Artificial intelligence is often presented as a technology of expansion. More languages, more voices, more people able to interact with complex systems. Yet behind that promise lies a prior condition. To participate, you first have to become legible to the machine. And the rules of that legibility are not necessarily written by those who participate.
The three stories in this Radar explore that tension from different places. A statement on AI and Indigenous self-determination argues that inclusion is not enough if communities cannot decide what happens to their data, knowledge and territories. A study of generative music finds that models can do more than compress diversity within genres. Each system also carries its own learned assumptions about what a musical category ought to sound like. And in India, a telephone agent designed to make form-filling easier shows how a speech-recognition error can become a material obstacle to accessing a service.
Across all three cases, the issue is not simply what AI can do. It is who defines the conditions under which a person, a language or a culture can enter the system.
1. Indigenous Peoples — Sovereignty also includes deciding whether to use it
For the International Day of the World’s Indigenous Peoples, the Indigenous Commission of the International Federation of Social Workers published a statement that shifts the debate away from a relatively comfortable idea —making AI more inclusive— towards a more demanding one. Recognising the right of Indigenous Peoples to decide whether they want to use these technologies, how they want to use them and under what conditions.
The issue directly concerns data. Languages, stories, images and traditional knowledge can end up in digital repositories and training datasets. But the fact that knowledge is technically accessible does not mean there is permission to transform it into data for an artificial intelligence system.
The statement also goes beyond the contents of models. It stresses that digital infrastructure requires water, energy and land, placing that consumption within much longer histories of who controls resources. In the United States, it also notes that only around 40 per cent of the Indigenous population has access to broadband. Expanding technological infrastructure does not necessarily mean that those who bear some of its material costs share equally in its benefits.
The document also introduces a significant expression, Ancestral Intelligence. It uses the term for knowledge transmitted through relationships, ceremonies, stories, land and generations. A form of intelligence whose validity does not depend on first being digitised, classified or entered into a database.
This raises a question that comes even before the model is built.
Who has the authority to transform a culture into data?
Technological sovereignty means more than having access to one’s own tools. It can also mean retaining the right not to participate, setting limits around particular forms of knowledge, or deciding that some materials should not circulate beyond their cultural context.
Inclusion assumes that the system should find a way to bring in those who were outside it. Self-determination introduces another possibility. Communities themselves decide whether they want to enter and under what conditions.
2. Music — When a genre becomes its statistical average
A new study by Zoe Slendebroek and Danaé Metaxa compares music generated with Suno and Lyria 3 against human-produced music in four genres. Afrobeats, K-pop, Dance Pop and Heavy Metal.
The study analyses one hundred tracks per system and genre using 72 music information retrieval, or MIR, features to investigate two different questions. How much diversity exists within each category and how distinguishable the genres remain from one another.
The results reveal two different logics of homogenisation.
Lyria 3 tends to reduce diversity within some genres. Its outputs cluster more tightly around a particular acoustic profile. Suno, by contrast, retains somewhat more variation inside each category while pulling closer together genres that remain more distinct in human recordings.
Not every model, then, homogenises in the same way.
There is another particularly important detail. Some of the tests used nothing more than the genre name as the prompt, with no additional stylistic instructions. The patterns observed therefore cannot be explained solely by what the person using the system asks for. They also point to the models’ priors —learned tendencies acquired during training— about what Afrobeats, K-pop, Dance Pop or Heavy Metal ought to sound like.
That matters because a musical genre is never merely a combination of sonic features.
It is also a history, a scene, an industry, a geography, an audience, a way of dressing, a set of distribution circuits and a collective negotiation over what belongs —or ceases to belong— to a particular tradition. Its boundaries change because people who make and listen to music argue over them, mix them and transform them.
A generative system learns that diversity from large numbers of examples. Yet to produce something recognisable, it also has to extract regularities.
The risk begins when what is statistically most representative starts to replace the tradition itself.
What happens when a genre stops being a social practice in transformation and starts becoming the statistical centre of what a machine has learnt it ought to be?
