After a long period of hidden development, deepfake technology has made a qualitative leap that surprised even experts. Artificial intelligence has reached such a level of realism in generating faces, voices, and movements that it can now easily deceive people without special training or even automated verification systems.
By 2026, a radical change is expected: the end of the era of pre-recorded clips and the beginning of the era of real-time synthesis. This means that fake content will be able to react and interact with a person during live communication.
Previously, a deepfake was a static video that could be analyzed later. However, new generation models ensure temporal consistency: movements look logical from frame to frame, facial features do not flicker or distort around the eyes or jaw, and the personality remains stable even when the 'performance' changes. This allows for the free combination of movement characteristics and identity, for example, applying one gesture to different faces or giving multiple expressions to one fake personality.
The practical consequence of this is that the deepfake ceases to be just a file and begins to function as a 'digital actor' capable of responding in real time to a video call or conversation. This significantly complicates manual verification, as many visual signs of manipulation, such as flickering, artifacts, or lighting inconsistencies, disappear.
Similar to the development of video, voice cloning has also overcome a barrier. Only a few seconds of audio sample are needed to create a convincing voice with natural intonation, rhythm, pauses, and even breathing, which can mislead the human ear in most everyday situations.
This progress already has direct consequences: large retailers report receiving over 1000 fraudulent calls generated by AI daily. Such a volume was unimaginable recently and demonstrates how synthetic voice fraud has turned from a theoretical risk into an industrial operation.
The rapid spread of this phenomenon is explained by three factors. Firstly, technical realism: the most modern video models are designed specifically to maintain consistency between frames, eliminating visual errors that previously served as forensic evidence. Secondly, indistinguishable voice: voice cloning has exceeded what researchers call the 'threshold of indistinguishability,' making it unreliable to rely solely on hearing. Thirdly, almost zero technical barrier: the availability of generation tools and the automation of the process using AI agents allow almost anyone to create coherent and large-scale fake content without needing advanced technical knowledge.
According to security sector estimates, the growth rate is impressive: the number of detected deepfakes is projected to grow from approximately 500 thousand in 2023 to nearly eight million in 2025, corresponding to an annual increase of about 900%.
In light of this situation, the traditional defense method—analyzing pixels for visual errors—is losing its effectiveness. Researchers point out that current defense is shifting from human judgment to infrastructure protection, including:
For businesses, the immediate risk is fraudulent attacks via voice or video, ranging from 'digital kidnapping' to a fictitious executive authorizing fund transfers. For public discourse, the threat lies in real-time disinformation: simulations of interviews, bulletins, or polls that appear authentic enough to spread and influence before any verification can take place.
The central point for 2026 remains that the line between 'real' and 'synthetic' is no longer a matter of technical quality but a matter of trust in infrastructure—here, the source signature, authentication, and digital literacy will carry more weight than a person's ability to 'feel' that something is fake.
Artificial intelligence (AI) that executes commands without understanding the implications must be supervised, not revered. Generative AI is already involved in tasks such as writing, coding, research, teaching, and executive decision-making. Those who see it merely as a shortcut to increase productivity are missing crucial elements, initiating a dispute over how to integrate it.
It is essential to transform AI into a 'cointelligence,' approaching it with method and responsibility, rather than allowing the machine's fluidity to replace human discernment. In 2022, OpenAI launched ChatGPT, a conversational system capable of answering, recognizing errors, and questioning incorrect premises. Global research indicates that 88% of companies regularly use AI in some business function, and 62% have already tested AI agents.
Ethan Mollick, a professor at the University of Pennsylvania and author, studies the effects of AI on entrepreneurship, education, and work, coining the term 'cointelligence.' This expression rejects both the idea that AI is an infallible oracle and the view that it is merely a sophisticated calculator. The machine can function as a colleague, mentor, or teacher, provided it is guided by someone with knowledge. Although it writes accurately, summarizes quickly, simulates scenarios, and errs with great eloquence, the inattentive leader confuses clarity with truth; the prepared leader, on the other hand, expands options and tests assumptions.
Economists such as Ajay Agrawal, Joshua Gans, and Avi Goldfarb from the University of Toronto argue that AI reduces the cost of prediction, but judgment remains an essential human component in decision-making. In IT, AI agents can assist in classifying tickets, suggesting fixes, creating automation routines, and supporting incident analysis. However, defining acceptable risk, operational priority, customer impact, regulatory exposure, and cost of failure still belongs to a human being. AI accelerates the path, but leadership defines the destination.
