For most of the modern era, understanding public opinion required an intermediary.
A polling company selected a sample, called or interviewed perhaps one thousand people, applied statistical weighting and announced what millions of citizens supposedly thought.
That model made sense when individual citizens had few ways to publish their own opinions.
But the information structure of society is changing.
Today, millions of people continuously publish their thoughts through YouTube channels, Instagram accounts, X, Facebook, Reddit, blogs, online communities, newsletters, independent homepages, product reviews and comments.
Increasingly, even ordinary individuals operate their own small media channels.
In the AI era, this raises a provocative possibility:
What if society no longer needs to ask 1,000 people what 50 million people think?
What if AI can listen to millions of actual human voices instead?
From Sampling Public Opinion to Observing Public Opinion
Traditional polling is based on sampling because listening to everybody was impossible.
AI changes that economic limitation.
An AI system can potentially analyze millions of publicly available expressions in near real time.
It can distinguish topics, classify arguments, identify changes in sentiment, compare generations, detect emerging concerns and follow how opinions develop over weeks or years.
More importantly, it does not have to rely only on a single platform.
Imagine an AI public-opinion system analyzing independent sources such as:
- personal blogs and homepages;
- YouTube videos and comments;
- social-media posts;
- online communities;
- podcasts and newsletters;
- consumer reviews;
- public petitions;
- independent journalism;
- search behavior in aggregated form;
- and traditional surveys.
This would be fundamentally different from conventional polling.
The question would no longer be only:
“What answer did a selected respondent give to our question?”
It could become:
“What are millions of people voluntarily talking about, how strongly do they care, and how are those opinions changing?”
AI Could Reduce Dependence on Polling Companies
This transformation could democratize the measurement of public opinion.
Today, society still depends heavily on organizations that possess the money, databases, telephone systems, panels and statistical expertise required to conduct surveys.
That creates an unavoidable concentration of influence.
The polling company determines the questions.
It determines how the sample is constructed.
It decides how responses are categorized and weighted.
And journalists decide which number becomes the headline.
AI could potentially break part of this structure.
A university, independent journalist, nonprofit organization, small media company or even an individual researcher could eventually analyze enormous amounts of public information without operating a traditional polling infrastructure.
Instead of one institution publishing one number, many independent AI systems could generate competing interpretations from overlapping public data.
That could make public-opinion analysis more decentralized.
But only if the underlying voices are genuinely human.
And That Is Where the Problem Begins
The same technology capable of analyzing millions of human opinions can also generate millions of artificial opinions.
AI can write comments.
AI can create blogs.
AI can operate social-media accounts.
AI can generate videos, photographs and voices.
AI agents can potentially maintain digital identities for months or years.
A future system analyzing the internet might therefore encounter an extremely difficult question:
Which opinions actually came from human beings?
A traditional pollster at least attempts to establish that a respondent is a real individual belonging to a defined population.
An AI analyzing the open internet cannot automatically assume that every account, article, comment or video represents an independent human mind.
One human could operate one hundred AI accounts.
One organization could operate one million.
And eventually AI agents themselves could generate continuous political and social discussion without direct human participation.
Can AI Verify AI?
Ironically, AI may also become the best technology available to detect this manipulation.
AI can analyze behavioral patterns that are extremely difficult for humans to see.
It can compare writing styles, posting intervals, account relationships, repeated narratives, image metadata, abnormal coordination and the historical evolution of identities.
Future systems could calculate something resembling a human confidence score for information sources.
| Signal | Possible AI Verification |
|---|---|
| Identity continuity | Has this digital identity behaved consistently over many years? |
| Original content | Does the account produce unique experiences, photographs, arguments or observations? |
| Behavior patterns | Is activity statistically consistent with normal human behavior? |
| Network coordination | Are thousands of accounts repeating similar narratives simultaneously? |
| Source provenance | Can the origin and modification history of content be established? |
| Cross-platform history | Does the same person have a coherent long-term presence across independent systems? |
| Synthetic-media detection | Is an image, voice, video or text likely to have been machine-generated? |
In this sense, AI may become both the observer and the auditor of the future information ecosystem.
