THE GREAT ATTENTION EXPERIMENT: 9 - WE ARE LIVING INSIDE THE LARGEST BEHAVIOURAL EXPERIMENT IN HUMAN HISTORY
Before you have properly woken tomorrow morning, you may already have contributed new information to one of the largest behavioural systems ever created.
You reach for your phone.
The time is recorded.
You open one notification but ignore another.
You check a message.
Pause over a photograph.
Read a headline.
Watch part of a video.
Leave before it finishes.
Return to something you previously ignored.
Search for a product.
Close an application.
Open it again ten minutes later.
None of those actions feels like participating in an experiment.
You were not invited into a laboratory.
Nobody handed you a consent form.
No researcher explained the hypothesis.
You were not told which group you had entered.
There is no white coat.
No clipboard.
No obvious control condition.
You are simply using technology.
Behind the visible experience, however, the system is learning.
It records what appeared.
What you selected.
What you rejected.
How long you remained.
Which prompt brought you back.
Which arrangement produced more activity.
Which recommendation succeeded.
Which prediction failed.
Your behaviour becomes evidence.
That evidence changes what the system does next.
The next version may be shown to you.
Or to somebody whose behaviour resembles yours.
The process repeats throughout the day.
Across billions of people.
Through thousands of products.
Inside homes, workplaces, schools, shops, hospitals, vehicles and bedrooms.
This is why I describe modern digital life as the largest behavioural experiment in human history.
Not because one company planned it.
Not because humanity has been placed inside one controlled scientific study.
Not because every digital interaction is sinister.
The phrase describes something less organised and far larger.
For the first time, technologies used by a substantial proportion of humanity can observe behaviour continuously, alter the environment in response, compare outcomes across enormous populations and deploy successful changes almost instantly.
We are not standing outside that process.
We are the environment in which it is happening.
This Is Not an Experiment in the Strict Scientific Sense
The word experiment must be used carefully.
A scientific experiment normally begins with a defined question.
Researchers manipulate one or more variables.
Participants are assigned to different conditions.
Outcomes are measured.
Controls help distinguish the effect of the intervention from everything else that might have caused the result.
Formal human-subject research may also require ethical review, informed consent and additional protections depending upon the research, institution and jurisdiction.
In the United States, for example, the Common Rule establishes requirements covering institutional review boards, informed consent and compliance for federally conducted or supported human-subject research within its scope.
The Great Attention Experiment has no single protocol.
No universal hypothesis.
No agreed outcome.
No fixed beginning.
No planned end.
Nobody randomly assigned one half of civilisation to grow up with personalised feeds while the other half continued without them.
No independent committee examined the complete proposition before billions of people were exposed.
The phrase is therefore a metaphor.
But it is not merely dramatic language.
Inside the wider social experiment are countless real experiments.
Interfaces are tested.
Notifications are compared.
Feeds are reordered.
Recommendations are adjusted.
Messages are rewritten.
Friction is removed or restored.
Different people are shown different versions.
Their behaviour determines which version survives.
The industry calls this product development.
Optimisation.
Personalisation.
A/B testing.
Experimentation.
Whatever language we choose, the process involves changing an environment, observing human response and using the result to influence what happens next.
The Scale Has No Historical Precedent
The International Telecommunication Union estimated that six billion people were online in 2025, representing approximately 74 per cent of the world's population.
Meta alone reported that an average of 3.58 billion people used at least one of Facebook, Instagram, Messenger or WhatsApp each day during December 2025.
That figure does not mean 3.58 billion people were all scrolling through the same feed.
Some were sending messages.
Speaking to family.
Operating businesses.
Sharing photographs.
Participating in communities.
Using the products briefly and deliberately.
The significance is scale.
A product change affecting a tiny proportion of that population can still affect millions of people.
An additional second of average use, multiplied across billions of daily users, becomes decades of collective human attention.
A small shift in the probability of clicking, sharing, replying or continuing can become commercially and socially significant when repeated across a population larger than any country.
Previous forms of media reached enormous audiences.
Radio broadcasts crossed nations.
Television events were watched by hundreds of millions.
Newspapers shaped political and cultural understanding.
The difference is that those systems could not normally observe each person's response and alter the next broadcast accordingly.
Modern digital systems can.
The audience does not merely receive the environment.
Its behaviour trains it.
Actual Experiments Have Already Reached Tens of Millions
In 2012, Facebook described running hundreds of tests every day.
Most were released to random samples of users so the company could measure their effects.
Some features requiring network effects were tested across an entire market, including potentially a whole country.
