THE GREAT ATTENTION EXPERIMENT: 10 - THEY KNEW. THAT CHANGED THE QUESTION.
Imagine a company discovers that one of its products may be making a vulnerable group feel worse.
The evidence is incomplete.
It does not prove the product caused the problem.
Most customers report no harm.
Some say the product actively helps them.
The negative effect appears concentrated among a smaller group already experiencing difficulty.
What should the company do?
Remove the product?
Redesign it?
Warn people?
Introduce stronger protection for the vulnerable group?
Conduct more research?
Release the findings?
Invite independent scientists to examine the data?
Continue operating while it investigates?
There is no simple answer.
Uncertainty does not automatically require withdrawal.
Almost every useful product carries risk.
Cars kill people and create extraordinary mobility.
Medication causes side effects and saves lives.
The internet can expose children to harm and connect isolated young people to support they cannot find anywhere else.
The moral problem does not begin merely because a risk exists.
It begins when the organisation possessing the greatest knowledge of that risk also controls the evidence, the product, the protection and the commercial incentive determining how quickly anything changes.
That is where the attention economy reached its turning point.
For years, technology companies could reasonably argue that they were building unprecedented systems inside an uncertain environment.
Nobody knew exactly what would happen when billions of people began carrying personalised, connected devices throughout almost every waking hour.
Nobody could predict every effect of algorithmic recommendation, visible social metrics, infinite feeds, targeted advertising and permanent connectivity.
Then the companies began researching those effects.
Employees raised concerns.
Internal documents recorded risks.
Whistleblowers came forward.
Regulators demanded information.
Parents described what they were seeing.
Young people reported harms directly.
The public evidence remained complicated.
The internal evidence was complicated too.
But the defence of complete ignorance became impossible.
They knew enough to recognise the possibility of harm.
From that moment, the question changed.
It was no longer:
Could anybody have predicted this?
It became:
What did they do once the warning signs appeared?
Knowledge Is Not a Switch
The phrase they knew sounds absolute.
It suggests a moment when executives received a definitive report establishing that their products caused harm and deliberately chose to ignore it.
The evidence does not support such a simple story.
Knowledge develops in stages.
A company may first become aware of complaints.
Then observe a pattern.
Then conduct interviews.
Then run surveys.
Then identify associations.
Then test product changes.
Then discover that different users experience the same feature differently.
One study raises concern.
Another produces a neutral result.
Another identifies benefits.
Researchers disagree about causation.
The organisation learns gradually.
Knowledge is rarely a switch moving cleanly from off to on.
It is an accumulation.
This matters because overstating what companies knew makes the accountability argument easier to dismiss.
Meta did not possess a scientific document proving that Instagram caused a population-wide mental-health crisis.
TikTok did not possess a single universally accepted diagnostic test demonstrating that every frequent user was addicted.
YouTube could not trace every emotional difficulty experienced by a young viewer to one recommendation system.
Social media interacts with family life, school, poverty, discrimination, temperament, sleep, existing mental health, friendship, puberty and wider culture.
No platform controls every cause.
No serious investigation should pretend otherwise.
But uncertainty does not equal ignorance.
A company can lack proof of universal causation while possessing credible evidence that specific designs, content pathways or user groups face particular risks.
That narrower knowledge can still create responsibility.
The Internal Research Was Not the Scandal
In 2021, former Facebook employee Frances Haugen disclosed thousands of internal company documents to regulators and journalists.
She later testified before the United States Senate.
Her central allegation was not merely that Facebook and Instagram sometimes produced harmful outcomes.
She argued that company leadership repeatedly encountered conflicts between safety and profit, possessed information unavailable to the public and failed to make changes that could have reduced harm.
Those were whistleblower allegations.
They should not automatically be treated as established fact.
Haugen had worked within Facebook, including on algorithmic products and civic misinformation, but she did not make every decision across the company.
Meta disputed her interpretation.
The leaked documents represented selected parts of a much larger internal record.
Still, the disclosures revealed something important.
Facebook had not ignored wellbeing entirely.
It had conducted research.
Researchers had spoken to teenagers.
They had examined body image, social comparison, loneliness, anxiety, sadness, eating issues and suicidal thoughts.
They had tried to understand where Instagram helped and where it might make existing problems worse.
That research was not evidence of negligence by itself.
It was evidence that people inside the company were taking the questions seriously.
A company unwilling to investigate harm would know less.
The problem was not that Meta had performed the research.
The problem was the gap between what the company could see internally and what outsiders could evaluate independently.
One in Three Teen Girls
The most famous finding from the disclosures was reduced to one sentence:
“We make body image issues worse for one in three teen girls.”
That sentence travelled around the world.
It sounded like an admission that Instagram harmed one third of all teenage girls using the platform.
Meta argued that this interpretation was inaccurate.
