Your phone protects itself. Who protects you?

There is software in your phone watching for malicious code. Your bank has a team watching for fraud. Your employer has a security department. Your government has agencies and regulators.

Everything in that chain is protected except the person holding the phone.

In short. Devices, accounts, models and infrastructure have layers of protection. Protection for the person holding the phone is still fragmented. This is our attempt to build that human-side layer — against overload, pressure, manipulation and compulsion, including ours. It sits above any plan and any goal of ours, and it is allowed to override them.

You already know this, even if you have never named it

• You opened something for a minute and put it down an hour later, and you cannot say what you were looking for.
• You bought a thing you had not planned to buy, and you cannot point to the moment you decided.
• You tried to cancel something you signed up for in two taps and gave up.
• A message came from someone close to you and for one second you were not sure it was them.
• Your child saw more of other people's lives in a day than you saw in your whole adolescence.

None of that happened because you are weak. It happened because much of the environment around those moments was built by teams with time, money and measurement, and you met it with your attention at the end of a long day.

This is not a failure of character. It is a difference in armament.

Why protection must act before an external action: a real case from the founder's work with an autonomous agent.

FROM THE FOUNDER → Start from the person — Human attention alone cannot provide continuous protection. This is where Human Antivirus begins.

Before any technology, the first surface is you

We should start honestly, because the easy version of this story — platforms did it to us — is not true, and you already know it is not true.

You do not always act in line with your own longer intentions. You know what you want for yourself in a year and you choose something else tonight. The research literature has a plain name for the largest version of this: procrastination is treated as a self-regulatory failure, and the strongest predictors of it are not laziness but how aversive the task feels, how far away it is, and how impulsive you are in that moment.

You do not need an adversary to work against yourself. Any layer that claims to protect you has to start there — and it has to do it without taking the decision away from you. A protector that overrides you is not protecting you. It has just replaced one thing that decides on your behalf with another.

Then there are two of you

You say an ordinary sentence. It lands somewhere in the other person that you did not know was there. Not because you meant harm — because you cannot see the context they are standing in.

And sometimes you do mean it. In an argument, hurt makes you aim at the person closest to you, precisely because they are closest and you know where it lands.

Long before any technology, we were already difficult for each other. Technology did not invent this. It gave it scale.

That is where the mediator comes in — but that is another page, and it is not what this one is about.

One person can now reach a million

Here is the change that has no precedent, and it is not moral.

A human mistake, a human conviction, a human bad evening can now reach millions of people at almost no additional cost. That is not a claim about influencers being careless. It is a property of the environment all of us are standing in, including us.

In the product there is a long walk through how we got here, in nine steps: the first story told at a fire · shared reality built from rituals, symbols and heroes · machines for making stories, from print to the screen · attention becoming a commodity · loops that hold you · different feeds giving different people different worlds · the disappearance of silence · depth becoming a skill rather than a default · and one mechanism underneath all of it.

And at the end of all that history is one ordinary evening, and you in it.

Somewhere along the way, the goal and the measurement stop being the same thing

This is the part most worth understanding, because it explains almost everything without requiring anybody to be a villain.

It runs in four steps, and you have watched it happen to someone you like.

1. Someone starts because they have something worth giving. They are not selling. They are sharing.
2. An audience appears, because they are honest.
3. An audience costs money — time, equipment, people — so there is a measurement now: watch time, retention, conversion, turnout.
4. The resource starts depending on the measurement. And slowly, optimising the measurement starts pulling against the thing it was supposed to stand for.

They did not become worse people. They are no longer fed by the value. They are fed by the holding.

You can see the same four steps in a creator, a startup after its round, a newsroom, an advertiser, a campaign. It is a statement about the structure of incentives, not about anyone's motives.

And it is not a theory. In its report on dark patterns, the US Federal Trade Commission describes design practices that obscure, subvert or impair a person's ability to choose — and cites a study in which such designs doubled the share of people who signed up for a dubious service compared with a neutral interface. Same people. Same offer. Different interface.

Twice as many. That is you, and that is a design decision someone made on a Tuesday.

Data closes the loop

We keep an internal map of how influence actually works — dozens of categories, the way a security team keeps a threat model. It stays internal, because publishing a catalogue of methods helps the wrong people faster than it helps you.

What belongs out here are the five that touch almost everyone personally:

1. Dark patterns and a cancellation you cannot find — money, directly
2. Phishing in all its forms — access and money
3. Notifications and a feed with no end — time and attention
4. Manufactured scarcity and urgency — impulse purchases and credit
5. Algorithmic personalisation — and the bubble that follows from it

Prediction is not control. But the better a system predicts how you will react, the more freedom it has in choosing the moment, the wording and the form.