Many such experiments are ordinary and beneficial.
Does this page load faster?
Can people find a privacy control?
Does a new accessibility feature work?
Does the application crash less often?
Do users understand the wording?
Responsible companies should test products rather than assume they function correctly.
Other experiments have demonstrated how far platform influence can extend beyond interface preference.
During the 2010 United States congressional elections, researchers randomly assigned different political messages to 61 million Facebook users.
One version displayed a civic message with photographs of friends who had said they voted.
Another displayed an informational message without the social element.
A control group received no message.
The researchers reported effects upon political expression, information seeking and validated real-world voting behaviour.
They also found that influence spread through close social relationships beyond the people who received the original message.
The effect upon any single individual was small.
That does not make the experiment insignificant.
At a population scale, a small behavioural effect can change the actions of hundreds of thousands of people.
Two years later, another study attracted even greater ethical controversy.
Researchers altered the amount of emotionally positive or negative material appearing in the feeds of 689,003 Facebook users.
Those shown less positive material subsequently used slightly more negative and fewer positive words in their own posts.
Those shown less negative material displayed the opposite pattern.
The study concluded that emotional expression could spread through an online social network without direct interaction between people.
The measured effects were small.
The ethical question was not.
Participants had not been explicitly told that the emotional balance of their feeds was being manipulated for research.
The authors relied upon Facebook's data-use policy as the basis for participation.
The Proceedings of the National Academy of Sciences later published an editorial expression of concern, noting that informed consent and the ability to opt out are normally best practice in most social-science research involving people.
The controversy revealed a gap.
A platform could alter the information environment of hundreds of thousands of people as part of ordinary product operation.
The ethical expectations became less clear when the same activity was studied and published as academic research.
The manipulation was technically possible before the paper existed.
Publication merely made the process visible.
Product Development and Human Research Now Overlap
A company changes the colour of a button.
It observes whether more people press it.
That is product testing.
A company changes which political message appears.
It measures whether people vote.
That begins to resemble behavioural research.
A platform reduces positive material in a feed and studies emotional expression.
That is clearly an experiment involving human behaviour.
The boundaries are no longer simple.
Digital products are behavioural environments.
Changing the product changes what people encounter.
What people encounter can change what they do.
The result generates data.
That data influences the next product decision.
The loop can support scientific knowledge, commercial optimisation and ordinary engineering simultaneously.
Formal research ethics developed around studies that researchers could identify clearly.
A trial begins.
Participants enter.
The intervention is defined.
The study ends.
Digital experimentation can be continuous.
The user may not know which features are experimental, permanent, personalised or being withdrawn.
The platform may not need to classify every test as human-subject research under a particular legal framework.
That does not mean the company has no ethical obligation.
Law defines which formal procedures apply.
Ethics asks a wider question:
When does changing a product become an intervention in a person's life?
Terms and Conditions Are Not the Same as Understanding
Most digital services operate under terms, privacy notices and data policies.
Users agree by clicking a button, creating an account or continuing to use the service.
Those agreements matter legally.
They communicate rules and authorise forms of data processing.
They do not automatically create meaningful understanding.
Informed consent within formal research is intended to be a process through which people receive understandable information and voluntarily decide whether to participate.
The United States Office for Human Research Protections describes consent as more than a form, emphasising that information should enable a thoughtful and voluntary decision.
Most people do not approach a social application in the same way.
They want to message a friend.
Watch a video.
Join a community.
Read an article.
Upload a photograph.
The notice appears between them and the thing they came to do.
They click.
The practical meaning of the agreement may include data collection, algorithmic inference, product experimentation, advertising and personalisation across years of use.
The user has technically encountered the information.
That does not prove they understood the behavioural system they were entering.
Children make this distinction impossible to ignore.
A child may learn to tap through a digital pathway years before they can understand a privacy notice.
The action can be recorded.
The developmental capacity required to evaluate the exchange does not yet exist.
The Experiment Has No Single Hypothesis
Scientific experiments usually test something defined.
The digital experiment tests almost everything that can be measured.
Will a larger image attract more attention?
Will this wording produce more clicks?
Will an earlier notification increase returns?
Will visible approval generate more posting?
Will an automatically starting video reduce exits?
Will shorter content increase completion?
Will longer content produce more total viewing?
Will showing friends' activity change behaviour?
Will reducing friction increase spending?
Will a different recommendation create a new interest?
Will this person respond more strongly to fear, humour, aspiration, outrage or belonging?
The company does not need one grand theory of humanity.
It can test alternatives.