Its published explanation said the figure referred specifically to surveyed teenage girls who already reported struggling with body-image issues.
Among that subgroup, approximately one third said Instagram made the issue worse.
Twenty-two per cent said Instagram made it better.
Approximately 45.5 per cent said it made no difference.
That context matters enormously.
One third of a vulnerable subgroup is not one third of all teenage girls.
The study also relied on self-report.
It did not prove that Instagram caused the underlying problem.
The participants were being asked how they perceived the platform's influence.
Meta further argued that, across eleven of twelve wellbeing categories examined, more teenage girls experiencing difficulties said Instagram made those difficulties better rather than worse.
Body image was the exception.
The responsible conclusion is therefore not:
Meta proved Instagram was toxic to teenage girls.
It did not.
Nor is the responsible conclusion:
The finding was meaningless because most users were unaffected or helped.
It was not.
The research identified a subgroup of teenage girls already struggling with body image, and a substantial proportion of them reported that Instagram made the experience worse.
That is precisely the sort of signal a responsible company should investigate further.
The fact that many other teenagers reported benefits does not erase the vulnerable group.
The fact that the vulnerable group exists does not erase the benefits.
Both belong inside the evidence.
The Suicidal-Thoughts Finding Was Also More Complicated
Another internal presentation was reported as finding that 13 per cent of British teenagers and 6 per cent of American teenagers experiencing suicidal thoughts traced those feelings to Instagram.
Again, the denominator mattered.
Meta said those percentages applied to the smaller group of surveyed teenagers who had reported suicidal thoughts.
Across everyone surveyed, approximately one per cent said they had experienced suicidal thoughts that they believed began on Instagram.
Meta also said that, among teenage girls who reported struggles with suicidal thoughts or self-harm, 38 per cent believed Instagram made the experience better and 49 per cent reported no impact.
Those numbers do not establish causation.
A teenager may believe a feeling began on a platform without the platform being the sole or primary cause.
The sample was limited.
The result cannot be applied automatically to all teenagers.
Yet approximately one per cent of a platform operating at enormous scale is not trivial.
When billions use a service, small percentages represent large numbers of human beings.
The company acknowledged that even one person reporting such an experience mattered and pointed to investments in suicide-prevention resources and protective interventions.
The disagreement was not therefore about whether any young people experienced harm.
Meta accepted that some did.
The argument concerned scale, interpretation, causation and whether the company's public statements accurately reflected the complexity of its own findings.
The Research Did Not Say Instagram Was Simply Good or Bad
Meta's internal studies appear to have revealed a pattern familiar throughout the wider evidence.
The same platform could help one person and harm another.
A teenager feeling isolated might find community.
Another might experience exclusion more intensely.
One girl might discover support around body acceptance.
Another might encounter repeated comparison that deepened insecurity.
A young person struggling with identity might find language and belonging.
Another might be drawn towards harmful content.
The platform was not producing one universal effect.
It was interacting with the person already there.
Meta said some of the internal work involved small focus groups, including one sample of forty teenage users selected partly because they already struggled with issues such as body image, self-esteem or negative mood.
It said the work was intended to identify the most negative perceptions and inform internal product discussions, not establish causal population-level conclusions.
That defence is scientifically reasonable.
Qualitative research is not useless because the sample is small.
It reveals experiences that larger surveys may miss.
But it cannot determine prevalence reliably.
Internal research can identify questions without answering them conclusively.
The ethical issue is what happens next.
Does the company investigate more rigorously?
Does it introduce protection?
Does it share enough information for independent researchers to test the concern?
Does it treat the vulnerable minority as an unavoidable cost?
The Company Saw What Outsiders Could Not
This is where the information imbalance becomes critical.
A university researcher can ask teenagers how often they use social media.
They can measure reported symptoms.
Track groups over time.
Run experiments involving temporary limits.
Study particular content.
Analyse public data.
But they cannot normally see the complete ranking system.
They do not know every recommendation shown.
Every candidate rejected.
Every internal risk score.
Every product experiment.
Every change in session length.
Every pathway leading from one topic to another.
Every subgroup responding unusually strongly.
Every metric considered by a product team.
The company can.
Haugen told the Senate that almost nobody outside Facebook could see what happened inside the system and that independent researchers and regulators lacked access to the data required to test the company's claims properly.
The United States Surgeon General later identified the same structural problem.
The 2023 advisory on social media and youth mental health stated that lack of transparency and access to company data had obstructed understanding of the full scale and nature of potential harms.
It said there was not enough evidence to conclude that social media was sufficiently safe for children and adolescents, while also recognising meaningful benefits for connection, support, identity and self-expression.
The public debate was being asked to reach conclusions without access to the strongest evidence.
The companies possessed that evidence because the systems belonged to them.
That asymmetry created a responsibility greater than publishing occasional safety updates.