If generative music becomes more prominent on platforms, in advertising, games or audiovisual production, this homogenisation ceases to be merely an aesthetic issue. Models can also influence which sounds circulate more widely, which are easily recognised, and which cultural forms are pushed towards the margins of a category.
A machine does not need to ban a variation to make it less visible. It may be enough to learn that it is less representative.
3. India — When a machine that does not understand your voice can block a right
In India, a large number of benefits and services begin with a form. For people who face literacy barriers or have little familiarity with digital interfaces, that form can itself become an obstacle.
FormBharo —“fill the form” in Hindi— proposes replacing part of that written interaction with a telephone conversation. The system conducts a spoken dialogue, extracts the information it needs and uses it to complete the relevant fields.
It is being evaluated with ARMMAN, an organisation working in maternal and child health, in a particularly sensitive context. Helping low-income mothers enrol in antenatal and postnatal care programmes.
Here the interface is not a website or a written chatbot.
It is the voice.
And that difference matters.
The research team developed FormVoiceAgentBench, a benchmark comprising 3,760 evaluations across 960 simulated calls. When models work from clean reference transcripts, results can be very high. But when they have to process the speech that actually comes through a telephone call —noise, pronunciation, pauses, voice variation and automatic speech-recognition errors— correct form completion can fall by as much as roughly 41 percentage points.
A mistranscribed word becomes an incorrect piece of data. Incorrect data can produce an incorrect form. And an incorrect form can affect access to a service.
The study also reveals an interesting technical paradox. GPT-5.5 reaches roughly 99.8 per cent accuracy in some turn-by-turn extraction tasks using reference transcripts, yet that strong result does not automatically make it the best system when the entire conversation is evaluated.
Errors accumulate, propagate and sometimes interact.
For that reason, smaller models combined with deterministic controls —rules that check and constrain particular stages of the process— can match or outperform more powerful systems when the final outcome is measured.
The lesson travels beyond FormBharo. Optimising every component of a system does not guarantee that the whole system will work better.
But for the Radar, the more important question lies elsewhere.
Digital administration has spent years turning certain forms of knowledge —reading instructions, navigating a website, having an email address, understanding bureaucratic categories— into silent requirements for exercising rights. A voice interface can reduce some of those barriers.
It can also create new ones.
What happens when making yourself understood by a machine becomes one of the steps required to make yourself understood by an institution?
Accessibility does not depend simply on whether a model “knows” Hindi. It depends on how it hears Hindi as people actually speak it, which variations it considers acceptable, which errors it tolerates and, above all, what happens when it gets something wrong.
Because once AI becomes an intermediary for accessing a service, the decisive question is not merely how many errors it makes.
It is who bears their consequences.
One question shared across all three — The conditions of legibility
A community deciding which forms of knowledge may become data. A musical genre whose diversity is compressed as it passes through a model. A person who needs a system to recognise their voice correctly in order to complete a form.
These are different situations, but they share a structure.
Artificial intelligence systems need to transform the world into representations they can process. Languages, sounds, identities, forms of knowledge and cultural practices have to acquire a form that is legible to the machine.
That process is never entirely neutral.
Deciding what counts as data, which variations are treated as “noise”, which examples stand for a category, or which information can be used to train a system means drawing boundaries.
And legibility does not affect those on either side of those boundaries in the same way.
For a technology company, it may be a condition required for a model to function. For a community, it may mean deciding which parts of its knowledge become exposed. For a musical tradition, it can change what counts as representative. For someone trying to complete an administrative process, it can become a condition for accessing a service.
The anthropological question, then, is not simply whether AI represents human diversity accurately.
There is an earlier question.
Who decides the conditions under which that diversity has to become legible to the machine?
References
- International Federation of Social Workers — https://www.ifsw.org/artificial-intelligence-ai-and-indigenous-peoples-self-determination/
- Zoe Slendebroek and Danaé Metaxa — https://arxiv.org/abs/2608.06106
- Aman Dalmia et al. — https://arxiv.org/abs/2608.06027