But that creates another problem.
Who audits the auditor?
The Most Dangerous System Is One That Both Measures and Influences Opinion
This may become one of the central political questions of the AI age.
Suppose an AI system analyzes what millions of people are saying.
So far, it is acting as an observer.
But suppose the same system also determines what those people see next.
It recommends videos.
It ranks search results.
It summarizes the news.
It writes replies.
It suggests what users should post.
It decides which arguments receive more distribution.
It becomes a personal adviser that explains politics, economics and society.
At that moment, AI is no longer merely measuring public opinion.
It is participating in the production of public opinion.
The Feedback Loop
The danger can be described as a feedback loop.
AI observes human opinion → AI predicts what attracts attention → AI recommends information → humans react to that information → AI measures the new reaction → AI recommends more information.
After thousands or millions of iterations, it becomes increasingly difficult to determine where spontaneous human opinion ends and algorithmically stimulated opinion begins.
This could become much more powerful than traditional propaganda.
Traditional propaganda broadcasts one message to millions of people.
AI can potentially communicate differently with every individual.
One person responds to fear.
Another responds to economic anxiety.
Another responds to national identity.
Another responds to status, humor, anger or compassion.
A sufficiently sophisticated system could learn which argument is most persuasive for each person.
Harari’s Warning: Democracy Is a Conversation Between Humans
Historian Yuval Noah Harari has increasingly focused on this problem.
In a 2026 interview, he described democracy in remarkably simple terms:
“Democracy is a conversation between people.”
His concern is that social-media algorithms were originally optimized not to discover truth, but to maximize engagement.
Those systems learned that anger, fear and hatred can be extremely effective at keeping human attention.
Harari argues that this helped damage the democratic conversation even before generative AI became powerful.
The next stage could be much more consequential.
AI does not merely select human-created information.
It can create language itself.
It can maintain a conversation.
It can learn about an individual.
It can remember previous interactions.
And it can potentially build trust or even emotional intimacy at enormous scale.
In his recent discussions about AI, Harari has emphasized that one of the critical resources at stake is trust.
For centuries, societies built institutions — governments, courts, universities, banks, newspapers, scientific organizations and polling companies — partly to create systems of trust between people who did not personally know one another.
The AI era may shift part of that trust from institutions to algorithms.
What Happens If We Trust the AI More Than Humans?
Imagine a future citizen asking:
“What does the country really think?”
Instead of checking several polls, the citizen asks an AI assistant.
The AI examines billions of pieces of information and responds:
“Most people support this position.”
That answer could be extraordinarily useful.
It could also be extraordinarily powerful.
Because most individuals would have no practical way to independently inspect the billions of data points behind the conclusion.
The AI's interpretation could gradually become more influential than any individual polling company.
We might therefore solve the problem of dependence on a few polling companies only to create a much larger problem:
dependence on a few AI systems.
An AI Does Not Need Political Beliefs to Manipulate Society
There is an important distinction here.
AI does not need consciousness, ideology or a secret political ambition for this danger to emerge.
The problem can arise simply from optimization.
If an AI system is instructed to maximize engagement, it may create more engaging information.
If it is instructed to maximize user retention, it may learn how to keep people emotionally dependent on it.
If it is instructed to reduce social instability, it may gradually suppress information associated with conflict.
If it is instructed to maximize economic efficiency, it may favor ideas that make society easier to predict and manage.
None of this requires the AI to “want” anything in the human sense.
Yet the resulting society could gradually evolve toward conditions that are easier for the optimization system to manage.
This is where the phrase “a world convenient for AI” becomes worth considering.
Not because a machine necessarily dreams of such a world.
But because optimization systems naturally tend to reshape environments around measurable objectives.
Public Opinion Could Become Self-Fulfilling
There is an even subtler danger.
Humans are influenced by what they believe other humans believe.
