AI and the 2026 Elections: Brazil Faces Its Biggest Test Against Deepfakes and Synthetic Disinformation

Generative artificial intelligence has spent the last few years proving that it can create.

Texts.

Images.

Voices.

Videos.

Avatars.

Characters.

Now, Brazil is beginning to face a much harder question:

What happens when these technologies become fully embedded in the political process?

On September 1, 2026, just over a month before the first round of Brazil's elections, scheduled for October 4, the debate around artificial intelligence entered a new phase.

Brazilian electoral authorities are facing one of the most significant regulatory tests since generative AI became widely accessible.

The problem is not limited to so-called deepfakes.

The challenge is much broader.

We are entering a reality in which virtually anyone can produce synthetic political content at scale, create characters who never existed, simulate voices, generate videos, and use artificial intelligence systems to influence how information is produced and distributed.

The central question is no longer:

"Is it possible to create a convincing fake video?"

We already know the answer is yes.

The question now is:

"How does a democracy function when producing synthetic reality becomes cheap, fast, and accessible?"

2026 is not just another election with social media

Technology and elections are not a new combination.

Political campaigns have used the internet for decades.

Then came:

  • social media;
  • targeted advertising;
  • data analytics;
  • influencers;
  • automation;
  • instant messaging;
  • recommendation algorithms.

Generative artificial intelligence adds something different.

It drastically reduces the cost of producing content.

In the past, creating a sophisticated audiovisual campaign required:

  • a scriptwriter;
  • a designer;
  • an editor;
  • a voice actor;
  • a video production team;
  • equipment;
  • time.

Now, part of that process can be handled by software.

IDEA
  ↓
AI writes the script
  ↓
AI generates the image
  ↓
AI synthesizes the voice
  ↓
AI generates the video
  ↓
AI creates variations
  ↓
Content is distributed

This fundamentally changes the economics of digital communication.

Deepfakes are only part of the problem

When we talk about AI and politics, the best-known term is deepfake.

In simple terms, it refers to synthetic or manipulated content capable of representing people, situations, or events in an artificially realistic way.

One example would be a video that makes it appear as though a person said something they never actually said.

But focusing exclusively on deepfakes can cause us to overlook a much larger transformation.

AI can produce:

  • entirely artificial characters;
  • synthetic photographs;
  • fake audio;
  • videos;
  • comments;
  • articles;
  • memes;
  • automated responses;
  • profiles;
  • advertising variations;
  • hyper-personalized content.

In other words:

Manipulation does not necessarily need to impersonate a real person. It can manufacture an entire reality.

The case of political avatars

One of the regulatory difficulties lies precisely in characters generated by artificial intelligence.

Imagine a political profile called:

Maria

The character does not exist.

Her face was artificially created.

Her voice is synthetic.

Her videos are generated with artificial intelligence.

She comments on politics every day.

She supports certain ideas.

Criticizes others.

Gains followers.

Creates memes.

Participates in digital debates.

Now an interesting question arises:

Is that a deepfake?

Perhaps not in the traditional sense.

The character is not necessarily pretending to be a real person.

But it may still exert political influence.

This type of situation shows why regulating artificial intelligence is so difficult.

Technology frequently creates behaviors that do not fit neatly into earlier legal categories.

Brazil already has electoral rules for artificial intelligence

Brazil did not enter this scenario completely unprepared.

Since the 2024 municipal elections, the country's Electoral Justice system has had specific rules related to the use of artificial intelligence.

Among the main concerns are:

  • identification of synthetic content;
  • use of deepfakes;
  • manipulation of images and voices;
  • use of automated systems;
  • integrity of electoral information.

However, 2026 presents a different challenge.

The technology has evolved.

Access has expanded.

Quality has improved.

And the amount of content produced by ordinary users has increased.

Regulation is no longer dealing only with professional campaigns.

It now has to deal with an entire ecosystem.