It created an obligation to make meaningful independent scrutiny possible.
Research Without Transparency Creates a Trust Problem
Companies often argue that internal research is misunderstood when removed from context.
They are right.
Internal documents use shorthand.
They may highlight negative findings precisely because employees are trying to solve problems.
A slide designed for colleagues familiar with a study does not contain every methodological caveat.
A researcher speaking frankly inside a company should not fear that every unfinished hypothesis will later be treated as a public admission.
Total transparency could discourage honest internal work.
It could expose personal information, security systems and legitimate trade secrets.
There are valid reasons not to publish every document immediately.
But complete secrecy creates the opposite danger.
The company can decide which research outsiders see.
Which findings receive context.
Which caveats matter.
Which risks deserve attention.
Which product changes count as sufficient.
The organisation becomes researcher, subject, evaluator and publisher simultaneously.
It also earns money from the product being assessed.
That does not mean every internal conclusion is dishonest.
It means the conflict must be managed structurally rather than through trust alone.
Independent researchers need secure access to appropriately protected data.
Regulators need expertise capable of auditing ranking and design systems.
Research methods need external challenge.
Companies cannot reasonably ask society to accept:
We have studied the product and concluded that our interpretation of our private evidence is correct.
Not when the product shapes childhood at global scale.
Warning Signs Existed Before the Leaks
The 2021 documents did not create concern about social media.
Researchers had already examined cyberbullying, sleep disruption, body image, compulsive use, social comparison and mental health.
Parents had already raised concerns.
Platforms had introduced reporting tools and wellbeing resources.
Facebook said it had worked with suicide-prevention experts since 2006 and repeatedly reassessed how its services handled self-harm content.
By May 2021, before the Haugen disclosures became public, Meta was openly acknowledging that social comparison and pressure to look a particular way could affect young people.
It was working with mental-health organisations and creators on resources intended to help teenagers manage those pressures.
This history complicates any claim that Meta discovered the issue only when newspapers reported the leaked documents.
The company knew social comparison, body image, bullying and self-harm were relevant risks because it had already built interventions around them.
That is not proof that it concealed a causal mental-health crisis.
It is proof that the broad category of risk was understood.
What Did Meta Change?
An accountability essay must examine action, not merely knowledge.
Meta points to a substantial record of safety and wellbeing changes.
It introduced stronger restrictions on graphic suicide and self-injury content.
Created dedicated reporting routes.
Developed tools allowing people to limit bullying without obvious retaliation.
Made teenage accounts private by default.
Restricted adults from messaging teenagers who did not follow them.
Reduced the recommendation of potentially sensitive material.
Introduced parental supervision.
Added reminders encouraging people to take breaks.
Began nudging teenagers towards different topics when they appeared to dwell repeatedly on one category.
In 2024, Meta launched Instagram Teen Accounts with protections enabled automatically.
These included private accounts, stricter content settings, restrictions on unwanted contact, sleep-mode notifications and parental tools.
In 2025, Meta expanded versions of Teen Accounts to Facebook and Messenger, saying at least 54 million Instagram Teen Accounts were active globally by April of that year.
By January 2026, Meta was publicly presenting these developments as evidence of a long-term commitment to teen safety and parental support.
These are not cosmetic changes.
Defaults matter.
Private-by-default accounts create a different environment.
Restrictions on direct contact can reduce exposure to strangers.
Nudges and time tools can interrupt automatic use.
Stricter recommendation settings can prevent some harmful pathways.
Meta deserves credit for building them.
The accountability question does not disappear.
It becomes more precise.
Why were many protections introduced only after years of concern, public pressure, internal disclosures, lawsuits and regulatory attention?
Were they introduced as soon as the company possessed enough evidence?
Could stronger defaults have arrived earlier?
How much harm is acceptable while a company waits for certainty?
The Instagram Kids Decision Revealed the Conflict
Meta had been developing a version of Instagram intended for children under thirteen.
The company argued that children were already obtaining phones, misrepresenting their ages and entering services intended for older users.
A child-specific product with parental controls, it said, could be safer than leaving underage children inside the adult platform.
That argument had merit.
Children were already present.
Banning them in the terms did not make them disappear.
A purpose-built experience could potentially offer stronger protections, limited communication and greater parental visibility.
In September 2021, Meta paused the project amid intense criticism following the internal-research controversy.
It maintained that the concept still had value but said it would consult parents, experts and policymakers before proceeding.
The dispute revealed the difficulty of interpreting corporate solutions.
A safer child-specific product could reduce harm.
It could also bring younger users formally into a commercial ecosystem and begin a relationship earlier.
The same proposal could be understood as protection and market expansion.
The business model made motive impossible to separate cleanly.
That does not prove bad faith.
It shows why independent oversight matters.