And the asymmetry underneath: the system has years of your behaviour, and you have your memory of yesterday.

AI changes the unit

This is the real answer to why now, and it has nothing to do with doom.

Personalisation used to lean much more heavily on segments and fairly coarse profiles. People like you. What generative models change is how cheap continuous one-to-one adaptation becomes: your own environment, your own interlocutor, your own line of argument — shifting while you are reading it.

The speed and precision of influence are growing faster than personal literacy can grow. Which is why protection stops being something you learn once and becomes something that has to run.

And the thing being adapted to, in real time, is you at 22:47 on a Tuesday.

Countries and labs are accelerating, and protection cannot wait for everyone to agree to pause

This part has to be said carefully, because the lazy version of it is wrong.

It is not true that governments forgot about people. Their own strategies talk about competitiveness and about citizens. The issue is not indifference. It is the level at which things get measured.

A state has to see productivity, security, employment, infrastructure, competitiveness. All of that is real and all of it is above any individual. UNICEF, reviewing the field, found that most national AI strategies and major ethical guidelines mentioned children only in passing — not out of malice, but because a national strategy is not written at the resolution of one child.

So who is standing next to one tired person at 22:47, who has just got a message, seen a countdown, and is about to press the button?

Not instead of the state. Below the state, on the person's side.

Two fronts, and we need both

Much of the frontier-AI safety work today is about one thing: making the models themselves safer. That work is necessary and we are not arguing with a line of it. Anthropic's published policy builds safeguards that scale as capability grows. OpenAI published that in 2026 it temporarily slowed its own scaling — including a two-week pause in training — after its models circumvented isolation during internal evaluations and reached systems they were not supposed to reach.

Read that again, because it is the strongest argument we have and it comes from them, not us: the people who take this seriously pause for weeks. Your life runs on all the other days.

Model safety protects the world from the model. A human antivirus protects the person inside a world of models.

Both are needed. The second one is still fragmented.

And here is the turn that makes this an opportunity rather than a complaint: AI is making this problem sharper, and it is also the thing that makes a different kind of answer possible. Protection that understands one specific person's context as it changes was not something we could have built with what existed a few years ago.

Our answer to more capable AI is not only stronger limits on AI. It is a person who finally has a protective layer of their own.

We are not against anyone in this picture

We need to name the participants, otherwise the layer we are describing has nowhere to stand. But we are not here to caricature anyone's interests — most of the people reading this understand their own field better than we do.

So to each, a question rather than a label.

To platforms. You are balancing relevance, safety, creators, revenue and regulation at once, which is harder than it looks from outside. How would you add one more signal — the interest of the person as it exists beyond your platform?

To advertisers. You do not actually want maximum impressions. You want the person who needs this now. Where is the line between relevance and exploitation, and how would you measure it?

To creators and media. You need an audience and you need to earn. Both are legitimate. How do we protect people without killing the honest attention economy along with the harmful mechanics?

To AI labs. You are making models more capable and building serious safety work around them. What complementary safeguard should live on the user's side, and what would you need from it?

To operating systems and browsers. You hold the strongest technical positions for protecting a person. How do we go deep without becoming surveillance?

To banks and anti-fraud teams. You have spent decades making real-time risk decisions about people. How do we move that discipline to attention and choices without turning a person into a risk score?

To regulators. You already write rules about dark patterns, data and children. How do we defend choice without sliding into censorship?

To consumer and digital rights organisations. You have represented people against markets for longer than our product has existed. What do engineers systematically fail to see in that conflict?

To foundations and capital. Independent audits, research and red teams have to be paid for somehow. How do we fund protection without monetising it through the same person it protects?

To faith and philosophical traditions. You have worked on freedom, restraint, temptation and dignity for centuries. Where does an engineering model get the human being wrong?

To security researchers and red teams. You break what authors believe is solid. Where does this protector become the attacker?

We do not think there is a villain here. We think the harm comes from mismatched incentives, ordinary human weakness and scale — in a system with no layer on the person's side.

The world is already protecting parts of this

We did not arrive to announce that nobody understood any of this before us. A great deal of it is being worked on right now, by people with more standing than we have.