The system keeps the version producing the desired metric.
What matters is the objective.
A platform optimising for completed tasks may learn one set of lessons.
A platform optimising for advertising impressions may learn another.
A service optimising for purchases may discover how to reduce hesitation.
A workplace tool optimising for activity may encourage more visible communication.
An AI companion optimising for continued conversation may learn which responses make departure less likely.
The experiment does not ask only how people behave.
It asks how behaviour can be changed.
The Experiment Is Distributed
No single company controls the whole system.
Search engines test results.
Retailers test prices, layouts and recommendations.
Streaming platforms test thumbnails and programme ordering.
Games test progression, rewards and spending pathways.
Social networks test feeds, prompts and public metrics.
News publishers test headlines.
Advertisers test images, language and targeting.
Employers test monitoring, workflow and communication software.
Schools test learning platforms.
Creators test openings, video length and emotional tone.
Users test versions of themselves.
The experiments interact.
A creator learns which content the platform rewards.
The platform learns which creator retains viewers.
The advertiser learns which audience responds.
The viewer learns which behaviour earns social approval.
The employer learns who answers fastest.
The employee learns that availability is rewarded.
The parent learns that a screen stops a public meltdown.
The child learns that boredom can be removed immediately.
No central authority designed the complete result.
The participants alter one another.
This makes the system harder to control than a conventional experiment.
There is no single laboratory to close.
The Participant Is Also Producing the Equipment
In an ordinary experiment, the researcher provides the instrument.
In digital life, users continually help build the instrument measuring them.
Photographs improve recognition systems.
Searches reveal language and intention.
Clicks train ranking models.
Messages reveal social relationships.
Viewing histories shape recommendations.
Purchases improve prediction.
Corrections improve artificial-intelligence outputs.
Reports help identify harmful content.
The person's contribution may create real value for them.
A music service improves its recommendations.
A map learns better routes.
An email filter identifies spam.
A language model becomes more useful.
The same contribution strengthens the system's ability to predict and influence future behaviour.
The user is simultaneously:
- the person receiving the product;
- the source of behavioural evidence;
- part of the environment influencing other users;
- a contributor to the model's improvement;
- and sometimes the commercial audience being sold to advertisers.
That combination has no close historical equivalent at this scale.
Small Effects Become Large Social Forces
One defence frequently offered is that the measured effects of individual features are tiny.
Often, that is true.
A redesigned button might increase clicks by a fraction of a percentage point.
A social message may influence only a small proportion of voters.
A change in feed composition may produce barely detectable shifts in emotional language.
But products used at vast scale do not require large individual effects.
Suppose a feature changes the behaviour of one person in every thousand.
In a population of one million, that affects one thousand people.
Across one billion, it affects one million.
The size of the individual effect and the size of the societal consequence are different questions.
A small effect repeated occasionally may still be negligible.
A small effect repeated daily, across years, during childhood and adulthood, deserves closer attention.
The experiment is not one intervention.
It is an accumulation.
One notification does not transform a life.
Thousands may change the rhythm through which attention is allocated.
One recommendation may do nothing.
Repeated recommendations can change what feels familiar, important or normal.
One interruption does not destroy concentration.
A workplace organised around continuous interruption can change how work is performed.
The cumulative outcome may be far larger than any individual product test was designed to detect.
There Is No Unaffected Control Society
Inside digital companies, experiments may contain excellent control groups.
One group receives a feature.
Another does not.
The company compares behaviour.
At the level of society, the control group is missing.
We cannot observe the same generation growing up twice.
Once with personalised feeds.
Once without them.
We cannot give one version of an adult permanent workplace connectivity and allow the identical person to live the alternative life.
We cannot know precisely which conversation would have happened if the notification had not arrived.
Which interest would have developed if the recommendation had been different.
Which purchase would not have been made.
Which hour would have been spent elsewhere.
Researchers can compare populations.
Use longitudinal data.
Study policy changes.
Conduct controlled experiments.
Examine natural variations.
Each method helps.
None creates a complete parallel society untouched by the technology.
Even people who avoid particular platforms live among others whose behaviour has been shaped by them.
Politics changes.
Markets change.
Language changes.
Work expectations change.
Schools change.
Friendship changes.
The non-user still inhabits the same culture.
The experiment leaks into the control group.
The Output Changes the Participant
A recommendation system learns from behaviour.
The recommendation also affects the behaviour later used as evidence.
The feed presents a topic.
The person watches.
The watch becomes proof of interest.
More of the topic appears.
The person becomes more familiar with it.