A Safety Tool Is Not the Same as a Safe System
Platforms frequently respond to risk by giving users more control.
Time limits.
Break reminders.
Notification settings.
Content filters.
Parental dashboards.
Keyword controls.
Ways to reset recommendations.
These tools are useful.
But they often place the final burden upon the person already being influenced.
A teenager receives a reminder.
They can continue.
A parent receives a dashboard.
They must understand what the numbers mean.
A user can reject a recommendation.
They must notice the pattern and act.
The platform continues operating around the intervention.
A warning placed inside a product designed to encourage continued use can become another screen to dismiss.
That does not make the tool fraudulent.
It limits what the tool can achieve.
The strongest protections usually change the environment by default.
They do not merely ask the user to resist it more successfully.
This is why Meta's later Teen Account defaults matter more than optional wellbeing advice.
The company accepted greater responsibility for the architecture.
TikTok Faced Similar Allegations
Meta was not the only company accused of understanding risks while publicly minimising them.
In 2024, New York and other American states brought legal actions against TikTok.
New York's complaint alleged that TikTok knew its design could produce compulsive use among young people, understood that minors were particularly vulnerable and used features including the For You feed, notifications and visible social feedback to increase engagement.
The complaint cited internal documents, although substantial sections were redacted from the public version.
These remain allegations made by a government plaintiff.
They are not final judicial findings.
The complaint's language sometimes adopts simplified descriptions of dopamine and addiction that require scientific caution.
The existence of internal quotations inside a legal claim does not independently prove the state's interpretation.
TikTok has consistently said it prioritises teen safety and points to age-specific protections, family controls and wellbeing tools.
The distinction between allegation and verdict matters.
Still, the complaint demonstrates that concern about corporate knowledge extends beyond one whistleblower and one platform.
Regulators now claim that internal records show companies understanding the behavioural effects of their designs in greater detail than their public messaging suggested.
That claim deserves adjudication, not automatic acceptance or dismissal.
TikTok Also Changed the Architecture
TikTok introduced a sixty-minute daily screen-time limit by default for users under eighteen in 2023.
Teenagers could continue after reaching it, but they were required to enter a passcode, creating an additional moment of choice.
TikTok said its tests showed that prompting teenagers to set limits substantially increased use of its screen-time tools.
The platform also restricted nighttime notifications for teenagers, expanded Family Pairing and introduced more controls over content and communication.
By 2026, TikTok said teen accounts had more than fifty preset safety, privacy and security features, including private defaults, messaging restrictions, content filters, sleep-related protections and screen-time limits.
Again, these changes matter.
They show that the platform can redesign prompts, introduce friction and differentiate the experience according to age.
They do not prove the allegations in the New York complaint.
Nor do they prove the platform had previously intended harm.
They establish that the architecture is changeable.
Nighttime notifications are not an inevitable feature of digital communication.
Infinite recommendations do not have to operate without interruption.
Teenagers can be placed into more protective settings by default.
The design always contained choices.
The Safety Features Create Another Question
When a company introduces a protective feature, several interpretations are possible.
It may have discovered a previously unknown risk.
It may have developed better technology.
It may be responding responsibly to emerging evidence.
It may be reacting to public pressure.
It may be anticipating regulation.
It may be attempting to protect reputation.
Several motives can be true simultaneously.
A company can genuinely care about safety and understand that safety failures threaten the business.
The more important question concerns timing.
What threshold of evidence triggered the change?
Was the change made when internal risk became visible?
After public criticism?
After regulation became likely?
After litigation created financial exposure?
Outsiders often cannot answer because the decision process remains private.
We see the announcement.
Not the years of internal debate preceding it.
The FTC Looked Across the Industry
In 2024, the United States Federal Trade Commission published a staff report based on information demanded from major social-media and video-streaming companies.
The FTC concluded that the companies collected and retained enormous quantities of information, often fed that information into automated systems and provided users with limited ability to avoid such uses.
It described many data-minimisation and retention practices as inadequate and said business models based on targeted advertising created incentives for extensive surveillance.
The report also found that many services treated teenagers much like adults and lacked sufficient restrictions or consistent safeguards.
The report was not limited to one company.
It examined an industry structure.
That distinction matters.
If several organisations independently produce similar weaknesses, the explanation may lie less in the character of one executive and more in the incentives examined throughout this series.
Collect more information.
Improve prediction.
Increase engagement.
Expand monetisation.
The behaviour makes sense inside the model.
Safety measures that limit data or engagement must compete with the model.
The Public Science Remained Uncertain
By 2023, the United States Surgeon General had reviewed enough evidence to issue a formal advisory on social media and youth mental health.
The document was deliberately cautious.
It recognised benefits including connection, support, identity formation, creativity and access to information.
It emphasised that effects differed according to the child, content, design, context and pattern of use.
It also concluded that there was not enough evidence to say social media was sufficiently safe for children and adolescents.
