The intelligence behind the intelligence
Artificial intelligence can now generate convincing conversations, extraordinary images, computer code and increasingly realistic simulations of the physical world. Yet the more capable these systems become, the more they depend on something they cannot simply invent: reliable information about reality.
A conversation between humans. A surgeon performing a procedure. An engineer repairing a machine. A linguistic nuance missing from the world’s dominant languages. The millions of small decisions that allow people to navigate complex environments.
And behind every piece of information lies another question: who owns it, and who has the right to use it?
These questions have defined much of Daniela Braga’s professional life.
When Building Creative Machines met the Portuguese-born founder and CEO of Defined.ai at Web Summit in Lisbon on 10 November 2026, she had just presented an expanded data marketplace containing nearly 2,000 datasets and approximately ten petabytes of information.
The company says its marketplace now generates 70 per cent of its revenue and supplies major frontier AI laboratories. Its latest initiatives aim to make proprietary information more accessible to model developers while creating new opportunities for organisations and content creators to monetise their data.
But our conversation quickly moved beyond the announcement.
Braga, a linguist and speech-technology specialist who has spent more than two decades working in artificial intelligence, spoke about the predictions she got right, the ones she got wrong, the changing balance of power in AI, and the future of human work.
One particularly revealing exchange concerned synthetic data.
Defined.ai does not offer synthetic datasets in its marketplace. Not because Braga rejects them, but because, as she explained, its customers are sophisticated enough to generate synthetic information themselves.
What they need from Defined.ai is something different.
This apparently simple distinction illuminates one of the most important emerging markets in artificial intelligence: the supply of original, traceable, legally usable human data.
Daniela Braga | Press Conference @ Web Summit 2026. Credits: Building Creative Machines
From ALPHA to the global AI industry
The story begins a decade earlier.
In 2016, when Web Summit first arrived in Lisbon, DefinedCrowd — the original name of Defined.ai — was a young company founded in Seattle, with research and development activities in Portugal.
Braga had built her career in computational linguistics, speech synthesis and language technologies, including experience at Microsoft and VoiceBox Technologies.
Her conviction was that computers would eventually interact with humans through natural language.
At the time, that remained a difficult engineering ambition. Digital assistants were improving, but speech recognition, multilingual processing and the collection of reliable training information presented substantial obstacles.
DefinedCrowd was established in 2015 to address some of those obstacles.
It began by providing the human-generated data needed to train machine-learning systems, particularly in speech and language.
Its journey through Web Summit reflected its progress: from the early-stage ALPHA programme in 2016 to BETA in 2017 and the Growth Summit in 2018.
Investment followed, including a $50.5 million Series B in 2020. The business subsequently became Defined.ai, expanding beyond its original speech-data foundations into multimodal datasets, annotation, evaluation and data licensing.
Today, the company reports having raised more than $85 million.
But the most revealing aspect of this trajectory emerged when Braga reflected on her original technological predictions.
She was right about natural language becoming a fundamental interface between humans and computers.
She was wrong about how quickly keyboards and mice would disappear. And she had not anticipated how rapidly generative AI would become mainstream.
After decades in a specialised business-to-business industry, the speed of mass adoption genuinely surprised her.
Her experience illustrates a recurring paradox of technological forecasting: we may correctly identify the direction of change while fundamentally misunderstanding its timing.
The marketplace: three businesses in one
At Web Summit 2026, Defined.ai presented an expanded version of a marketplace concept it first introduced in 2022.
The platform now contains approximately ten petabytes of data across nearly 2,000 datasets. The catalogue includes speech, multilingual text, images, video and multimodal information, alongside specialised collections for healthcare, finance, manufacturing, robotics and other applications.
An AI-powered assistant allows buyers to search for relevant datasets through natural language, while downloadable samples enable technical evaluation before purchase.
According to figures presented at the launch, 280 companies had explored the marketplace during its first two weeks, with 344 new registrations and 333 sample downloads.
The expansion has three interconnected components.
1. A marketplace for frontier AI developers
The first is the catalogue itself, giving model developers access to information acquired through private sources and licensing arrangements rather than unrestricted internet scraping.
Braga says Defined.ai serves the leading frontier laboratories, although individual client relationships are not fully disclosed.
These customers increasingly require more than the general-purpose text that helped build the first generation of large language models.
Healthcare systems need information about patient journeys, clinical conversations and medical procedures. Robotics developers need real-world demonstrations. Industrial models need records of manufacturing processes, maintenance and failures.
Such information is often difficult to obtain because it exists inside organisations rather than on the public internet.
The challenge is therefore not simply collecting more information. It is gaining access to information that is both useful and legitimately available.
2. Strategic Access Program: flexibility instead of buying everything
The second component is the Strategic Access Program, designed to make the acquisition and evaluation of training data more flexible.