Infrastructure also requires maturity. A generative AI without reliable data, adequate security, observability, network, storage, and prepared servers results in a mere technological spectacle. The model may seem advanced, but the operation fails during the first peak of use, in cases of data leaks, or when the response lacks traceability. Data centers, accelerators, access policies, curated databases, and continuous evaluations constitute the least visible part of cointelligence, distinguishing real demonstration from practical outcome.
The Stanford University AI Index Report 2026 reports that generative AI reached 53% adoption in three years, surpassing the pace of personal computers and the internet, and that 70% of organizations use it in at least one business function. Although the speed is remarkable, true value emerges when processes are reformed. Simply copying answers into old documents yields little return; redesigning workflows, validation criteria, and professional roles brings greater benefit.
The labor market already reflects this pressure. The Global AI Jobs Barometer 2026 analyzed over one billion jobs across six continents and concluded that junior positions more exposed to AI are seven times more likely to require leadership and strategic reasoning. This is a critical point because early careers lose some of the repetitive training that forged judgment. Therefore, companies and educational institutions need to equip people to make decisions proactively, review better, and challenge systems with justification.
There is also the issue of trust. A Pew Research Center 2026 survey showed that 49% of adults in the United States use AI chatbots, compared to 33% in 2024, while 63% express concern about the speed of AI advancement. Stuart Russell, a computer scientist and professor at the University of California, Berkeley, identifies this dilemma as a control problem. Mustafa Suleyman, co-founder of DeepMind, argues that AI progress depends on the parallel evolution of governance structures responsible for supervising it. Kate Crawford, a researcher renowned for her studies on the social, political, and economic impacts of AI, reminds us that AI entails material, labor, and political costs, making AI governance a matter of design, management, and culture.
The regulatory timeline reinforces this message. In the European Union, obligations for general-purpose AI model providers came into force on August 2, 2025, with the Commission gaining oversight powers starting August 2, 2026. In Brazil, a bill presented in 2025 proposes the National System for Development, Regulation, and Governance of Artificial Intelligence (SIA), assigning a role to the National Data Protection Authority (ANPD).
The competitive advantage of the next decade will lie in the ability to interact with machines without yielding them final authority. This requires training, architecture, processes, ethics, and executive courage, demanding that AI be viewed as a positive force for society and businesses, with the same seriousness given to its risks. The leader who delegates judgment to the machine trades authority for mere convenience; the one who transforms AI into cointelligence expands the organization and empowers individuals. Between the assistant and the oracle, the responsible choice is clear: AI must think with us, and the decision must always have human backing.
A recent study published in the journal PNAS demonstrated that geometric intuition and the ability to recognize abstract shapes are not exclusive to the human species, but rather a cognitive trait shared with other primates. Researchers found that both monkeys and humans use analogous mental processes to classify and process geometric figures.
To map how each group perceives spatial patterns, scientists examined the behavior of monkeys, preschool children, and adults in visual matching tests. The collected data indicated that the brains of non-human primates can integrate complex visual representations with abstract symbolic properties.
Previously, the theory of symbolic singularity maintained that only humans possessed the aptitude to encode geometric relationships through specific rules, such as identifying right angles, parallel sides, or symmetry. The assumption was that animals were limited to using superficial and continuous perceptual characteristics to differentiate figures in the environment.
To test this theoretical distinction, scientists applied visual matching tasks to participants, requiring them to identify geometric figures even when presented in different orientations. By analyzing the choices using statistical and computational models, the team noted that monkeys favored abstract structural concepts, rather than just simple perceptual details.
The crucial point of the experiment was observed when analyzing the animals' response to rotated figures. When exposed to objects rotated in space, monkeys showed a dependence on structural and symbolic cues even more pronounced than preschool children when viewing static images. The results suggest that geometric representation derives from the visual system's capacity to process rotational invariance, establishing a neurological basis for the development of geometric intuition in primates. This discovery refutes the notion of an exclusive evolutionary leap in the human mind, indicating that basic geometric concepts have shared biological origins and that spatial abstraction is a gradual process among species.