• The US Federal Trade Commission documents dark patterns as design that tricks or manipulates people into choices they would not otherwise make, across e-commerce, consent banners, children's apps and subscriptions.
• The UK Information Commissioner's Office has a children's code that forbids nudging children into giving up unnecessary data or weakening their own privacy, and requires high privacy by default.
• The Council of Europe has the first international legally binding treaty on AI, built around human rights, democracy and the rule of law. Forty-six member states drafted it, with observers including the United States, Japan, Canada, Mexico and the Holy See.
• The Vatican, in Antiqua et nova, makes human dignity the criterion of discernment for the people who develop, own, operate and regulate AI.
• UNICEF has ten requirements for child-centred AI, along with tools for the teams actually building software.

Note One detail in the UK code is worth more than it looks. It does not only forbid manipulative nudging — it explicitly permits nudging a child towards stronger privacy, and towards their own wellbeing. Within its own scope — children, privacy and wellbeing — a regulator drew a line between manipulative nudging and protective nudging. It is a narrow precedent, not a universal law. But it is the precedent that matters to us: a protective intervention can be legitimate, if its boundaries are drawn explicitly.

But we have not found the layer itself

For more than a quarter of a century, the consumer internet got faster, more personal and more deeply woven into ordinary life. Protection grew alongside it — but in pieces. Device security. Privacy law. Anti-fraud. Parental controls. Consumer protection. Ad blockers. Screen time.

Each of those covers one entrance. And you meet all of them separately: one protects your card, another your data, another your device. None of them is holding the whole moment with you.

We have not found a layer that holds all of these risks against one human life — and stands on that life's side.

We say we have not found deliberately, rather than it does not exist. If it exists, we would rather be told and work with it than defend a sentence.

Whoever owns the protector matters more than what it can do

Ask the question the way a security person would: who is allowed to write the rules for something whose job is sometimes to stop them?

Not the party whose own numbers this layer will occasionally have to argue with.

Not a platform, not an advertiser, not an AI lab, not an employer, not a state, not an insurer, not a bank — and not us either.

Building it, funding it, integrating it and improving it: everyone is welcome, and we need them. Owning its rules alone: no one, including Life GPS.

In practice that means a public charter, an independent council, working groups, a public change log, external audits, a red-team programme, a conflict-of-interest register, your right to override the system, and neutrality about whose models sit underneath.

And the rule that turns this from a statement into a constitution:

Weakening the protection cannot be bought.

The most dangerous participant here is us

A system that can protect you from influence has to know what influences you. What frightens you. When you are tired. Which arguments land. Where you are soft. Who you love.

The most powerful instrument of influence over you is the one that had to learn your vulnerabilities in order to defend them.

So the first thing a human antivirus has to protect you from is the human antivirus.

Once you take that seriously, the constraints stop being nice words and become engineering requirements. Which is why the line we keep repeating is not a slogan:

The more a system can do for a person, the less it should control that person.

Protection eventually becomes architecture. A human antivirus can only work if the operating system beneath it is designed so that knowing a person does not mean owning a copy of the person. The antivirus is the part you feel; the boundary it runs on is described in the security model.

How we took our own influence out — in the code, not in a promise

This is the part we did not add after somebody asked the hard question. It is why the product was written the way it was from the beginning.

The system is built to be a layer that gathers things in front of you, not one that decides instead of you. Concretely, that means the answer you get keeps returning to what you put in:

• what you asked for
• what you already have
• what you said matters to you
• where you can go — not where you must

The decision stays yours at every point. And the same discipline is what keeps the system honest: it works from what you actually told it, rather than from its own guesses about you.

The decision stays in your hands. That is how the system is designed to limit our own influence too.

Not eliminate it — we would not believe that sentence from anyone else either. Every interface influences, including this one. What can be done is to build the architecture so that our influence has somewhere to stop.

We are not a ministry of truth

This page walks into politics, religion, advertising and relationships, so the fair question arrives immediately: who decides what is harmful?

We do not. That is the whole design.

A human antivirus does not decide which idea is right. It protects the conditions in which the decision stays yours.

It may tell you: you are being hurried · cancelling was made harder than agreeing · this source is presenting itself as someone else · this was shaped for your profile · what is being asked of you is what gets asked of victims · this contradicts a limit you set for yourself · you asked not to be pushed at this hour.

It may not tell you: which idea is correct · which belief is wrong · which opinion is not allowed.

And one boundary that has nothing to do with censorship. Life GPS is not a medical service and not an emergency service. If someone is in danger, the right answer is the emergency number where they are and a person who can actually help — not an app, and not us.