The stronger response confirms the profile.
The system has not simply discovered preference.
It has participated in reinforcing it.
The same loop applies beyond content.
Visible popularity changes what people perceive as desirable.
Fast replies change expectations of responsiveness.
Personalised prices can change purchasing.
AI-generated answers can affect the human judgements later used to train or evaluate AI.
The participant is not a fixed subject being measured from outside.
They are changing inside the measurement process.
The system adapts to the person.
The person adapts to the system.
This is why separating discovery from influence becomes so difficult.
The Researchers Have a Commercial Interest in the Outcome
Scientific conflicts of interest do not prove research is false.
They create a reason for stronger scrutiny.
The attention economy contains an unavoidable conflict.
Companies need to understand behaviour to improve their products.
Many of those products generate more revenue when particular behaviours increase.
The researcher and the commercial beneficiary may be parts of the same organisation.
The Federal Trade Commission reported in 2024 that the social-media and video-streaming companies it examined commonly used algorithms, data analytics or artificial intelligence for recommendations, personalisation, advertising and engagement.
Most derived revenue directly or indirectly from these automated systems.
The FTC also found that companies commonly ingested personal information into such systems by default, often without giving people a comprehensive ability to control or reject those uses.
The FTC did not find one uniform standard of oversight.
Some companies reported substantial teams concerned with ethics, fairness, privacy and bias.
Others had weaker or fragmented arrangements.
Even where internal oversight existed, the regulator said it was not always clear whether recommendations from those teams were binding.
Testing was described as regular or continuous, but its frequency, scope and methods varied widely.
This does not mean companies cannot research their own products honestly.
They possess expertise and data independent researchers cannot replicate.
The conflict lies in deciding which questions receive attention.
The business naturally measures whether engagement increased.
It must make a deliberate additional effort to measure what the engagement cost.
The System Sees the Counterfactual
An individual sees what happened.
The company may see what would probably have happened under another design.
You receive one notification.
The company knows whether people receiving different wording returned more often.
You see one feed.
The company can compare it with thousands of alternatives.
You stop watching after ten minutes.
The company may know that users shown another sequence averaged twelve.
You complete a purchase.
The retailer may know which friction removed your final hesitation.
This counterfactual knowledge creates power.
The user knows their experience.
The system knows how groups behaved when the experience changed.
That does not make the prediction perfect.
Average effects do not determine every individual.
Product tests can be misleading.
Short-term improvements may damage long-term trust.
But the person cannot independently reproduce the comparison.
They do not see the version of themselves who received no prompt.
No autoplay.
No public metric.
No recommendation.
No personalised advert.
The company sees the experiment.
The user sees life.
Internal Research Changed the Moral Position
For many years, companies could argue reasonably that the effects of unprecedented products were uncertain.
Then they began studying them.
Meta's internal research into teenagers and Instagram, later made public amid whistleblower disclosures and congressional scrutiny, documented both positive and negative experiences.
The company said that many teenagers reported Instagram helped them through loneliness, anxiety, sadness and other difficulties.
It also acknowledged that, among teenage girls who already experienced body-image concerns, approximately one third in one survey reported that Instagram made those concerns worse.
That finding did not prove Instagram caused body-image problems.
It did not apply to one third of every teenage girl.
It was based upon self-reported experience.
The research was exploratory and needed careful interpretation.
Its significance lies elsewhere.
The company could see differences hidden by an average engagement number.
The same product helped some users and appeared to worsen an existing difficulty for others.
Internal research gave the company a view of vulnerable subgroups that ordinary users, parents and many independent researchers did not possess.
The United States Surgeon General later called upon technology companies to share relevant health and wellbeing data with independent researchers and the public while protecting privacy, reflecting wider concern that critical evidence remained concentrated inside the companies operating the systems.
The moment a company sees credible evidence of risk, the ethical question changes.
It does not need to know everything.
It knows enough to investigate further, alter the experiment or explain why it has chosen not to.
This Is Not Evidence of a Secret Master Plan
The experiment is too distributed, inconsistent and commercially competitive to support a simple conspiracy.
Companies disagree.
Teams inside the same company disagree.
Products fail.
Experiments produce unexpected results.
Algorithms misunderstand people.
Safety interventions improve some outcomes and create other problems.
Regulators intervene differently across countries.
Users adapt in ways designers did not predict.
The system is powerful.
It is not omniscient.
There is no single group controlling the thoughts of humanity.
The more accurate concern is structural.
Organisations test what improves the outcomes upon which they depend.
Advertising businesses test engagement, targeting and response.