The advisory cited evidence linking more than three hours of daily social-media use among adolescents with double the risk of poor mental-health outcomes in one longitudinal study, while acknowledging that much of the wider research remained correlational and that causation was difficult to establish.
This is the position the evidence demands.
Concern without certainty.
Action without pretending every question has been answered.
The lack of proof that social media caused a particular mental-health trend does not prove the products are safe.
The existence of association does not prove the products caused the trend.
Both overstatement and complacency are intellectually dishonest.
For companies, however, the standard cannot be identical to the standard applied to an uninvolved observer.
They control the intervention.
They collect the richest data.
They run the experiments.
They design the defaults.
Their duty begins before universal scientific consensus.
How Much Proof Should Be Required?
Suppose a company suspects that a feature may worsen body-image distress among some teenage girls.
Waiting for perfect evidence could take years.
Removing the feature immediately could deny benefits, damage expression and produce unintended consequences.
There is a space between those extremes.
Reduce exposure among younger users.
Test an alternative.
Remove public metrics by default.
Create stronger recommendation limits.
Introduce natural stopping points.
Share anonymised data with independent researchers.
Measure regret, not only engagement.
Inform parents honestly about known uncertainties.
Warn vulnerable users.
Design more conservatively while evidence develops.
This is the principle of precaution.
It does not demand that innovation stops whenever someone raises a concern.
It asks organisations to respond proportionately when potential harm is serious, the affected group is vulnerable and the company has enough control to reduce risk.
Children should not carry the burden of proving damage conclusively before the system protecting them changes.
The Company Knows Which Change Costs Money
External researchers can recommend safer design.
Only the company usually knows the full commercial consequence.
What happens to engagement if visible like counts disappear?
What happens to session length if recommendations stop earlier?
What happens to revenue if teenagers receive fewer notifications?
What happens to growth if age verification becomes stricter?
What happens to creator activity if potentially harmful but highly engaging material receives less distribution?
What happens to advertiser value if less data is collected?
These are not abstract questions inside the company.
They can be measured.
That is why internal evidence matters.
The company may know both that a feature creates risk and that changing it will reduce performance.
The moral decision occurs where those two datasets meet.
Profit Does Not Need to Be Spoken Aloud
There may be no meeting in which an executive says:
Accept the harm because the revenue matters more.
Organisations rarely operate through such blunt declarations.
The trade-off can happen indirectly.
A safety team proposes a change.
The evidence is considered insufficient.
More testing is requested.
The rollout is delayed.
A narrower intervention is chosen.
The feature becomes optional rather than default.
Responsibility is assigned to parents.
A warning appears instead of a redesign.
The company waits for industry-wide standards so it will not move alone.
Every step can be defended.
The commercial outcome remains unchanged.
Delay favours the existing model.
This is how profit can win without anyone explicitly voting for harm.
The burden of proof is simply set higher for the change that might reduce growth.
What Did Executives Personally Know?
This is one of the most difficult questions.
A company employing tens of thousands of people produces vast amounts of research, experimentation and internal discussion.
The existence of one document does not prove every senior executive read it.
An employee's warning does not establish that leadership accepted the conclusion.
A research team may see risk differently from a product team.
The phrase Facebook knew or TikTok knew can hide enormous organisational complexity.
Companies are not single minds.
Accountability must therefore examine communication.
Who received the report?
Which meetings discussed it?
What recommendations were made?
What decision followed?
Which executive possessed authority?
Was the issue escalated?
Did leadership ask for more evidence?
Did a commercial objective influence the response?
Without that chain, broad corporate knowledge can be difficult to convert into personal responsibility.
This is one reason litigation and regulatory disclosure matter.
They can compel documents, testimony and timelines that public relations statements do not provide.
Courtrooms Are Beginning to Test the Claims
In March 2026, a Los Angeles jury found Meta and Google's YouTube negligent in the design or operation of products used by a young plaintiff and concluded that the companies' conduct contributed substantially to her mental-health injuries.
The jury awarded $3 million in compensatory damages and a further $3 million in punitive damages, allocating responsibility between Meta and YouTube.
The case was the first major social-media addiction product-design claim to reach a United States jury.
Both companies denied that their products caused the plaintiff's conditions and have appealed.
The verdict is important.
It is not the final scientific answer.
A jury decides the specific case presented to it under a legal standard.
It does not establish that social media causes mental illness generally.
The verdict may be altered or reversed on appeal.
Other cases may reach different conclusions.
But for the first time, a jury examining extensive evidence concluded that platform design itself, rather than only third-party content, could create legal responsibility for harm experienced by a young user.
That changes the accountability landscape.
Legal Liability Is Not the Same as Moral Responsibility
Companies often wait for law to define the boundary.
That is understandable.
Clear rules create consistency.
