When AI Disappears & When It Shouldn't

On invisible AI, visible AI, privacy, and what products are getting wrong about all three.

I have been using Naukri.com to find a job.

I uploaded my resume. Filled in my preferences. Explicitly told the platform what kind of roles I want, what industries interest me, what level I am at. The system has more information about my career than most people I have met in person.

And yet, every day, it sends me AI-generated job suggestions that feel completely disconnected from everything I told it.

Recently, Naukri introduced a feature that allows "AI to apply to jobs on my behalf." I did not use it. Not because I am resistant to automation. But because something felt off. And it took me a while to figure out exactly what.

This blog is my attempt to think through that feeling, as a product designer, and as someone who is increasingly thinking like a product manager.

Before I get back to Naukri, I want to set up a frame.

In the last two years, two of the most consequential AI product decisions were made by Apple and Google. They were building in the same space,  on-device AI that understands your context, respects your privacy, and makes your phone smarter. But they arrived at completely opposite answers about how AI should present itself to a user.

Understanding why they made those choices, and what those choices look like on an actual screen, is useful for anyone building a product with AI right now.

Apple: Designing AI You Cannot See

Apple's guiding principle with Apple Intelligence is deceptively simple, the best interface for AI is no interface at all.

Everything is processed on the device, inside Apple's own silicon chips. Your data does not travel to a server. There is no cloud round-trip, no loading state, no moment where your information leaves your hands. But the more interesting design decision is not the privacy architecture. It is what Apple chose to show the user.

Almost nothing.

What this looks like on screen

When your notifications are summarized, there is no "AI Summary" banner with a sparkle icon. There is just a shorter, cleaner notification stack. The reduction in clutter is the feature. The AI is the plumbing.

When Writing Tools appears, the feature that lets you rewrite, summarize, or change the tone of any selected text, it surfaces through the same contextual menu you have always used for copy and paste. No new button. No new screen. An existing interaction quietly made more capable.

And then there is Call Screening, one of the most quietly radical things Apple has shipped.

When an unknown number calls, your iPhone does not ring. It silently answers on your behalf. A Siri voice asks the caller to state their name and reason for calling. As they speak, a real-time transcription appears on your screen. You read why someone wants to talk to you, decide if it is worth your time, and choose to answer or decline, all before your phone has made a sound.

Everything is processed on device. No audio leaves your phone.

Notice what Apple chose not to do here. No confidence scores. No "this call is 73% likely to be spam" gauge. The AI handled the interaction. The judgment is yours.

This is what designers call ambient UX, technology so embedded in the environment that it stops feeling like technology. The interface does not change. Your mental model of the product does not change. It just starts doing more.

Why Apple made this choice, and what it costs

Invisibility is Apple's brand. Privacy, simplicity, and seamlessness are not features for Apple, they are the product. On-device processing is a statement: your data never leaves this device, and you do not need to think about it.

From a design perspective, invisibility reduces anxiety. Job seekers, patients, anyone using a product in a high-stakes context, they do not want to feel surveilled. The moment a product announces "AI is watching," users start wondering what exactly it is watching for. Invisibility removes that question before it is asked.

But there is an honest cost. When AI is completely hidden, users cannot learn from it. They cannot understand why the system made a particular decision. When something goes wrong, when the summary is wrong, when the suggestion is off, there is no visible thread to pull. The opacity that creates trust in normal conditions can create confusion in edge cases.

Apple also kept its AI models closed to outside developers. Third-party apps work within predefined templates, not direct access. Consistent experience across the platform, but at the cost of what builders can create.

The question Apple is betting on: do users want to feel AI, or do they want to not feel it at all?

Google: Designing AI You Can See

Google answered the same question differently.

With Gemini Nano embedded into Android through a system service called AICore, Google also chose on-device processing, the privacy foundation is similar to Apple's. But unlike Apple, Google made its AI visible by design and open by design.

What this looks like on screen

When you are reading a text that mentions a restaurant, a small contextual chip fades in at the bottom of the screen. It does not interrupt what you are doing. It just sits there, a gentle offer: open in Maps, make a reservation. You tap it or ignore it. Either way, the primary task continues.

When Google summarizes your notifications, it does not just shorten them invisibly. It surfaces a sentence with a light AI indicator, subtle, but present. You know the system did something. You can expand if you want to understand how.

And then there is Scam Detection, which is, deliberately, nothing like either of those.

Powered by Gemini Nano running entirely on the device, Scam Detection listens to active phone calls in real time. It analyses conversational patterns, phrases like "urgent transfer," "your account has been compromised," "you must act now", that are statistically associated with fraud. Not caller ID. Not a blocklist. The actual content of the conversation processed locally.

When it detects a likely scam, it does not send a subtle notification. It triggers a simultaneous audio beep, haptic vibration, and a loud high-contrast warning that occupies a significant portion of your screen: Likely Scam. You must actively acknowledge it to continue the call.

Compare this to the restaurant chip, a peripheral whisper. Scam Detection is a full interruption, designed to be impossible to miss. Same platform, same AI, completely different visual intensity, and the reason is deliberate. A missed restaurant suggestion costs you nothing. A missed scam warning can cost you everything.

This is the design principle Google embedded into every AI decision: the visibility of AI should be proportional to the consequence of the action it is influencing. Low stakes, stay quiet. High stakes, be loud.

Why Google made this choice, and what it costs

Google's philosophy has always leaned toward explicitness. Android gives users more control, more visibility into what the system is doing. Visible AI is consistent with that. And opening Gemini Nano to developers through the Android AI Edge SDK is a strategic bet, any developer can now build AI-powered features without cloud API costs, so the platform gets smarter as more builders use it.