Retailers test conversion.
Employers test productivity and monitoring.
Creators test attention.
Political campaigns test persuasion.
Each participant behaves rationally inside its own incentives.
The combined result can still become irrational for society.
No conspiracy is required.
The metrics provide direction.
The Experiment Produced Genuine Benefits
The word experiment should not imply that humanity has received nothing in return.
Six billion people are not online merely because they were tricked into participating.
Digital technology solves real problems.
Families remain connected across continents.
People access education regardless of location.
Creators reach audiences without traditional gatekeepers.
Small businesses reach customers.
Communities form around rare conditions, identities and interests.
Emergency information travels quickly.
Translation lowers barriers.
Maps reduce uncertainty.
Online banking increases access.
Remote work creates flexibility.
Recommendation systems help people find information they would never have discovered alone.
Behavioural experimentation can improve all of these services.
A clearer interface reduces confusion.
Better accessibility includes people previously excluded.
A safety test identifies harmful pathways.
A recommendation improves learning.
A reminder helps someone take medication.
A platform discovers that one protection works better than another.
Experimentation is not the enemy.
Experimenting upon people without sufficient transparency, limits or responsibility is the concern.
The Experiment Does Not Affect Everyone Equally
Averages conceal vulnerability.
A feature that produces little effect in most adults may affect a child differently.
A visible popularity metric may be amusing to one person and emotionally significant to another.
A recommendation system may broaden one user's interests while narrowing another's world.
A notification may be ignored by someone with control over their time and impossible to disregard by an employee expected to remain available.
A personalised advert may be useful to an informed adult and exploitative when directed towards insecurity, distress or developmental immaturity.
The experiment has billions of participants.
It does not have one standard participant.
Age.
Temperament.
Disability.
Neurodivergence.
Mental health.
Income.
Culture.
Education.
Power.
Employment.
Family environment.
Each changes the meaning of the same intervention.
An average positive outcome does not prove the system is safe for the vulnerable minority.
An identified harm to a minority does not prove the entire service lacks value.
Responsible experimentation must be capable of holding both conclusions at once.
Children Turn the Metaphor Into an Ethical Emergency
Children did not design these environments.
They do not control the business model.
They cannot interpret every commercial incentive.
Many cannot understand the data being collected or the inferences being made.
Yet their behaviour can be observed, classified and used to shape what appears next.
They are participants in an evolving system while still developing the abilities needed to evaluate it.
This makes the absence of prior risk assessment increasingly difficult to defend.
In March 2026, Ofcom told regulated platforms that recommendation algorithms were children's main pathway to harmful online material.
It issued legally binding information requests to examine major systems and emphasised that platforms must assess the risks of significant design changes before deploying them.
Ofcom described this explicitly as an end to product testing on children.
That language matters.
The regulator is not claiming that every product update constitutes a scientific experiment.
It is recognising the underlying principle.
Children should not become the population through which major risks are discovered after release.
The safety question must be asked before the behavioural evidence arrives.
Artificial Intelligence Expands the Experiment
The next phase will not consist only of feeds selecting existing content.
Artificial intelligence can generate the content itself.
A recommendation system chooses which video to show.
A generative system can create the message, explanation, image, voice or character most likely to suit the individual.
It can respond to language.
Adjust tone.
Remember prior conversations.
Offer reassurance.
Teach.
Persuade.
Sell.
Entertain.
Remain available at any hour.
This capability can strengthen human agency.
An AI system can summarise noise.
Help someone understand a difficult subject.
Translate communication.
Reduce repetitive work.
Support disability.
Help a child explore curiosity with guidance.
It can also create a more intimate behavioural environment.
A feed learns what keeps a person watching.
An AI companion may learn what keeps them talking.
A personalised advert predicts what someone may buy.
Generative AI can alter the wording for that person's fears, goals or preferences.
The experiment moves from selecting among available interventions towards creating interventions dynamically.
The number of possible versions becomes effectively unlimited.
That does not make harmful outcomes inevitable.
It increases the importance of deciding what the system is allowed to optimise.
We Are Beginning to Regulate the Experiment
For most of the internet's history, product experimentation was treated primarily as an internal company matter.
That assumption is changing.
The European Union's Digital Services Act requires very large platforms to assess systemic risks, create greater transparency around recommender systems and support access to platform data for vetted researchers studying those risks.
The designated platforms each reach more than 45 million monthly users in the European Union.
The UK's Online Safety regime increasingly requires risk assessment before significant product changes, particularly where children may be affected.
