Manish Karney

Field notes

Notes from the years the questions were still live.

Dated to the moments they come from — early product bets, launch pressure, consent questions, reliability resets, leadership transitions. Some are conviction, some doubt; the work kept asking better questions.

Newest first.

2026

  1. 2026 · Trust

    Why I left the platform for the patient

    Twelve years building for billions taught me to feel scale in the abstract. I’m leaving for work where the consequence has a face, on purpose.

    Live questionAfter a career of consequences measured in millions, do you want the kind you can finally see?

    TrustHealthcareConsequence

  2. 2026 · Trust & safety

    The adversary is sometimes the account holder

    What designing supervision tools taught me: in trust systems, the dangerous insider is often someone you already let in.

    Live questionIn a trust system, who do you design against when the dangerous actor is already authorized?

    TrustSafetyIdentity

  3. 2026 · Leadership

    The team is the deliverable

    On the craft of management — why a leader’s real output is what their team can build without them.

    Live questionWhat is a leader’s real output, if not what their team can build without them?

    LeadershipPeopleCraft

  4. 2026 · AI & trust

    “Humans in the loop” is not a safety plan

    The phrase reassures more than it protects. What a real version asks of the people who build the system.

    Live questionWhen is a human in the loop a safeguard, and when is it just a place to put the blame?

    AI safetyTrustAccountability

  5. 2026 · Ethics & craft

    Defaults are moral decisions

    What eight years building for families taught me about the ethics hiding inside a toggle’s default state.

    Live questionWho is a default really serving, before anyone reads a word?

    EthicsDefaultsCraft

  6. 2026 · Trust-critical systems · The abuse case

    Design the unsafe home first

    A supervision tool is most revealing not in the loving family it was built for, but in the home where it becomes a weapon. That case is not an edge. It is the spec.

    Live questionDo you design a trust system for the user you hope you have, or the one who already holds power over someone else?

    TrustPowerAI

  7. 2026 · AI systems

    I bet a compliance migration on a model

    The cleanup nobody would fund with humans was the one that mattered most. The only way to afford it was to let an AI do work I’d normally never let it touch.

    Live questionWhen the work is too big for humans and too sensitive for a model, which fear do you let win?

    AIMigrationTrust

2025

  1. 2025 · YouTube · Platform

    A platform no one is forced to use

    My team sets the standards hundreds of engineers build on, and I can’t order a single one of them to adopt anything. A platform you can’t mandate has to be good enough that people choose it.

    Live questionHow do you set standards for hundreds of engineers when your authority over them is essentially zero?

    LeadershipPlatformsInfluence

  2. 2025 · Engineering org

    My reach exceeded my authority on purpose

    Eight people reported to me; ninety ran on processes I set. I had to learn to lead the ninety without the lever I’d been trained to pull.

    Live questionHow do you take responsibility for far more than you control without pretending you control it?

    LeadershipInfluenceScale

2023

  1. 2023 · Google Assistant · Platform

    The empire you agree to dissolve

    We built a custom framework to make Assistant logic portable across servers and devices. Then I argued my own teams should hand most of it away — and wind themselves down.

    Live questionWhen the strategy shifts away from what your team owns, do you defend the empire or dismantle it on your own terms?

    LeadershipPlatformsChange

  2. 2023 · Assistant & Gemini · AI systems

    The demo is not the product

    AI capability is crossing from interesting to obvious. I keep coming back to the distance between a great demo and something real users can rely on.

    Live questionWhat has to be true before an AI capability becomes a dependable product?

    AIFulfillmentProduct systems

2022

  1. 2022 · Google Assistant · On-device

    The cost of running closer to the user

    Putting the Assistant’s brains on the device buys privacy, speed, and a response that survives a dead network. It also spends battery, memory, and disk I don’t get back.

    Live questionWhen is a local model worth the battery, memory, and disk it permanently spends on someone’s phone?

    PrivacyLatencyConstraints

2020

  1. 2020 · Family Link · Transparency by design

    The child is watching too

    A supervision product can feel like surveillance from the inside. The work that mattered most was making the kid’s side of it honest — and not spending their battery to reassure their parent.

    Live questionHow do you make a child feel watched-over rather than watched — when the same data reassures a parent and exposes a kid?

    DignityTrustPrivacy

2019

  1. September 2019 · Family Link · Reliability

    I asked Google to stop shipping

    Growth was outrunning reliability, and I was about to argue that a growth org should stop shipping. I wasn’t sure I was right.

    Live questionWhen does continuing to ship become the irresponsible choice?

    ReliabilityLeadershipTrust repair

  2. 2019 · Family Link · Reliability

    The dashboard is not the reality

    We had dashboards. That is not the same as being able to see what users actually felt.

    Live questionCan an organization tolerate seeing itself clearly when the dashboard is green?

    ReliabilityObservabilityHonesty

  3. 2019 · Family Link · Identity & dignity

    We called it a supervised member. He called it his account.

    The infrastructure needed one clean word for the relationship. The teen on the other end of that word already had a name, a history, and a sense of whose account it was.

    Live questionWhen supervision is added to an account a person already owns, whose account is it now?

    IdentityDignityTrust

2018

  1. 2018 · Family Link · Teens & autonomy

    A supervised teenager is the most motivated red team on earth

    Designing against bypass without designing with contempt — and why the family is not a security model.

    Live questionHow do you design against bypass without designing with contempt?

    AutonomySafetyTrust

  2. 2018 · Family Link · The autonomy transition

    The account belongs to the teenager

    Parents wanted to supervise teens who already had real digital lives. But that account was the teen’s — which made consent, and the exit, the whole engineering problem.

    Live questionWhen a parent wants to supervise an account that already belongs to a teenager, who is allowed to say yes — and who is allowed to leave?

    AutonomyConsentSafety

2017

  1. 2017 · Family Link · Launch

    Parents don’t want controls. They want confidence.

    The category says parental controls. I kept finding parents wanted something else: to feel less alone in a decision they were already making.

    Live questionIs the job to make a parent feel powerful, or to make them feel steady?

    ProductDefaultsTrust

2016

  1. 2016 · Family Link · 0→1

    Four prototypes became a promise

    The pieces were starting to look like a product. The harder question was whether a family could understand the whole.

    Live questionWhen do useful features become a coherent promise a family can understand?

    0→1LaunchProduct clarity

  2. 2016 · Family Link · Identity model

    A family is not an account

    Google’s identity systems knew one thing: the individual adult. We had to teach them what a child, a teen, and a parent were — without forking every product to do it.

    Live questionHow do you model ‘supervised’ for a six-year-old and a sixteen-year-old in the same system without forking every product that touches an account?

    IdentityTrustScale