.
  fetch still can’t report upload progress in a first-class way. If you have a file uploader with a progress bar, that’s a real reason to keep a library.

I personally heavily rely on interceptors in my interactive online courses, such as Learn JavaScript, and I have solved that for years using a custom class on top of fetch. I’ve shipped this to millions of users and have seen lots of success with it.

None of these are hard to rebuild, and most apps only use one or two of them. But this is exactly the kind of cluster where you shouldn’t do a blind find-and-replace. Look at how you actually use your HTTP client first. If it’s plain GETs and POSTs, dropping axios for a thin fetch wrapper saves you about 17 KB gzipped.

Cluster 3: UI Primitives

This cluster has some of the most satisfying swaps, because the platform features don’t just match the libraries, they’re often more accessible than what teams ship by hand.

The libraries here are modal dialogs (something like a11y-dialog, 1.8 KB gz), tooltip and popover libraries (tippy.js, 14 KB gz, which bundles Popper for positioning), focus-trap (6.6 KB gz), and body-scroll-lock (1.3 KB gz). They get replaced by three platform features: the <dialog> element, the Popover API, and CSS anchor positioning.

The <dialog> Element

A huge amount of modal-related code exists to solve accessibility problems: trapping focus inside the modal, closing on Escape, restoring focus to the previous element when the dialog is closed, and rendering above everything else. The <dialog> element, Widely available, does all of that for you.

<dialog id="confirm">
  <form method="dialog">
    <p>Delete this file?</p>
    <button value="cancel">Cancel</button>
    <button value="delete">Delete</button>
  </form>
</dialog>

const dialog = document.querySelector("#confirm");

dialog.showModal(); // focus moves in, background goes inert, Escape closes it

dialog.addEventListener("close", () => {
  console.log(dialog.returnValue); // "cancel" or "delete"
});

Calling showModal() does the work that focus-trap was installed for: focus moves into the dialog, the rest of the page becomes inert so you can’t tab out of it, Escape closes it, and focus returns to the element that opened it. The dialog renders in the browser’s Top layer, so you don’t fight z-index. You also get a ::backdrop pseudo-element to style the overlay.

That single element can replace your modal library and focus-trap. The one piece it doesn’t handle on its own is locking the background from scrolling, which is what body-scroll-lock was for. That’s now one line of CSS:

body:has(dialog:modal) {
  overflow: hidden;
}

If you’re wondering why we’re using dialog:modal instead of dialog[open], it’s because the open attribute is set as soon as you call show(), but the dialog isn’t actually modal so you don’t want to lock scrolling yet. The :modal pseudo-class is only true when the dialog is actually modal, which is the case when you call showModal().

So three libraries collapse into one element and one CSS rule.

Popover API And Anchor Positioning

For things that aren’t full modals (dropdown menus, tooltips, the small floating panels that tippy.js handles), the Popover API gives you light-dismiss behavior, top-layer rendering, and Escape-to-close with no JavaScript at all:

<button popovertarget="menu" id="options">Options</button>

<div id="menu" popover>
  <!-- menu content -->
</div>

Clicking the button toggles the popover. Clicking outside it closes it. It’s Baseline Newly available (since January 2025).

The other half of what a tooltip library does is positioning: keeping the floating element pinned to its trigger and flipping it when it would overflow the viewport. That’s what Popper (bundled inside tippy.js) handles, and it’s now a CSS feature called anchor positioning. Here it pins the same #menu popover directly under its trigger button:

#options {
  anchor-name: --trigger;
}

.tooltip {
  position-anchor: --trigger;
  position-area: top;
  margin: 0;
}

Anchor positioning is the newest feature in this article. It became Baseline Newly available in January 2026, when Firefox 147 shipped it (Chrome had it since version 125, and Safari since version 26). Because it’s this fresh, it’s squarely a question-1 feature: great for modern audiences, but check your traffic, and note that some of the more advanced parts (like position-try fallbacks) have uneven support across versions. Keep a sensible fallback for older browsers.