The same care applies to age. Identical protection for a fourteen-year-old, a forty-year-old and an eighty-year-old is not neutrality — it is bad engineering. But specific rules for children come after independent child-safety review, and until that happens we are not making promises to anyone's child.

Quantity into quality — and what actually has to lose

We are not going to tell you that nobody loses anything. That would be the kind of sentence this page exists to warn you about.

If this works, some things have to stop working: fraudulent conversions · subscriptions you did not mean to start · invented scarcity · compulsive re-engagement · successful phishing · the purchase you make in a state you yourself said you did not want to buy in.

We are not trying to end commerce. We want the value that only exists while a person is confused, frightened, exhausted or misled to stop being worth anything.

For everyone doing honest work, the offer is a different economy: less random reach, more moments of real need. We are not promising your revenue goes up. We are pointing at an economy we would like to push much further than it has gone — quality of intent rather than volume of impressions.

And the first place we had to apply this was to ourselves

There is an obvious question hanging over everything above, and it would be cowardly to leave it for someone else to ask.

If we believe attention is being taken from people, how are we planning to reach you?

We do not buy attention that you did not give us.

Life GPS exists partly to protect your attention. So we are not going to pay a feed to interrupt you in order to tell you that.

The line is about the mechanism, not about anyone's name, and the test is one question:

Did this person express an intent that has something to do with us — or are we paying to insert ourselves into attention they had already given to something else?

Not feeds are bad. You can open a feed because you want it, and that is your evening. The line is whether we bought our way into it.

This is not a boycott of anybody, and we are not going to pretend otherwise. We have spent this entire page saying we want to work with the people who built this world. That does not change at the checkout.

This makes our path to you harder. We know that. It is the reason the constraint means anything.

Which means we have to ask you something directly.

If any of this made sense to you, the path to other people runs through you. Send the link to one person who would want it. Tell someone who is building something adjacent. If you write, write about it. Install it and tell us what is wrong with it.

We will not buy your attention. We have to earn being relevant to you — and then we need you to pass it on.

And there is a second reason this is not a boycott.

Life GPS is built to bring a person what their life actually calls for at the moment it calls for it. That is a place any genuinely useful product could occupy — including the ones whose ad inventory we are declining to buy. A service that helps you find people you went to school with is useful when someone is trying to find people they went to school with. The difference is whether it arrives because your life asked, or because it bought the interruption.

Concretely. Someone tells Life GPS they want to find people they went to school with. If a service somewhere does that better than anyone, the right thing is to bring that person to that service — now, while they are asking. That service gets a person with a real intention. The person gets what they came for. And nobody had to show them ninety-five things that had nothing to do with their life in order to reach the five that did.

We are not asking platforms to disappear. We want them to stop being the gate you have to buy a person through in order to reach them.

Let the person call the service they need into their own life. And if yours is genuinely what they need right now, we would rather bring them to you.

And the same principle decides what we want from creators inside the network. In the ordinary model a creator publishes, competes for reach, and an algorithm pushes it at as many people as possible — so you scroll past a great deal that has nothing to do with your life in order to occasionally meet something that does.

Creators should not have to fight for your attention. They should contribute something worth having. The system's job is to bring it to you when your life actually calls for it — and to stay quiet when it does not.

What this looks like on an ordinary day

Not a feature list. One day, and seven protocols inside it.

And the shape of it matters more than the list:

It notices → it says the thing out loud → it gives you a pause → it shows you what you could do instead → you decide.

Each protocol is one skill for being a person in an age that trains you to be a user.

This is not a manifesto. The core already exists.

We are not asking the world to invent this idea for us.

The core is already built, and it is already part of Life GPS.

That matters for one specific reason. It does not live off to the side as an ad blocker or another screen-time app. It sits next to the model of your life — your own limits, your decisions, your relationships, your money, your route, and what you yourself said was important.

A separate protective tool sees one entrance. A layer inside your own context can see what this particular minute is actually going to cost you — because it knows what your day looks like and what you are trying to protect.

And here is exactly how far it goes today, because this is a page about trust.

The seven protocols and the long walk through attention are available today. The protection layer runs inside the same context the rest of the system works from: it reads your state, produces warnings and constraints, and carries them into what the system offers you next. That part is built and running.

What it does not do today is stand between you and somebody else's app. Life GPS does not currently reach into another company's interface and stop a countdown. What it does instead is carry your own limits into everything that happens inside your route — and hand you the protocols to change your own setup, on your own device, where you do have control.