It introduces research and commercial licensing arrangements. According to the company’s announcement, research licences are priced at 20 per cent of the corresponding commercial licence, allowing developers to test whether particular datasets actually improve their models before making a larger commercial commitment.
In our interview, Braga clarified another important detail.
The programme operates through different tiers, using an allocation-based mechanism comparable to tokens. Customers choose which datasets they need within the allowance associated with their tier.
It is not an unrestricted subscription giving every buyer simultaneous access to everything.
This matters commercially.
Different frontier laboratories can acquire the same underlying information and use it in different ways, combining it with their own architectures, training procedures and proprietary knowledge.
The competitive advantage does not necessarily come from exclusive possession of an entire dataset. It can come from how effectively a laboratory selects and uses the information.
3. Partner Acceleration Program: turning knowledge into an asset
The third component is the Partner Acceleration Program.
It targets content creators, SMEs, publishers and larger organisations that possess potentially valuable proprietary information but may lack the technical or commercial capacity to sell it to AI developers.
Defined.ai offers to help evaluate, prepare, license and commercialise that information.
For an SME, the proposition is significant.
A manufacturer might possess thousands of equipment maintenance records. A specialist publisher might own decades of structured expertise. A healthcare organisation may hold valuable clinical information, subject to strict privacy and licensing conditions.
Such assets could acquire new economic value if properly organised and legally made available for training AI systems.
The company reported in January that several marketplace partners were generating more than $1 million annually from licensing their datasets. That figure is company-reported, and it should not be interpreted as a typical outcome for partners.
Nevertheless, the underlying ambition is compelling: moving some of the economic value of AI training towards the organisations and people who originally generate knowledge.
The real business is what happens between supply and demand
One of the most interesting moments of our conversation came when Braga described what Defined.ai actually does between a data supplier and a buyer.
It is tempting to imagine the marketplace as a digital warehouse: suppliers upload datasets, customers search, select and purchase.
Reality is substantially more complicated.
Information rarely arrives in the form a model developer needs.
A supplier’s archive may be fragmented, incomplete, poorly annotated or structured for purposes entirely unrelated to machine learning.
And buyers rarely want a dataset exactly as it was originally produced.
Braga used an unexpectedly effective analogy: it is not like opening Netflix and watching an entire film. Customers might want a minute from one film, another minute from a different one, and a carefully selected fragment from a third.
They want to slice and dice.
This reveals three activities at the heart of Defined.ai’s business.
First, the company connects demand with supply, understanding what AI developers need and identifying potential suppliers of relevant material.
Second, it works on the supplier’s data: cleaning, restructuring, organising, annotating and preparing it for particular applications.
Third, it adapts and selects information for the buyer’s requirements, sometimes helping prepare highly specific training collections for smaller or specialised models.
That last capability is becoming increasingly important.
As Braga emphasised, smaller models do not necessarily need enormous quantities of generic information. They often benefit from carefully chosen, high-quality examples.
The implications extend beyond the marketplace.
If model development becomes increasingly dependent on the selection and preparation of specialist knowledge, then the data supplier becomes part of the intelligence-development process itself.
The economic opportunity lies not simply in selling data, but in transforming information into something a machine can learn from.
Why organic data matters in a synthetic world
The distinction between original and synthetic data deserves particular attention.
Synthetic information is becoming an important resource for AI development. Language models can produce training examples, reasoning exercises and simulated conversations. Robotics systems can learn in virtual environments. Synthetic datasets can expand rare scenarios or supplement expensive real-world collection.
There is substantial evidence that carefully designed synthetic data can improve machine-learning performance.
But synthetic generation also introduces risks.
Research published in Nature in July 2024 demonstrated how indiscriminately training successive generations of models on model-generated outputs can produce model collapse, progressively weakening their representation of the original data distribution.
The danger is particularly pronounced for uncommon but important cases.
Later research has clarified that collapse is not inevitable. Synthetic and original information can be combined successfully, provided the training process is managed appropriately.
The important distinction is therefore not that synthetic data is bad and human data is good.
It is that synthetic information and original observations perform different functions.
Synthetic data can expand, reorganise or simulate knowledge. Original data provides an independent connection to the world.
Braga’s explanation for excluding synthetic datasets from Defined.ai’s marketplace is also commercially revealing: her customers can already generate them.
What is harder for those customers to reproduce is an authentic recording of an industrial task, a specialist conversation, or a properly licensed collection of human expertise.
And as AI-generated content becomes increasingly common online, identifying the provenance of information becomes more difficult.
The premium may shift towards data whose origin, rights and technical characteristics can be established.
In a world capable of producing almost unlimited artificial content, verified reality may become one of the most valuable inputs.