Between <dialog>, the Popover API, and anchor positioning, the UI primitives cluster (tooltip library, modal library, focus-trap, body-scroll-lock) adds up to roughly 24 KB gzipped, and you come out the other side with better accessibility defaults than most hand-rolled solutions.

Cluster 4: Lodash Utilities 

Lodash is rarely imported whole anymore, but its individual functions show up everywhere, either as the full lodash package (25 KB gz) or as standalone installs like lodash.clonedeep and lodash.groupby. Several of the most common ones now have direct platform equivalents.

Grouping

lodash.groupby reorganizes an array into an object keyed by some property. Object.groupBy does exactly that:

const products = [
  { name: "Apple", category: "fruit" },
  { name: "Carrot", category: "vegetable" },
  { name: "Banana", category: "fruit" },
];

const grouped = Object.groupBy(products, (product) => product.category);
// {
//   fruit: [{ name: "Apple", ... }, { name: "Banana", ... }],
//   vegetable: [{ name: "Carrot", ... }],
// }

There’s also Map.groupBy for when you want a Map instead of a plain object (handy if your keys aren’t strings). Both are Baseline Newly available, since March 2024, and on track to become Widely available in late 2026.

Deep Cloning 

lodash.clonedeep makes a deep copy of an object. structuredClone is the platform version, and it’s Widely available:

const original = { user: { name: "Sam", roles: ["admin"] } };

const copy = structuredClone(original);
copy.user.roles.push("editor");

original.user.roles; // ["admin"] (unchanged)

structuredClone handles the tricky cases that trip up JSON.parse(JSON.stringify(...)): it clones Date, Map, Set, ArrayBuffer, and circular references correctly. The limit to know about (question 3 again) is that it can’t clone functions, DOM nodes, or class instances; it throws on functions and drops the prototype on class instances. For plain data, which is what most people deep-clone, it’s a clean replacement.

Set Operations

If you’ve ever pulled in a Lodash helper for union, intersection, or difference, the Set object now has these built in. They’re Baseline Newly available, since June 2024:

const admins = new Set(["sam", "alex", "jo"]);
const editors = new Set(["alex", "kim"]);

admins.intersection(editors); // Set { "alex" }
admins.union(editors); // Set { "sam", "alex", "jo", "kim" }
admins.difference(editors); // Set { "sam", "jo" }

The full set of methods is union, intersection, difference, symmetricDifference, isSubsetOf, isSupersetOf, and isDisjointFrom.

What’s Worth Keeping 

Not all of Lodash has moved into the platform. debounce and throttle still have no native equivalent, and they’re genuinely useful, so cherry-picking lodash.debounce is reasonable. The point of this cluster isn’t “delete Lodash,” it’s “stop shipping the parts the browser already has.” Dropping lodash.clonedeep and lodash.groupby alone is about 8 KB gzipped, and if you were importing the full lodash for a handful of functions, replacing the platform-covered ones can let you drop it entirely.

Cluster 5: Temporal, A Case Study In Not Dropping A Library Yet

Every cluster so far has ended in “go ahead, drop it.” This one is the opposite, and that’s why it’s worth including: it shows the framework telling you to wait.

Temporal is the long-awaited replacement for JavaScript’s Date, and it’s a genuinely better API: immutable objects, sane time zone handling, and no more month indexes starting at zero. It reached TC39 Stage 4 in March 2026 and is part of the ES2026 specification. Firefox shipped it in version 139 (in 2025), and Chrome shipped it in version 144 (January 2026). Safari hasn’t shipped it in a stable release yet; it’s in Safari Technology Preview, with stable support expected later in 2026.

If Temporal is news to you, check out the Temporal Cheatsheet for a quick overview of the API and a comparison to Date.

However, Temporal is not Baseline. It’s still in limited availability, because Safari users don’t have it. To use it across all browsers today, you need a polyfill, and this is where the math turns against you.