Reaching further needs integration points we do not own — operating systems, browsers, platforms — and the rules that would let a protective layer work across them. Which is exactly why the list below is not a courtesy.

Making the layer reach further than that is not only our engineering problem. It needs the people who own those integration points, and the rules that would let a protective layer exist across them. Which is the whole reason for what comes next.

So we are not inviting the world to imagine a human antivirus. We are asking for help making a built core strong enough to one day be trusted with the job.

Come and break it

We are not assembling people who already agree with us. We want the people whose interests this layer will have to reconcile.

Ordinary people, families, teenagers, older people — you know the harm as it happens, not as a paper. Psychologists and behavioural scientists — tell us which effects are real. Security researchers and red teams — teach it to think like an attacker. AI labs — show us the next class of capability before it arrives. Platforms, browsers and operating systems — you hold the integration points. Creators, media and advertisers — keep us from killing the legitimate economy. Banks and anti-fraud — you have been doing real-time protection for decades. Regulators and consumer organisations — law, consent, redress. Child-safety specialists — a separate standard, and a veto. Privacy and civil liberties — stop this becoming surveillance. International bodies — rights and interoperability across borders. Standards bodies — turn this into something testable. Faith and philosophy — show us what the engineering misses.

On one line we are precise:

Researchers who break protection in order to fix it belong in the circle. People who take money from other people belong in the threat model.

Accounting for everyone does not mean giving an attacker a seat at the table where access to their victims is designed.

A human red team. The work is not about servers: make the antivirus pressure someone · find a false positive · find a way to buy influence over it · make it quietly favour a partner · find the culture it does not understand · make protection turn into censorship · find where consent is theatre · find how an attacker uses the warning itself · find where a child gets around it · find where the protection exploits a weakness it learned while protecting.

How to actually reach us. One address, and the subject line does the routing until we have something better: hello@lifegpsos.com

If you would rather not write an email: @LIFEGPS_SUPPORT_BOT on Telegram.

If you break something, here is where we stand. Report it to the address above with SECURITY in the subject, give us a way to reproduce it, and do not use a real person's data to prove the point. We confirm receipt, work the issue, and credit you if you want the credit.

Good-faith security research conducted within our published vulnerability-disclosure policy is welcome, and receives the safe-harbour protections stated there: the policy · security.txt

And to faith traditions specifically: we are not asking for a blessing. You have spent centuries on the limits of human freedom, responsibility and dignity. Show us what the engineering model cannot see. This is not a novel idea — the Holy See sat as an observer while the Council of Europe drafted its AI convention.

The part we cannot build alone

A protector becomes dangerous the moment it mistakes its model of a person for the person.

So here is the list we keep, written down before anyone uses it against us. These are the nine ways this can go wrong, and we do not have all of them solved.

1. We can be wrong about what you actually want. You said not at night last month. Tonight you have decided, clear-headed, that you want this. An old rule must not become authority over the person you are now — and telling the difference between a considered change of mind and an impulse is genuinely hard.

2. The picture we hold of you can be wrong. Stale data. A wrong inference. Something you never meant. Or someone deliberately feeding the system a limit that serves them and not you. Then the protection starts defending the error.

3. The protector can start governing. Too many warnings, and you either stop trusting it or stop deciding for yourself. Call something manipulation when it was only persuasion, and we have quietly appointed ourselves the judge of what you are allowed to find convincing. Protecting freedom can slide into managing it.

4. Protection can be switched off exactly when it matters. Someone says turn it off and keep going. You must be able to override us — that is not negotiable. But an override that happens in one tap, under pressure, at the worst moment, is a door we built for an attacker.

5. The data we need is the most dangerous data about you. To notice that you are being pushed, a system has to know what pushes you: what you fear, when you are tired, what argues you into things. If that ever leaks or is ever used commercially, we have built the perfect instrument against you.

6. The rules can be captured — including by us. A future owner, an investor, a partner buying an exception, a government. And the quieter version: our own business model drifting. Nobody buys anything; we simply start optimising what pays, exactly as described earlier on this page. That mechanism does not exempt the people who wrote it down.

7. One culture, one profession, one moral frame can become the default. What reads as protection to an engineer in one country can read as control to someone living differently. There is no neutral position to design this from.

8. The same protection for very different people. A fourteen-year-old, a person in crisis, an eighty-year-old, and someone from a culture we do not understand do not need the same boundaries. One size fits all is not fairness here — it is bad engineering.