They prevent one company from accepting a disadvantage while competitors continue unchanged.
But legality is a minimum standard.
A design can be lawful and still irresponsible.
A risk can be real before a court recognises it.
A child can be harmed before regulation catches up.
Moral responsibility begins where power, knowledge and foreseeable consequence meet.
Technology companies possess extraordinary power.
Their knowledge is incomplete but far greater than that of the ordinary user.
Some consequences are now foreseeable.
The argument can no longer rest entirely upon whether a particular law prohibited a particular feature at the moment it was built.
They Also Knew the Products Helped People
The evidence cannot be reduced to harm.
Meta's research found teenagers who believed Instagram made loneliness, anxiety, sadness and other difficulties better.
The Surgeon General documented benefits for friendship, connection, identity, creativity and support, particularly for some marginalised young people.
TikTok has created routes for education, cultural participation, creativity and community.
YouTube has placed an extraordinary library of instruction and experience within reach of children around the world.
The companies knew this too.
That knowledge matters because it complicates the moral decision.
Redesigning or restricting a product can remove benefits alongside risks.
A strong safety measure can isolate a vulnerable teenager from a community that matters to them.
Age verification can protect children while increasing privacy intrusion.
Parental controls can support families while creating danger for children living with controlling adults.
Recommendation limits can reduce harmful content and make valuable minority content harder to discover.
The right answer is rarely remove everything.
The existence of trade-offs does not justify doing nothing.
It requires better design and better evidence.
Safety Became a Competitive Feature
By the middle of the 2020s, platforms were no longer presenting safety merely as a regulatory obligation.
They were marketing it.
Teen Accounts.
Family Pairing.
Private defaults.
Sleep reminders.
Content controls.
Time limits.
Age-assurance systems.
Safety became part of the product proposition.
This is positive.
Markets can reward companies for protecting users.
Parents may prefer services with stronger safeguards.
Advertisers may avoid unsafe environments.
Reputation can align with responsibility.
But marketing safety creates another duty.
The claims must reflect how the system operates in practice.
A sixty-minute default means little if the barrier is easily dismissed and the platform knows most teenagers continue.
A private account does not prevent harmful recommendation.
A parental dashboard does not protect a child whose parent never enables it.
The feature announcement is not the outcome.
The relevant question is what happened after deployment.
Did exposure decrease?
Did sleep improve?
Did unwanted contact fall?
Did vulnerable users receive less harmful material?
Did the company publish the results?
Safety should be measured through harm reduction, not feature count.
Why Did It Take Exposure to Accelerate Change?
This question runs through the evidence.
Meta had researched teen wellbeing before 2021.
It had already introduced safety features.
After the internal documents became public, product changes accelerated and received greater prominence.
TikTok had information about teen use before state litigation.
It had safety tools already.
Later, defaults, nighttime protections and family controls expanded.
This pattern does not prove companies acted only because they were exposed.
Large product changes take time.
Technical capacity evolves.
Evidence becomes stronger.
Public expectations change.
Regulation creates consistency.
The pattern does show that external pressure matters.
Whistleblowers.
Journalists.
Parents.
Researchers.
Regulators.
Courts.
Public scrutiny alters the cost of inaction.
Inside the business model, safety may reduce engagement.
Outside pressure creates another calculation.
Failing to act may damage trust, invite regulation, create liability and threaten long-term growth.
Accountability changes the incentive.
Whistleblowers Should Not Be the Safety System
Society should not depend upon an individual employee secretly copying thousands of documents before regulators understand a product used by billions.
Whistleblowing is risky.
The employee may be wrong.
Selective documents may produce distorted conclusions.
Confidentiality can protect legitimate interests.
The public cannot build a stable safety regime around exceptional insiders choosing to disclose.
The system needs ordinary transparency.
Independent access.
Standardised risk reporting.
Auditable design changes.
Legal protection for good-faith internal concerns.
Research teams with enough authority to stop or alter launches.
Regulators capable of understanding machine-learning systems.
Clear records showing how known risks were assessed.
A healthy industry should not require betrayal of the company to make the company accountable.
We Need to Know What Was Not Measured
Internal research can reveal risk.
Its absence can conceal risk.
Companies measure what the business model makes visible.
Engagement.
Retention.
Clicks.
Views.
Advertising performance.
Growth.
They can also measure safety events, survey satisfaction and reported harms.
More difficult outcomes remain outside the platform.
Sleep lost.
Homework abandoned.
Family conflict.
Attention fragmented.
Anxiety experienced but never reported.
The conversation not held.
The child who quietly became more dissatisfied with their body.
The adult who regrets an hour but continues behaving similarly tomorrow.
If the company does not measure those effects, it may genuinely know less about them.
That ignorance is not necessarily innocent.
A system chooses what to investigate.