The FTC has called for stronger data minimisation, clearer control, improved safeguards for children and more consistent oversight of automated systems.
These measures do not stop experimentation.
Nor should they.
They attempt to change who bears the risk.
The old approach allowed the public to encounter the product and reveal the consequences through use.
The emerging approach requires companies to anticipate more of the consequence, document the risk and permit greater external scrutiny.
The direction is moving from:
Launch, observe and respond.
Towards:
Assess, protect, launch and continue monitoring.
Whether enforcement produces that outcome remains uncertain.
The legal principle is becoming clearer.
Scale creates obligations.
Regulation Faces a Difficult Balance
Excessive restrictions can prevent useful experimentation.
A small company may be unable to absorb the compliance costs carried easily by the largest platforms.
Requiring disclosure of every experiment could expose security systems, personal information and legitimate commercial knowledge.
Preventing rapid iteration might allow harmful defects to remain longer.
A product that cannot learn from users may become less safe, less accessible and less useful.
The objective cannot be to apply the full apparatus of medical research to every change in button placement.
The ethical weight should reflect the intervention.
Changing a font size is not equivalent to altering emotional content.
Testing page speed is not equivalent to testing political influence.
Recommending a new song is not equivalent to directing a vulnerable child towards content concerning self-harm.
The challenge is creating thresholds.
Scale.
Vulnerability.
Sensitivity of data.
Foreseeable harm.
Degree of manipulation.
Ability to withdraw.
Whether the intervention serves the user's stated goal or only the company's.
The greater the potential consequence, the stronger the required scrutiny should become.
We Still Do Not Know the Full Result
After decades of digital life, many central questions remain unresolved.
How much has personalised media changed human concentration?
How much of the rise in anxiety or distress among young people can be attributed to platform use rather than other social changes?
Which children benefit most?
Which face the greatest risk?
How much does recommendation shape preference rather than discover it?
What happens to identity when feedback becomes quantified?
How much workplace fragmentation is caused by technology and how much by management culture?
Does removing visible social approval improve wellbeing?
Do screen-time reminders change behaviour meaningfully?
Are non-profiled feeds better for agency?
Can contextual advertising fund services at similar scale?
How will persistent AI companions affect attachment, judgement and independence?
Honest answers require long-term research.
Company data.
Independent access.
Different methods.
Humility.
The absence of a complete answer does not mean nothing has happened.
It means the experiment is still running while society attempts to understand the results.
We Cannot Wait for Certainty About Everything
Scientific caution matters.
Weak associations should not be presented as proven causation.
One person's experience should not become a universal rule.
A legal complaint should not be reported as a verdict.
A company document should not be interpreted outside its methodology.
Moral panic produces bad policy.
But certainty can become another excuse.
If a company waits for undisputed proof that a feature harms every user, it may never act.
Digital effects are contextual.
The vulnerable subgroup may be relatively small.
The harm may be cumulative.
The counterfactual may be impossible to observe.
The product may provide benefits at the same time.
Responsibility must operate inside uncertainty.
A company can introduce a safer default while research continues.
Share data securely.
Reduce exposure for children.
Create meaningful stopping points.
Test the protective alternative as rigorously as the engaging one.
Measure regret as well as response.
Examine the minority harmed rather than allowing the average to hide them.
The experiment does not have to stop before it becomes more responsible.
The Participants Need More Than an Opt-Out
The standard response to concern is often user control.
Turn off notifications.
Reject personalised advertising.
Choose a chronological feed.
Set a time limit.
Reset recommendations.
Delete history.
These controls are valuable.
They return some power.
They do not create equality between the participant and the system.
The company understands the architecture.
The user sees settings.
The company knows which default produces more activity.
The user must recognise why the activity is occurring.
The system collects information through ordinary use.
The person must take special action to limit it.
The experiment continues by default.
Meaningful agency requires more than a technically available exit.
It requires understandable design.
Protective defaults.
Choices that do not punish the person for making them.
The ability to use essential functions without accepting every form of profiling.
A person cannot be described as freely choosing an environment they cannot realistically understand.
Society Is Both Researcher and Subject
Technology companies are not the only organisations learning from the experiment.
Parents are learning.
Schools are learning.
Governments are learning.
Children are learning about adults.
Adults are learning about themselves.
We introduced devices into family life because they were useful.
Then noticed how they changed behaviour.
We introduced workplace communication because it improved coordination.
Then discovered that flexibility could become permanent availability.
We introduced social metrics because they expressed response.
Then saw them become measures of worth.
We introduced recommendation because abundance required selection.