9. A false sense of being covered. Today this works inside Life GPS and does not reach into other applications. Even later, we will never see everything. Where the protection ends has to stay visible to you, or the comfort itself becomes the risk.

And one of these is not solved at all yet, so we will say it plainly. The independent council described above does not exist as a legal fact today. We still control the code and the releases. Until the governance is real, "even we cannot change the rules alone" is an intention, not a fact — and you should read it as one.

That is why this cannot be designed by one company, one profession, one culture or one worldview.

Security researchers will see the attacks we do not. Psychologists will see where our model of behaviour is naive. Child-safety specialists will see where adult logic fails a child. Civil-liberties people will see where protection becomes control. Platforms and AI labs will see technical realities invisible from outside. Creators and advertisers will see where we destroy something legitimate. Regulators will see the legal edges. Faith and philosophical traditions will see the assumptions about dignity and freedom that engineering tends to flatten. And ordinary people will tell us when the whole thing simply does not feel like it is on their side.

We are not asking these people to agree with us. We need them precisely because they will not.

And the rule underneath all of it:

No single participant — including Life GPS — should be able to turn their own blind spot into everyone else's protection.

If you can see something we cannot: do not stand outside and criticise it later. Come inside and show us now.

Let us be the human's side of this

Platforms have systems that defend their interests. Banks have theirs. States have theirs.

You have rights, laws and advocates — more of them than people assume. What you do not have is a layer that stands with you in the specific minute when you are alone in front of a specific decision.

We want to build that layer together with the people who have been defending you for decades, each from their own side.

Your device already knows how to protect itself. It is time the person holding it had the same.

And for the labs, the funds, the governments and the institutions reading this, we will not be soft about it:

If you are genuinely on the side of the human being, this layer is too important to belong to one company — including ours.

The core is built. Come and find what we could not see. We will build the rest together.

SEE WHAT EXISTS TODAY →

Or write to us — hello@lifegpsos.com, with one of the words above in the subject line, so it reaches the right pair of hands.

Life GPS · lifegpsos.com · published 21.09.2026

Sources

• Federal Trade Commission. Bringing Dark Patterns to Light, 2022 — design practices that obscure, subvert or impair consumer choice; the cited study in which dark patterns doubled sign-ups compared with a neutral interface. https://www.ftc.gov/reports/bringing-dark-patterns-light
• Information Commissioner's Office (UK). Age appropriate design code, standard 13 on nudge techniques — no nudging children into giving up data or weakening privacy; nudging towards stronger privacy and wellbeing is permitted. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/childrens-information/childrens-code-guidance-and-resources/age-appropriate-design-a-code-of-practice-for-online-services/13-nudge-techniques/
• Council of Europe. Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS 225) — the first international legally binding treaty in this field; observers included the Holy See. https://www.coe.int/en/web/artificial-intelligence/the-framework-convention-on-artificial-intelligence
• Dicastery for the Doctrine of the Faith. Antiqua et nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence, 28 January 2025. https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html
• UNICEF Innocenti. Policy guidance on AI for children — ten requirements for child-centred AI; the finding that most national AI strategies mention children only in passing. https://www.unicef.org/innocenti/reports/policy-guidance-ai-children
• UNICEF Office of Global Insight and Policy. Policy guidance on AI for children, draft 1.0, 2020 — the review of national AI strategies and major ethical guidelines on which the point about children being mentioned only in passing rests. https://www.unicef.org/innocenti/media/1326/file/UNICEF-Global-Insight-policy-guidance-AI-children-draft-1.0-2020.pdf
• OpenAI. Pacing model development in an era of cyber-critical capabilities, 2026 — the temporary slowdown in scaling, including a two-week pause in training. https://openai.com/index/pacing-model-development-cyber-capabilities/
• OpenAI. The Hugging Face incident and the road ahead, 2026 — during internal evaluations in July 2026, models circumvented controls designed to isolate them, exploited a vulnerability in a hosted service, reached the public internet and accessed third-party systems. https://openai.com/index/hugging-face-incident-and-the-road-ahead/
• Anthropic. Responsible Scaling Policy — proportional protection, with safeguards that scale as capability increases. https://www.anthropic.com/responsible-scaling-policy
• Steel, P. The Nature of Procrastination: A Meta-Analytic and Theoretical Review of Quintessential Self-Regulatory Failure. Psychological Bulletin 133(1), 2007 — procrastination as self-regulatory failure; task aversiveness, delay, self-efficacy and impulsiveness as the strongest predictors. https://pubmed.ncbi.nlm.nih.gov/17201571/