Once credible concern exists, failing to measure the possible cost can become a decision of its own.
Known Harm Is Not the Only Standard
A company does not need certainty that a feature harms users before accepting responsibility for risk.
Manufacturers test products because plausible failure matters.
Medicine monitors adverse events because rare effects matter.
Aviation investigates near misses because waiting for catastrophe would be absurd.
Digital products have often been treated differently.
Launch widely.
Observe behaviour.
Respond when evidence accumulates.
That approach may be acceptable for a new button colour.
It becomes harder to defend when the experiment concerns the emotional development of children.
The scale of deployment should increase caution.
Not reduce it.
The Standard Should Rise With Power
A small creator cannot conduct a global safety study before posting a video.
A parent cannot audit a recommendation model.
A school cannot inspect the data flows behind every application.
A technology company operating at global scale can do more.
It can test alternative designs.
Employ researchers.
Consult specialists.
Restrict high-risk features by age.
Monitor unintended effects.
Publish transparency reports.
Share protected data.
Accept slower growth.
The greater the capability, the weaker the defence that the user should manage the risk alone.
Power does not make perfect knowledge possible.
It makes deliberate ignorance less acceptable.
They Knew Enough
What, then, did the companies know?
They knew their products could produce extraordinary benefits.
They knew engagement was commercially valuable.
They knew particular design choices changed behaviour.
They knew children and teenagers used their services extensively.
They knew some young users reported body-image pressure, compulsive use, bullying, harmful content, unwanted contact and sleep disruption.
They knew vulnerable users could experience the same platform differently from the majority.
They knew recommendation systems shaped exposure.
They knew defaults mattered.
They knew notifications brought people back.
They knew protective changes could alter usage.
They did not know the complete causal effect of social media upon an entire generation.
Nobody does.
That missing certainty matters scientifically.
It does not erase what was already known morally.
The Question That Remains
Perhaps every company believed it was responding responsibly.
Perhaps employees genuinely disagreed about the evidence.
Perhaps some proposed changes created harms of their own.
Perhaps public criticism underestimated the benefits and exaggerated causation.
Perhaps safety teams achieved far more than outsiders understand.
All of that may be true.
The investigation still returns to one question:
When a company possessing unmatched behavioural evidence discovered that its profitable product might be harming a vulnerable minority, did it move as quickly and decisively as it would have moved if the same risk threatened revenue?
That is the test.
Not perfection.
Not hindsight.
Not absolute safety.
Priority.
Final Thought
Nobody meant to build the attention economy.
That explains the beginning.
It does not excuse the middle.
At first, the harms were uncertain.
Then researchers found patterns.
Employees raised concerns.
Teenagers described their experiences.
Internal documents recorded risks.
Safety teams proposed interventions.
Whistleblowers disclosed what outsiders could not see.
Regulators gathered evidence.
Courts began hearing cases.
The companies responded.
They introduced limits, warnings, private defaults, parental controls, recommendation safeguards and age-specific experiences.
Those changes prove something important.
The systems could be changed.
The architecture was never inevitable.
The question is why so many protections arrived only after the risk became public, political, legal or reputational.
They did not know everything.
They still do not.
They knew enough to understand that some users were being harmed.
They knew enough to recognise that design played a role.
They knew enough to build protections.
They knew enough to know the protections might reduce the behaviour their businesses had learned to value.
That changed the question.
The attention economy may have been created accidentally.
But once the people running it could see the consequences, every delay became a choice.
Sources & Research Gaps
Principal Sources
Meta, Internal Teen Wellbeing Research and Public Response
Meta published its explanation of two internal research decks in September 2021.
The company argued that reporting had overstated the findings and removed important context.
Meta acknowledged that approximately one third of surveyed teenage girls already experiencing body-image problems said Instagram made the issue worse, while 22 per cent said it improved the experience and 45.5 per cent reported no effect.
Meta also said that, in eleven of twelve wellbeing categories, more affected teenage girls reported improvement than deterioration.
The company emphasised that much of the research was exploratory, based partly on small samples and unable to establish causation.
Frances Haugen's Senate Testimony
Haugen's written testimony alleged that Facebook repeatedly prioritised profit over safety, possessed internal evidence concerning harms and withheld important information from the public and regulators.
Her testimony is direct evidence of what she alleged and believed based upon her experience.
It is not an independent judicial finding proving every allegation.
Meta Safety and Product Announcements
Meta's public announcements document protective changes including:
- private defaults for teenage accounts;
- restrictions upon adult contact with teenagers;
- Take a Break reminders;
- parental supervision tools;
- reduced recommendation of sensitive content;
- nudges towards different topics;
- Instagram Teen Accounts;
- expansion of Teen Accounts to Facebook and Messenger.
These announcements demonstrate that the company introduced substantial safeguards.
Independent evaluation is still required to determine how effectively each measure reduced real-world harm.