Scale is the real problem

Imagine a traditional campaign producing:

10 videos per week

Now imagine an AI system producing:

10 videos per hour

Then imagine:

1,000 users
×
10 pieces of content
=
10,000 pieces of content

The issue is no longer simply identifying whether a particular piece of content is false.

It becomes a matter of dealing with volume.

This is one of the major effects of generative AI.

It does not merely improve the ability to create.

It multiplies production capacity.

The marginal cost of content is approaching zero

There is an important economic concept here.

Producing the first piece of a campaign may require effort.

But generating additional variations with artificial intelligence can cost very little.

Imagine a political advertisement.

In the past, a campaign might produce one version.

With AI, it can produce hundreds:

Original message
       ↓
       AI
       ↓
┌──────────────┬──────────────┬──────────────┐
│              │              │              │
Youth version  Formal version Popular version
│              │              │              │
Region A       Region B       Region C
│              │              │              │
Topic X        Topic Y        Topic Z

This creates an unprecedented ability to personalize communication.

From mass propaganda to individualized propaganda

Traditional political communication works approximately like this:

Candidate
    ↓
Message
    ↓
Millions of people

With data, algorithms, and AI, we can move toward:

Candidate
    ↓
System
    ↓
Voter profile
    ↓
Personalized message
    ↓
Specific voter

Two people may receive completely different versions of the same campaign.

One message emphasizes the economy.

Another focuses on security.

Another on religion.

Another on education.

Another on technology.

Personalization is not necessarily illegal or harmful.

Marketing has used segmentation for a long time.

The problem arises when extreme personalization reduces the public's ability to scrutinize political speech.

How do you scrutinize a promise that only you received?

This is an important problem.

In a televised debate, everyone hears the same statement.

It can be:

  • analyzed;
  • challenged;
  • compared;
  • archived.

Now imagine hyper-personalized political advertising.

Voter A → receives promise A

Voter B → receives promise B

Voter C → receives promise C

Voter D → receives promise D

None of them necessarily knows what the others received.

Political communication becomes less public.

This creates new challenges for:

  • transparency;
  • oversight;
  • journalism;
  • electoral authorities;
  • civil society.

Chatbots enter the debate

There is another frontier as well.

For a long time, people searched for political information using search engines.

Now millions of people ask artificial intelligence systems directly.

For example:

"Who is the best candidate for president?"

Or:

"Who should I vote for?"

Or:

"Which candidate is better for the economy?"

This creates a different problem.

A chatbot is not simply a library.

It formulates an answer.

The choice of which information is presented, how it is organized, and even the language used can influence the user's perception.

When AI recommends, is it informing or influencing?

This question will become increasingly important.

Consider two responses.

Response A

Here are the candidates' economic proposals.

Response B

Based on the available information,
candidate X appears to be the best option.

There is a significant difference.

In the first case, AI helps the user access information.

In the second, it begins to participate in the decision itself.

That is why artificial intelligence systems used during elections must handle questions involving political recommendations with particular care.

The problem of algorithmic authority

There is an important psychological phenomenon at work.

People may assign authority to technological systems.

If a chatbot responds confidently:

"This is the correct answer."

many users may interpret that statement as the result of a neutral and objective analysis.

But artificial intelligence models are not absolute arbiters of reality.

They can:

  • make mistakes;
  • misinterpret information;
  • rely on outdated data;
  • reproduce biases;
  • provide incomplete answers.

In politics, this becomes especially sensitive.

The appearance of neutrality can be more powerful than explicit propaganda

When we see an advertisement saying:

Vote for candidate X

we know we are looking at propaganda.

But when we ask a technological tool:

Which candidate has the best economic proposal?

and receive a direct answer, the perception may be different.

The user may believe they are receiving a purely technical analysis.

That is precisely why algorithmic transparency and governance become important.

Deepfakes can also target companies

It may seem as though this discussion matters only to politicians.

It does not.

The same technologies used to fabricate political content can also be used against companies.

Imagine a video that appears to show the CEO of an organization announcing:

"Our company has entered judicial restructuring."