From language to machines that understand the physical world
Defined.ai’s January 2026 expansion into physical AI illustrates where this demand is heading.
The company announced multimodal datasets incorporating video, 3D point clouds, motion capture, LiDAR, force and torque measurements, and human demonstrations.
These resources support the development of robotics and other systems that must operate in physical environments.
At Web Summit, Braga described the shift in customer demand.
The industry is moving beyond relatively generic videos of people performing simple activities towards detailed recordings of complex real-world operations.
Repairing cars. Working on electrical systems. Carrying out industrial tasks. Performing medical procedures.
These activities require more than visual recognition.
A robot needs to understand movement, position, resistance, interaction and the relationship between an action and its consequences.
This connects directly with another observation from the interview: the emergence of smaller, customised models.
Rather than assuming that every application requires a larger general-purpose foundation model, an increasing number of organisations may find value in specialist systems trained on carefully selected information.
Braga’s argument was direct: for small models, a relatively limited quantity of excellent data can be more important than sheer volume.
That does not eliminate the importance of foundation models. It changes the opportunities available to businesses that cannot afford to build them.
Success, setbacks and the global competition
Defined.ai’s progress has not followed an uninterrupted upward trajectory.
The company raised substantial capital and expanded internationally, but also experienced difficult periods, including workforce reductions in 2021.
In January 2026, it reported 65 per cent year-on-year revenue growth for 2025, net revenue retention of 143 per cent and a 1,200 per cent increase in third-party partner datasets.
These figures suggested strong demand for its marketplace strategy, although the company did not disclose enough underlying financial information to independently evaluate profitability.
Then, in July 2026, Portuguese publication ECO reported a global restructuring involving approximately 30 per cent of the workforce, particularly affecting European operations.
Defined.ai confirmed the reduction and attributed it to a broader review of its operating structure and strategic priorities, rather than solely to sales performance.
The contrast between strong reported growth and significant restructuring is an important part of the story.
It reflects the volatility of the AI supply chain, where customer requirements, training methods and commercial priorities can change extremely quickly.
At the same time, Defined.ai faces increasingly sophisticated global competition.
Scale AI provides data, evaluation and expert services to major model developers. Other specialist businesses focus on licensed content, domain-specific information or human expertise.
In January 2026, Cloudflare acquired Human Native, a British marketplace connecting content creators and AI developers.
The acquisition demonstrated that large infrastructure companies also recognise the strategic importance of discovering, organising and commercialising original content.
Defined.ai’s potential differentiation comes from combining several capabilities: a sizeable catalogue, established supplier relationships, experience in multilingual and multimodal collection, data preparation, licensing and technical adaptation for individual customers.
Its collaboration with Getty Images, announced in January 2025, is one example of the licensed-content strategy.
But size alone is not a sufficient competitive advantage.
The company must demonstrate that its datasets improve model performance, that its licensing arrangements are trustworthy and that both suppliers and customers receive enough value to remain within the marketplace.
The decisive metric may ultimately be neither petabytes nor dataset counts.
It may be how much useful intelligence a customer can create from each properly licensed unit of information.
Daniela Braga and Gonçalo Perdigão during Web Summit 2026. Credits: Vanessa Amaral
An exclusive conversation with Daniela Braga
Building Creative Machines: Why does Defined.ai not have synthetic data in its marketplace?
Daniela Braga: We don’t have synthetic data in our marketplace simply because there is no demand for it. We would have it if there were a market. The market doesn’t look for it because it produces it itself. Our customers are sophisticated enough to generate their own synthetic data.
BCM: The marketplace operates between two parties: somebody looking for data and somebody supplying it. Are you essentially an intermediary?
Braga: Exactly, but we also add value to the data. The data that comes from suppliers has to be transformed. And on the customer side, they never want just the whole thing.
It’s not like going to Netflix to watch a film. They want a minute from one film, another minute from another, and another minute from another. It’s slice and dice.
BCM: Looking back over more than two decades in artificial intelligence, what surprised you most? What would you tell the people now creating companies in this new AI wave?
Braga: I made some predictions that were wrong, but I also made predictions that were right.
For example, when I started my career, I always believed that one day we would interact with machines through voice and natural language. That is the foundation of my academic background, and that is why I stayed on this path while everybody else was giving up along the way.
I knew that one day our interaction with machines would happen through natural language, through the language we speak.
And here we are.
My doctorate was in that field, and Defined.ai started very much from that background and from the gaps I saw in the market: the absence of a platform that could respond to the world’s languages.
And we need a less purist approach to language. I’m a linguist by training. If we all spoke Portuguese, or any other language, perfectly correctly, we would still be speaking Latin.
BCM: Do you believe keyboards and mice no longer belong on people’s desks?