Meta, Instagram Kids
Meta's September 2021 announcement explains its decision to pause the proposed Instagram experience for children under thirteen.
The company argued that a supervised, age-specific product could be safer than allowing children to enter the adult platform by misrepresenting their age.
Critics viewed the proposal as an expansion of Meta's ecosystem towards younger users.
United States Surgeon General, Social Media and Youth Mental Health
The 2023 advisory reviewed positive and negative evidence.
It concluded that current evidence was insufficient to determine that social media was adequately safe for children and adolescents.
The advisory also identified major limitations caused by lack of transparency and limited researcher access to company data.
It did not conclude that social media universally harms young people or that platform use alone explains the youth mental-health crisis.
Federal Trade Commission, A Look Behind the Screens
The FTC's 2024 staff report examined information supplied by major social-media and video-streaming companies.
It concluded that many collected and retained extensive personal data, fed information into automated systems and provided inadequate or inconsistent protection for children and teenagers.
The report also linked targeted-advertising business models with incentives for extensive surveillance.
New York v TikTok Complaint
New York's 2024 complaint alleges that TikTok knew its design could create compulsive use, understood the particular vulnerability of minors and used recommendations, notifications, badges and social rewards to increase engagement.
The public complaint contains substantial redactions.
Its claims remain allegations unless admitted, settled or proven through litigation.
Some descriptions of neuroscience and addiction in the complaint are simplified and should not be treated as settled scientific conclusions.
TikTok Teen Safety Announcements
TikTok documents the introduction of:
- a default sixty-minute daily screen-time limit for users under eighteen;
- Family Pairing;
- private teenage accounts;
- restrictions on nighttime notifications;
- communication controls;
- content filters;
- age-assurance measures;
- more than fifty preset teen protections by 2026.
These measures demonstrate active safety work and the technical ability to change platform defaults.
They do not resolve the allegations concerning earlier designs or determine the effectiveness of every safeguard.
K.G.M. v Meta and YouTube
In March 2026, a Los Angeles jury found Meta and YouTube negligent in a case involving one young plaintiff and awarded $6 million in compensatory and punitive damages.
The verdict concerned the specific evidence and legal standards presented in that case.
It does not establish universal causation between social media and mental illness.
Meta and YouTube have appealed, meaning the outcome is not legally final.
Research Gaps and Limitations
The phrase they knew refers to corporate knowledge distributed across researchers, safety employees, product teams, executives and internal documents.
It does not mean every employee or executive possessed the same information.
The existence of an internal document does not prove that senior leadership read, accepted or acted upon it.
A complete accountability analysis requires evidence showing:
- who received the research;
- how the findings were interpreted;
- which recommendations were made;
- who possessed authority to act;
- what decision followed;
- whether financial consequences influenced that decision.
Meta's internal teen research was not designed to establish population-level causation.
Some studies used small or deliberately selected samples.
The findings relied heavily upon self-report.
A teenager's belief that Instagram improved or worsened an experience is important but does not independently prove causal effect.
The wider scientific literature remains mixed.
Social media may benefit and harm different people under different circumstances.
The direction of causation may be bidirectional.
Young people experiencing distress may use platforms differently, while platform experiences may also influence distress.
Legal complaints contain allegations.
They should not be described as established facts until tested through a court process or admitted by the defendant.
Government reports may reflect staff interpretation and can be challenged by companies or individual commissioners.
Product announcements document what companies say they introduced.
They do not prove that features worked as intended or reached every eligible account.
The existence of safety changes does not prove earlier designs were negligent.
Technology, research, public expectations and regulatory requirements evolve.
Later safeguards may have been impractical or unavailable earlier.
The 2026 jury verdict concerns one plaintiff.
It is subject to appeal and should not be treated as definitive scientific evidence concerning every platform or young user.
Further independent research and disclosure are needed concerning:
- internal timelines showing when specific risks first became visible;
- the commercial impact of proposed safety changes;
- which protective interventions were delayed, rejected or narrowed;
- whether engagement targets affected safety decisions;
- the effectiveness of Teen Accounts and screen-time defaults;
- the number of teenagers who bypass or disable protective limits;
- how recommendation safeguards affect harmful-content exposure;
- whether public metrics such as likes worsen body-image or social-comparison outcomes;
- how companies measure long-term wellbeing and post-session regret;
- the extent to which internal researchers can halt or modify product launches;
- what information boards and senior executives receive about youth risk;
- the quality and independence of company-funded research;
- secure methods for allowing researchers to analyse platform data;
- the effects of pending appeals and future social-media product-liability trials;
- whether regulation changes design globally or only within particular jurisdictions.
This essay establishes that companies accumulated enough evidence to recognise meaningful risks.
It does not claim that they possessed complete proof, intended all harms or failed to introduce important safety measures.