But visibility has a cost. Every chip, every overlay, every indicator is one more thing on the screen. Done poorly, visible AI is clutter. It interrupts instead of assists. Google is betting its designers can calibrate visibility carefully enough that it feels like control rather than noise.

The question Google is betting on: do users want to understand what AI is doing, or do they want it to just work?

Privacy: The Design Decision Everyone Is Avoiding

It is worth pausing here because both Apple and Google made the same foundational infrastructure decision, and both chose to tell users about it explicitly. Not buried in terms-of service. In the product itself.

Apple's framing is brand-level: your data never leaves this device appears on product pages, in keynotes, in advertising. The privacy architecture is the product story.

Google's framing is contextual: when you enable Scam Detection, a visible notice in that settings screen tells you all processing happens in a secure, isolated environment on your device — no audio stored, nothing sent to servers. The disclosure is in the right place at the right moment.

Both understood something most product teams are still working out: privacy is not a legal checkbox. It is a user experience decision that affects how the product feels every time someone uses it.

When AI is running in the background using your data in ways you cannot see or understand, users feel it even if they cannot name it. Something feels off. The product feels like it is working for someone else.

Which brings me back to Naukri.

What Naukri Is Actually Doing, And Why It Feels Wrong

Naukri has shipped a meaningful amount of AI in the last year. AI Resume Maker. AI mock interviews. Skills insights. AI REX, an agentic platform that automates talent sourcing for recruiters. This is a company clearly moving on AI.

But as a designer, when I look at what has been built, I see one consistent pattern: every AI feature is a separate destination, and the quality of execution reflects that.

The AI Resume Maker is a tool you navigate away to, and it sits behind a paywall. When you get there, the output is a poorly formatted document that needs significant manual editing to be usable. Stretched profile picture, badly placed text. There is nothing AI about it. It is a resume template builder, you put information in, it puts it in a box.

For a feature asking job seekers to pay, built specifically to help them make a strong first impression, the artifact it produces would make most designers uncomfortable.

The feed looks the same. The profile works the same. The application flow is unchanged from what it was years ago.

AI has been placed on top of the product as a layer of optional tools. The underlying experience has not become more intelligent; a set of adjacent features has been built around it.

The auto-apply feature is where this becomes most visible as a design problem.

The promise is straightforward, AI applies to jobs on your behalf, in a few clicks. From a metrics perspective this looks like a win: applications sent, friction reduced, time saved. On a roadmap, it looks like an AI feature shipped.

But here is what actually happens.

You get a notification. You open the app. A chatbot-style window appears showing you the job. You tap apply. The system sends something to the employer and marks it as applied.

This is not automation. This is the same application flow with extra steps and an AI label on top. The friction was not removed, it was redistributed into a notification you have to respond to, a window you have to open, a button you still have to press.

And at the screen level, the more fundamental problem: a high-stakes action, submitting a job application in your name, to a company, with your resume, has been given the same visual weight as dismissing a notification. There is no moment that shows you what is being sent, to whom, or why the system selected this role for you. The design treats it like confirming a food order.

Compare this to Google's Scam Detection. Google identified a high-stakes moment, a potentially fraudulent call and deliberately made the UI interrupt. Full screen. Audio beep. Haptic vibration. You cannot continue without active acknowledgement. The stakes demanded unavoidable design.

Naukri's auto-apply does the opposite. It reduces the perceived weight of a decision that deserves weight.

And underneath all of this is the matching problem. If the AI were genuinely precise, if it understood your profile, your trajectory, your intent, this flow might be worth the trade-off. But when the system routinely suggests the wrong roles at the wrong seniority level, dressing that up as automation does not fix the precision problem. It just gives it a new name.

To be fair: this pressure is not unique to Naukri. Every product team right now faces the same incentives, ship AI visibly, ship it fast, have something to announce. The slower work of making AI foundational rather than decorative does not fit neatly into a quarterly roadmap. But that slower work is exactly what separates a product that earns trust from one that earns frustration.

What AI-Native Job Search Could Look Like

I want to think out loud here rather than prescribe, the teams building in this space are navigating constraints I do not fully see from the outside.

But when I think about what I actually want from Naukri, not as a critic, but as someone actively job searching, it is pretty simple. I want the system to show me it has been paying attention.

Not a separate AI Resume Maker. A profile that gets smarter the more I use the product, learning from the roles I spend time reading, the searches I abandon, the companies I keep coming back to. That behavioral signal is richer than anything I typed into a form.

Not invisible suggestions with no context. A light sentence on the ones that matter most: "This matches your recent searches and your stated interest in product roles at early-stage companies." That is the chip model Google uses, peripheral, present, one tap to act on. It does not require a heavy AI overlay. It requires one honest sentence.

And not auto-apply as it currently works. Before AI acts on my behalf on something this significant, I want to see exactly what is being sent, to which company, with which version of my resume, and why the system selected it. That confirmation is not friction for friction's sake. That is the Call Screening model, let the AI do the work, give the human the information to make the final call.

None of this requires the engineering investment of Apple's silicon architecture or Google's AICore. It requires a different question being asked before features ship.

The Question Worth Asking Before You Ship

Apple made AI invisible because their users trust invisibility when it comes from Apple. Google made AI visible because their platform thrives on openness and developer access. Both decisions were coherent. Both involved real trade-offs.

The question both teams asked before every feature, and the question most product teams are not asking deliberately enough, is this:

When your user looks at the screen, does the AI feel like it is working for them, or working for your metrics?

The answer shows up in small design decisions. Where a button sits. Whether an action has friction or not. Whether a suggestion comes with a reason. Whether the interface tells the user what it knows about them and why.

Those are not engineering decisions. They are design decisions. And right now, in most products, they are not being made deliberately enough.