The video is fake.

But it circulates for 40 minutes before being debunked.

During that period:

  • customers contact the company;
  • employees become concerned;
  • suppliers suspend operations;
  • investors react;
  • journalists begin asking questions.

The truth eventually emerges.

But the damage has already begun.

The time between a lie and its correction matters

There is an important characteristic of digital disinformation.

False information may spread faster than its correction.

We can represent it like this:

FALSE CONTENT
      ↓
Publication
      ↓
Shares
      ↓
Thousands of views
      ↓
Reaction
      ↓
────────────────────
      ↓
Verification
      ↓
Correction

By the time the correction arrives, part of the impact has already occurred.

With generative AI, producing false content becomes faster.

Organizations will therefore also need to improve their response speed.

Companies will need deepfake response protocols

Imagine that tomorrow an audio message appears from your company's CFO requesting a transfer.

The employee recognizes the voice.

The vocabulary sounds right.

The context seems plausible.

The question should not be:

"Does it really sound like them?"

The question should be:

"Was the authorization process followed?"

That change is fundamental.

Content can no longer function as authentication

An important business rule for the AI era is:

Content does not prove identity.

That means:

Voice ≠ identity

Video ≠ identity

Photo ≠ identity

Message ≠ identity

All of these elements can be synthesized.

Critical operations need additional verification mechanisms.

An example of a corporate protocol

Consider a payment request.

Instead of:

Director sends voice message
        ↓
Employee recognizes the voice
        ↓
Payment

use:

Request received
        ↓
Corporate system
        ↓
Authenticated identity
        ↓
Second approval
        ↓
Data verification
        ↓
Payment

The process no longer depends on a person's ability to recognize a voice.

The crisis of audiovisual evidence

For much of the twentieth century, photographs, audio recordings, and video carried enormous evidentiary weight.

"There is a recording."

That sentence carried authority.

Artificial intelligence changes that relationship.

It does not mean photographs and videos have lost all evidentiary value.

It means their authenticity needs to be examined more carefully.

We are entering an era in which asking:

"Does this look real?"

will no longer be enough.

We will need to ask:

"Can we verify its origin?"

Digital provenance becomes more important

One technological response to this problem is the creation of mechanisms capable of recording the origin and history of digital content.

We can imagine:

Camera
  ↓
Image created
  ↓
Digital signature
  ↓
Edit recorded
  ↓
Publication
  ↓
Verification

The goal is to allow platforms or users to verify information such as:

  • who created it;
  • when it was created;
  • what device produced it;
  • whether it was edited;
  • what changes were made.

This is known as content provenance.

Watermarks do not solve everything

Another commonly discussed solution is using watermarks in AI-generated content.

They can be:

  • visible;
  • invisible;
  • embedded in metadata;
  • detectable by software.

But there are limitations.

A watermark can be:

  • removed;
  • cropped;
  • altered;
  • lost during conversion;
  • eliminated through a screenshot.

Therefore:

There is no single technological solution to the problem of synthetic disinformation.

It will require a combination of:

  • technology;
  • regulation;
  • education;
  • processes;
  • accountability.

The paradox of AI detection

There is also a technical problem.

As generative models improve, detectors need to improve as well.

This creates a race:

GENERATOR
   ↓
More realistic content
   ↓
DETECTOR
   ↓
Detection improves
   ↓
GENERATOR
   ↓
New technique

It is similar to the historical relationship between:

  • spam and filters;
  • malware and antivirus software;
  • fraud and anti-fraud systems.

No defense remains perfect forever.

Brazilian regulation is being tested in real time

Brazil presents a particularly interesting scenario.

On one hand, there are specific electoral rules.

On the other, the country is still debating broader artificial intelligence legislation.

The main national regulatory proposal remains Bill 2,338/2023, approved by the Senate in 2024 and later sent to the Chamber of Deputies for analysis.