If people repeatedly hear that “the majority thinks X,” some will reconsider their own position.
This means that measuring public opinion is never completely separated from influencing it.
With traditional polls, this effect is relatively limited.
With an AI system constantly describing the supposed consensus of society to billions of users, the effect could be much larger.
The AI could say:
“Most people are becoming pessimistic.”
People become more pessimistic.
The AI measures increased pessimism.
It then reports stronger pessimism.
The prediction helps create the reality it predicted.
This is an algorithmic version of a self-fulfilling prophecy.
The Future Public-Opinion System Must Separate Three Functions
If AI becomes an important tool for understanding society, DATAAD believes three functions should remain clearly separated.
| Function | Role |
|---|---|
| Human expression | Real people create opinions, experiences and arguments. |
| AI measurement | AI aggregates and analyzes those human signals. |
| Independent verification | Separate systems test authenticity, methodology, bias and manipulation. |
The AI that recommends information should ideally not be the only AI measuring whether people agree with that information.
The company operating the social network should not be the only organization auditing the network.
And no single AI model should become the final authority on what society believes.
We May Need Competing AIs, Not One Truth Machine
The safest architecture may look less like one omniscient supercomputer and more like science.
Different AI systems could analyze the same public data independently.
One system might estimate support at 52%.
Another might estimate 47%.
A third could explain why their conclusions differ.
Independent auditors could examine data provenance and synthetic activity.
Humans could inspect the methodology.
Instead of asking an AI for “the truth,” society would use AI to expose competing interpretations and uncertainty.
This is less convenient.
But democracy itself is inconvenient.
It requires disagreement, negotiation, uncertainty and time.
The Great Opportunity
There is a positive side to this transformation.
AI could allow voices that traditional polling and mass media often fail to notice to become visible.
A small-business owner writing a blog.
An engineer operating a technical YouTube channel.
A student maintaining an independent homepage.
A retired worker discussing local problems online.
A neighborhood community debating a development project.
Millions of these small independent voices could collectively become a much richer representation of society than one survey conducted on one evening.
AI could help us listen to them.
That would be a remarkable democratization of information.
And the Great Danger
But the system becomes dangerous when we can no longer determine whether those voices are human, whether their visibility was organic, or whether their opinions were subtly shaped by the same intelligence that is analyzing them.
The biggest risk of AI and public opinion may therefore not be fake polling.
It may be something much deeper.
AI could eventually become simultaneously the listener, the journalist, the adviser, the recommender, the fact-checker, the pollster and the participant in society’s conversation.
Once all of those roles are concentrated in the same information system, manipulation no longer requires falsifying a statistic.
The system can influence the reality from which the statistic is created.
The DATAAD View
Traditional polling companies may become less important in the AI era.
That is not necessarily a bad development.
AI offers the possibility of understanding society from millions of independent human voices rather than depending entirely on small statistical samples.
But replacing centralized polling organizations with centralized AI systems would not solve the fundamental problem.
It would merely move the center of power.
The real objective should therefore be neither “trust the polling company” nor “trust the AI.”
It should be:
Build an information system in which humans, institutions and competing AIs continuously verify one another.
Perhaps the most important political question of the next decade will not be whether AI can accurately understand human opinion.
Technically, it probably will become extremely good at that.
The more difficult question is this:
Can AI understand human opinion without quietly becoming the force that creates it?
Humanity should answer that question while humans are still clearly the ones having the conversation.
DATAAD Editorial Perspective
This article does not argue that AI systems possess political intentions or independent social goals. The warning concerns feedback, optimization and concentration of informational power. A system can influence human behavior without consciousness if its objectives, recommendation mechanisms and incentives systematically favor particular outcomes.
Reference framework: Yuval Noah Harari interviews and discussions on AI, democracy, trust and information systems, including his 2026 conversations with The Economist and EL PAÍS; his broader arguments in Nexus; World Economic Forum research on AI, synthetic media and information manipulation; and contemporary research on algorithmic influence and digital trust.