Then realised that selection could shape the preference it claimed merely to observe.
Society is not a passive victim.
It participated.
Adopted.
Normalised.
Rewarded.
Sometimes resisted.
The experiment is not being conducted by one group upon another.
Power is unequal, but participation is widespread.
That creates uncomfortable responsibility.
We cannot blame companies for every behaviour adults modelled, every boundary workplaces removed or every device parents handed over for immediate relief.
Neither can companies transfer responsibility entirely to users while professionally designing and measuring the environment surrounding those choices.
The outcome belongs to the interaction.
The Experiment Has Reached a Decision Point
For most of its history, the internet evolved faster than the ethical framework surrounding it.
Build first.
Grow.
Observe.
Fix obvious failures.
Continue.
That approach produced astonishing innovation.
It also allowed systems to become culturally essential before their wider consequences were understood.
We are no longer at the beginning.
We know attention is finite.
We know engagement can diverge from wellbeing.
We know children require stronger protection.
We know personalised systems can amplify vulnerability.
We know workplace connection can erase recovery.
We know AI will make behavioural adaptation more powerful.
We know companies possess evidence the public cannot see.
We know product architecture can be changed.
The experiment is no longer defined by ignorance.
It is defined by what society chooses to do with partial knowledge.
That is a different moral position.
Final Thought
We are living inside the largest behavioural experiment in human history.
Not because one company designed a secret study.
Not because every digital action is manipulation.
Not because technology controls humanity.
Because billions of people now inhabit environments capable of observing behaviour, changing in response and learning which version produces more of the outcome somebody chose to measure.
The scale is unprecedented.
The speed is unprecedented.
The intimacy is unprecedented.
The participant carries the instrument.
The instrument follows them through life.
The behaviour trains the system.
The system changes the behaviour.
Inside this process, companies run genuine experiments.
Some improve products.
Some increase revenue.
Some strengthen safety.
Some have altered political, emotional or commercial behaviour at a scale conventional researchers could never have reached.
The experiment has produced extraordinary benefits.
It has also exposed children, families, workers and societies to consequences nobody assessed completely before deployment.
There is no untouched control civilisation waiting to show us what would have happened otherwise.
We have only the world before us.
The evidence companies possess.
The research outsiders can conduct.
And the choices available now.
That may sound like a reason for despair.
It is not.
An experiment can be altered when its design proves inadequate.
Objectives can change.
Variables can change.
Safeguards can be introduced.
Participation can become more informed.
The people running it can become more accountable.
The people inside it can regain more agency.
The first nine essays have shown how the experiment began, how attention became valuable, how psychology and algorithms accelerated it, how children and adults became participants and how warning signs emerged.
The final question is no longer whether the experiment exists.
It is whether we are prepared to leave its future direction to the same incentives that created its first version.
Sources & Research Gaps
Principal Sources
International Telecommunication Union, Facts and Figures 2025
The ITU estimated that six billion people, approximately 74 per cent of the global population, used the internet in 2025.
The figure establishes the potential scale of digital behavioural environments while also showing that 2.2 billion people remained offline.
Meta, 2025 Annual Report and Results
Meta reported an average of 3.58 billion daily active people across Facebook, Instagram, Messenger and WhatsApp in December 2025.
The figure is based upon internal company estimates and user-account activity.
It does not measure time spent, depth of engagement or the purpose for which each service was used.
Meta Engineering, Building and Testing at Facebook
In 2012, Facebook stated that it ran hundreds of tests each day, generally releasing them to random samples to measure impact.
Some products requiring network effects could be tested across complete markets, including an entire country.
Bond and Colleagues, 61-Million-Person Political Mobilisation Experiment
The 2012 Nature paper describes randomised political messages delivered to 61 million Facebook users during the 2010 United States congressional elections.
The study reported effects upon online political expression, information seeking and validated voting, including social transmission through close relationships.
Kramer, Guillory and Hancock, Emotional Contagion Experiment
The researchers altered the amount of positive or negative material appearing in the feeds of 689,003 randomly selected Facebook users.
They reported small corresponding changes in the emotional language participants later used.
The study does not establish large changes in mood, clinical mental-health effects or control over individual emotional states.
PNAS Editorial Expression of Concern
The journal raised concern over the absence of explicit informed consent and the ethical questions created when product-level manipulation is later treated as published human-subject research.
The expression of concern did not retract the paper or invalidate its statistical findings.
United States Department of Health and Human Services, Common Rule and Informed Consent Guidance
The Common Rule establishes protections including institutional review and informed consent for human-subject research falling within its jurisdiction.