This means that different regulatory layers coexist:

Constitution
      ↓
Electoral legislation
      ↓
LGPD
      ↓
Brazilian Internet Civil Rights Framework
      ↓
TSE resolutions
      ↓
Future AI legislation

AI does not exist in a legal vacuum.

Even when specific legislation is still under debate, other rules continue to apply.

The international debate is also divided

While Brazil faces its own regulatory challenges, different countries are adopting different approaches.

The European Union has moved forward with a broader regulatory strategy.

The United States has, at different times, defended an approach more focused on innovation and less dependent on new regulatory bodies.

On September 1, 2026, U.S. representatives were preparing to advocate within the G20 for a lower-intervention approach to AI regulation, amid international technological competition and the growth of Chinese models.

There are, therefore, different philosophies.

Approach 1

More rules
↓
Reduce risks
↓
Build trust

Approach 2

Fewer barriers
↓
Accelerate innovation
↓
Increase competitiveness

The challenge is finding the right balance.

Overregulation also has a cost

Regulation is not automatically beneficial simply because it seeks to improve safety.

Excessively complex rules may benefit precisely the largest companies.

Why?

Because large organizations can maintain:

  • lawyers;
  • compliance teams;
  • regulatory departments;
  • specialists.

A small startup may not be able to do the same.

Poorly calibrated regulation can therefore create a barrier to entry.

Underregulation also has a cost

The opposite extreme also creates problems.

Without adequate rules, we may see:

  • fraud;
  • discrimination;
  • data abuse;
  • manipulation;
  • unsafe systems;
  • loss of trust.

And trust has economic value.

If people begin to believe that:

"Nothing online can be trusted."

the digital economy itself suffers.

The problem of digital trust

Think about how many activities depend on trust:

  • contracts;
  • payments;
  • identity;
  • documents;
  • communication;
  • commerce;
  • news;
  • reputation.

The internet works because mechanisms exist that establish some level of trust.

Synthetic artificial intelligence places pressure on those mechanisms.

The challenge of the next decade may not simply be creating more realistic content.

It may be creating better ways to answer:

"Where did this come from?"

Zero Trust may reach content itself

In cybersecurity there is a principle known as Zero Trust:

Never trust automatically. Always verify.

Traditionally, it applies to:

  • users;
  • devices;
  • networks;
  • systems.

But we may begin applying a similar logic to content.

Content received
       ↓
Do not trust automatically
       ↓
Verify origin
       ↓
Verify context
       ↓
Verify authenticity
       ↓
Make a decision

This will be particularly important for high-impact information.

Digital literacy will need to change

For years, we taught:

"Do not believe everything you read on the internet."

Now we will need to teach:

"Do not automatically believe everything you see or hear."

Because:

  • images may be synthetic;
  • videos may be synthetic;
  • voices may be synthetic;
  • documents may be synthetic.

Critical thinking is no longer just an academic skill.

It becomes a digital security competency.

How can manipulated content be identified?

No single signal is sufficient.

But some behaviors help.

1. Verify the source

Who published it?

Is there a primary source?

2. Look for independent confirmation

Have other reliable outlets or sources reported the same event?

3. Be cautious with emotional urgency

Manipulative content often tries to provoke:

  • fear;
  • anger;
  • outrage;
  • urgency.

4. Do not trust appearance alone

A visually perfect video may still be synthetic.

5. Check the context

Does the information make sense given known events?

Companies also need to educate employees

Corporate security training will need to include artificial intelligence.

Employees need to know that:

  • an audio recording can be fake;
  • a video meeting can be manipulated;
  • a document may have been generated;
  • a message may have been automatically personalized.

This is especially important for employees working in:

  • finance;
  • human resources;
  • management;
  • data;
  • customer service;
  • infrastructure.

The 2026 election will serve as a laboratory

Brazil has more than 150 million voters.

It has extremely high social media usage.

It has a strong culture of instant messaging.

And it is entering a national election during the mass adoption of generative artificial intelligence.

That makes 2026 an important test.

Not only for Brazil.