Braga: Completely. But I thought keyboards and mice would disappear much earlier.
That proves that, despite more than twenty years of work in this field, the problem of speech recognition is still not solved.
Particularly code-switching, in a globalised world where people speak a mixture of English and their local language, and sometimes another language as well.
Certain geographies, such as the United States, are particularly prone to that. It’s still a problem.
BCM: If data is at the centre of everything, who is actually in control of artificial intelligence?
Braga: The large American technology companies are in control.
They consume data of every kind, ethical and unethical. There is still an enormous appetite for data.
But I also made a prediction that turned out to be wrong. I thought we would see consolidation among model developers, as we had seen with cloud providers and search engines.
That hasn’t happened in the same way.
And increasingly, fortunately, open source makes it possible to build small models and customised models.
With small models, the data is fundamental. What matters is a small amount of very good, high-quality data.
BCM: How can Portugal participate meaningfully in the AI revolution, rather than simply adopt technology developed elsewhere?
Braga: There are three pillars where we can make a difference.
There is the data pillar, where Defined.ai operates, although Defined.ai alone is not enough. There is the model pillar, which requires a great deal of capital. And there is the data-centre pillar.
Data centres are closely connected with energy, and Portugal has been a pioneer in renewable energy.
We also have a strategic geographical position, historical relationships with the United States, submarine cables and our role as an Atlantic gateway to Europe.
Portugal has a strong diplomatic tradition. We speak several languages and try to understand other people.
I think we have very special conditions to build on the data-centre pillar.
BCM: As AI transforms both white-collar and blue-collar work, where will the value of human labour remain?
Braga: Entry-level white-collar jobs are already being affected. This was the first year when I really saw the initial impact on the labour market.
Young people leave university and struggle to find entry-level work, whether in computer science, analysis, accounting or other areas.
But new jobs are also emerging: agent architects, data curators.
And then there is everything connected with the more human side of interaction: the emotional side, critical thinking and entrepreneurship.
An LLM can give you plenty of entrepreneurial ideas. Executing them is an entirely different problem. Managing people is another problem altogether.
I think human interaction will become extremely valuable.
We are talking about ageing populations, and there are opportunities for people to move towards work that addresses those needs.
Healthcare will advance very quickly. Education too.
Perhaps this is a utopian vision, but I believe we will see more opportunities.
In every technological revolution, there is a period of stress, and then humanity reorganises itself.
Europe, human work and the next decade
Three broader ideas emerge from the conversation.
The first concerns technological power.
Braga recognises the dominance of large American technology companies, but does not believe that building ever-larger foundation models is the only path to relevance.
Open-source technologies, specialist models and proprietary knowledge create opportunities for organisations operating outside the frontier-model race.
The second concerns Europe’s role.
When asked how Europe can exercise influence without leading the construction of the largest models, Braga emphasised historical memory, institutional credibility and the ability to impose limits on technologies when society considers them necessary.
It is a perspective that places Europe’s value not only in its computing capacity, but also in its experience of building institutions and governing technological change.
The third concerns work.
Braga sees early signs of displacement in entry-level professional roles, but also expects the emergence of new occupations and growing value in human interaction, emotional intelligence, critical thinking and entrepreneurship.
This is neither an argument that AI will destroy every profession nor a comforting promise that every displaced worker will easily find another role.
The transition will be uneven, and the economic opportunities will not automatically be distributed fairly.
That brings us back to the question of data.
If human experience becomes the resource from which increasingly capable machines learn, how will the people and organisations generating that experience participate in the resulting economy?
Will knowledge become another resource extracted at scale by a small number of powerful companies?
Or will licensing, attribution, specialist expertise and more accessible marketplaces create a broader distribution of value?
Defined.ai is attempting to build part of the commercial infrastructure for the second possibility.
Its success is far from guaranteed. The competition is formidable, the economics demanding, and the technological landscape continues to change.
But the underlying proposition has become more relevant over the decade since Braga first appeared at Web Summit with DefinedCrowd.
In 2016, the challenge was teaching machines to understand the human voice.
In 2026, it is helping increasingly intelligent systems understand the complexity of human activities, specialised knowledge and the physical world.
The ambition has expanded enormously.
The connection to human experience has remained.
Perhaps the next great advance in artificial intelligence will come from larger models, more powerful processors or increasingly sophisticated synthetic environments.
But whatever machines eventually become capable of generating, their intelligence will still need to be tested against something beyond their own predictions.
The future of artificial intelligence will depend not only on how intelligently machines simulate the world, but on how responsibly we preserve, organise and value their access to reality.
And that may be the most enduring lesson of Daniela Braga’s decade-long journey: sometimes the most valuable part of artificial intelligence is the part that remains unmistakably human.