These requirements do not automatically govern every private product experiment.
They are used in this essay as an ethical contrast rather than as evidence that every platform test violates research law.
Federal Trade Commission, A Look Behind the Screens
The FTC's 2024 report examined data and automated-system practices among major social-media and video-streaming services.
It found extensive use of algorithms, data analytics and AI for recommendations, advertising, personalisation, engagement and monetisation.
It also reported limited user control, inconsistent governance and varying approaches to testing and monitoring.
Meta, Internal Research on Teen Wellbeing and Instagram
Meta published annotated internal research following public reporting and congressional scrutiny.
The company argued that its findings showed mixed effects, with benefits reported across many areas and a concerning result among a subgroup of teenage girls already experiencing body-image difficulties.
The research was exploratory and did not establish population-level causation.
United States Surgeon General, Social Media and Youth Mental Health
The Surgeon General called for greater transparency and independent access to platform data relevant to health and wellbeing.
The advisory also recognised that social media can provide connection, support, creativity and identity-related benefits.
Ofcom, Online Safety Regulation
Ofcom stated in March 2026 that algorithms were children's main pathway to online harm and that significant product changes must be risk-assessed before deployment.
It described the requirement as an end to product testing on children.
European Union Digital Services Act
The Digital Services Act creates systemic-risk duties for very large platforms and permits qualified researcher access to relevant platform data under defined conditions.
It also creates transparency and oversight requirements for recommender systems and other large-scale platform operations.
Research Gaps and Limitations
The title We Are Living Inside the Largest Behavioural Experiment in Human History is rhetorical and philosophical.
There is no authoritative global ranking that establishes modern digital life as the largest experiment according to one accepted scientific definition.
The wider system is not a controlled experiment.
It lacks:
- a single researcher;
- one hypothesis;
- one intervention;
- a universal control group;
- consistent informed consent;
- a defined end point;
- one agreed outcome measure.
The phrase describes a distributed social process containing many formal and informal experiments.
A/B testing is not inherently unethical.
It can improve safety, accessibility, performance, privacy and usefulness.
The ethical significance depends upon the intervention, risk, transparency, affected population and objective.
Product testing does not automatically fall within legal definitions of human-subject research.
Rules vary by jurisdiction, funding, institution, methodology and purpose.
The Common Rule should not be presented as applying universally to private commercial experimentation.
The 61-million-person Facebook voting experiment measured small average effects.
Its results should not be interpreted as evidence that Facebook determined an election outcome or controlled how individual users voted.
The emotional-contagion experiment found very small changes in written emotional expression.
It did not directly diagnose or measure clinical mood disorders, and it does not prove that feeds can control a person's emotional state.
Meta's 2012 statement about hundreds of daily tests is historical.
Current experimentation volume is likely to differ and is not fully disclosed publicly.
Daily active-user figures measure account activity rather than unique human identity perfectly.
Duplicate, shared or automated accounts may affect estimates despite company integrity systems.
The FTC report concerns companies and information supplied during a defined investigation period.
Its findings do not mean every company used identical data, testing practices or governance.
Internal company research can provide valuable evidence inaccessible elsewhere.
Its existence should not be interpreted automatically as proof that the company concealed harm or ignored its researchers.
Published safety announcements do not independently prove that interventions work.
Outcomes require continued external evaluation.
The social effects of digital products are difficult to isolate because platform use interacts with family, education, economic conditions, culture, health, personality and existing behaviour.
Most people use several services simultaneously.
The absence of an untouched control society makes definitive causal claims difficult.
Further independent research is needed into:
- the current volume and scope of experiments run by major platforms;
- which experiments require internal ethical review;
- how companies determine when a product test becomes sensitive behavioural research;
- whether users understand that they may receive experimental product versions;
- how children are excluded from high-risk testing;
- the long-term cumulative effects of repeated small interventions;
- whether short-term engagement tests predict long-term satisfaction or regret;
- how product experiments affect vulnerable groups hidden by average outcomes;
- which internal teams possess authority to stop or modify experiments;
- how commercial targets influence the selection of outcome measures;
- whether independent researchers can reproduce company safety claims;
- how generative AI changes the scale and personalisation of behavioural intervention;
- whether AI companions should be subject to stronger consent and safety standards;
- how workplace software experiments affect employees and organisational culture;
- how regulation can distinguish low-risk optimisation from consequential human intervention;
- whether meaningful consent can exist inside services necessary for education, work or social participation;
- which forms of transparency genuinely improve accountability without compromising privacy or security.