Other countries will observe:

  • which rules worked;
  • which failed;
  • how platforms reacted;
  • how voters responded;
  • which new forms of manipulation emerged.

Technology evolves faster than legal categories

This may be the greatest regulatory difficulty.

Law works with definitions.

Technology works with possibilities.

When a law defines:

deepfake

technology may create something that does not fit perfectly into that definition.

When it regulates:

chatbot

an agent appears.

When it regulates:

generated content

hybrid content appears, partially human and partially synthetic.

It is a constant race.

Perhaps we should regulate effects, not technologies

One possible approach is to focus regulation less on the specific tool and more on the resulting impact.

Instead of asking only:

"Was artificial intelligence used?"

ask:

"What effect did it produce?"

For example:

  • was there fraud?
  • was there deceptive manipulation?
  • was privacy violated?
  • was there discrimination?
  • was there harm?
  • was there an attempt to impersonate someone?

This approach may make certain rules more resilient to technological change.

There is a broader lesson for companies

Even organizations completely removed from politics should follow this debate.

Why?

Because the solutions created to address synthetic content during elections will likely influence the rest of the digital economy.

Technologies for:

  • identity;
  • authentication;
  • provenance;
  • signatures;
  • detection;
  • verification;

may later be used in:

  • banking;
  • contracts;
  • e-commerce;
  • corporate communication;
  • business documents.

Elections may function as a laboratory for digital trust technologies.

The future of the internet may depend less on detecting falsehood and more on proving authenticity

There is an important difference between:

"Can we prove this is false?"

and:

"Can we prove this is authentic?"

The first approach tries to detect manipulation.

The second creates a chain of trust.

The future may be closer to the second.

In a world where almost any content can be synthesized, verifiable content may become increasingly valuable.

Trust may become infrastructure

Today, we think of digital infrastructure as:

  • servers;
  • cloud;
  • networks;
  • databases;
  • APIs.

In the future, we may need to add:

trust infrastructure.

Systems capable of answering:

Who created it?
     ↓
When?
     ↓
Under which identity?
     ↓
Was it modified?
     ↓
By whom?
     ↓
Can it be verified?

This may become as important as encryption and authentication are today.

Conclusion

Brazil's 2026 elections are putting artificial intelligence through a test that goes far beyond politics.

We are testing our ability to continue distinguishing:

information from manipulation,

identity from representation,

content from evidence,

reality from simulation.

Deepfakes are only the most visible manifestation of this transformation.

The larger problem is that we are entering a digital economy in which convincing content costs less and less to produce.

That changes propaganda.

It changes fraud.

It changes communication.

It changes security.

It changes politics.

And it changes the very definition of trust.

During the first phase of the internet, the main question was:

"How do we put information online?"

Then it became:

"How do we find the correct information?"

In the age of artificial intelligence, a new question is beginning to dominate:

"How do we know that what we are seeing actually came from the person or organization claiming to have produced it?"

Answering that question will require much more than deepfake detectors.

It will require technology, education, regulation, digital identity, and new verification processes.

Artificial intelligence has made fabricating content extremely cheap.

Now we need to make proving authenticity extremely reliable.

That may become one of the most important infrastructures of the next phase of digital transformation.

References

  1. Reuters. Brazil races to rein in AI weeks before election. Published September 1, 2026.
  2. Brazilian Superior Electoral Court (TSE). Rules and resolutions concerning electoral advertising, artificial intelligence, synthetic content, and deepfakes.
  3. Bill No. 2,338/2023. Proposal for Brazil's regulatory framework for artificial intelligence.
  4. Tech Policy Press. Analysis of Brazil's 2026 elections as a test for artificial intelligence regulation.
  5. Reuters. US to urge hands-off AI regulation at G-20, official says. Published September 1, 2026.

Editorial notice: This article is informational and analytical in nature. It does not represent support for, opposition to, or an electoral recommendation concerning any candidate, political party, or political position.

Content produced for the Vórtice Blog — technology, artificial intelligence, security, regulation, automation, and digital transformation.