The official @js-temporal/polyfill is about 44 KB gzipped. There’s a smaller polyfill that internally does not depend on BigInt and it weighs 19 KB gzipped. A lightweight date library like dayjs is about 3 KB gzipped. So if you swap dayjs for Temporal plus its polyfill right now, you’re not saving 3 KB, you’re adding roughly 41 KB to your bundle, unless you are able to load the polyfill conditionally.

Run it through the framework:

• Question 1 (audience): Temporal isn’t Baseline. For a broad audience, that’s a lot of people.
• Question 2 (cost): the polyfill is more than ten times the size of the library you’d remove. The swap makes your bundle bigger.
• Question 3 (feature gap): Temporal actually wins here; it does more than dayjs. But that doesn’t matter while questions 1 and 2 are failing.

The verdict is generally clear: keep dayjs (or date-fns) for now. The moment to revisit is when Safari ships Temporal in a stable release and it reaches Baseline. At that point you can use Temporal natively and conditionally load the polyfill for users on older browsers. This is a feature to write down and check again in a few months, not one to act on today.

How To Run This Audit On Your Own package.json 

The clusters above are a starting map, but your dependencies are your own. Here’s a repeatable process you can run this quarter.

Step 1: List Your Production Dependencies 

Start by listing what actually ships to users:

npm ls --omit=dev --depth=0

Step 2: Measure What Each One Costs 

For a quick per-package number, Bundlephobia gives you the minified and gzipped size of any npm package. For the real picture (what each dependency costs in your actual bundle, after tree-shaking and deduplication), run a bundle analyzer against your build. npx source-map-explorer works on most bundles, and npx vite-bundle-visualizer works for Vite projects.

Step 3: Check The Baseline Status Of Each Replacement 

For each candidate, find the platform feature that would replace it and check its Baseline status. The quickest way is webstatus.dev or the Baseline badge on the feature’s MDN page.

Step 4: Run The Three Questions

For each library with a platform replacement, go back to the framework: Is it Baseline-safe for your audience? What does the swap cost? Does the feature cover how you actually use the library? Most of your decisions will fall out of question 1 (check the feature’s status against your browserslist) and question 3 (check your own usage).

Step 5: Swap Behind Progressive Enhancement Where Needed #

For Widely available features, swap and move on. For Newly available ones, either confirm your audience is on modern browsers or guard the new code with a quick feature check and keep a fallback:

if (typeof Intl.DurationFormat === "function") {
  // use the platform feature
} else {
  // fall back to the library, or a simpler format
}

That way you ship less code to the users who can run it, without breaking the ones who can’t.

Wrapping Up

Add the clusters up, and the picture is concrete. The internationalization cluster is around 14 KB gzipped, HTTP is around 17 KB, the UI primitives are around 24 KB, and the Lodash utilities are 8 KB or more depending on how much of the library you were shipping. For a typical mid-sized app, that’s somewhere between 60 KB and 90 KB gzipped of dependencies you can hand back to the platform, and more if you were shipping the full lodash or several of these libraries at once. (The uncompressed numbers are two to three times larger, which is what you’ll see in a bundle analyzer before gzip.)

I’ve chosen relatively lean packages for most of these features, but some individual packages could still be heavy. Your dialog package, for instance, could alone weigh as much as 50KB gzipped depending on what you’re using.

A few features are worth keeping an eye on over the next year, because they’ll open up further swaps:

• Temporal going native.
  Once Safari ships it in a stable release and it reaches Baseline, you can drop both your date library and the polyfill, turning today’s regression into a real win.
• CSS anchor positioning maturing.
  It became Baseline Newly available in January 2026. As it ages toward Widely available, dropping tooltip and popover positioning libraries gets safer for broad audiences.
• Object.groupBy and friends crossing into Widely available.
  The 2024 batch (array grouping, Set methods) is on track to become Widely available in late 2026, which moves them from “check your audience” to “just use it.”

None of this is a one-time cleanup. The platform ships new features constantly, and the gap between “you need a library for this” and “the browser does this” keeps closing. The habit worth building is small: once a quarter, run the audit. List your dependencies, check what’s now Baseline, and hand back what you can.

Pick one cluster from this article, open your package.json, and see how much of it the browser already does for you.