Something connects three stories that emerged this week, and it has relatively little to do with artificial intelligence becoming ever more capable of generating things.

Remembering a conversation. Knowing how to practise a profession. Finding a film.

These are human, situated and relational practices. All three are beginning to become computational infrastructure. They also reveal the same prior operation. Before a task can be automated, the machine needs to make it legible. What was tacit, negotiated or distributed has to be translated into a format that can be processed, stored and queried.

A recording does not need to be kept for a social interaction to become data. A profession does not need to be automated for its knowledge to be extracted. An image does not need to be generated for a system to shape which imaginaries reach us.

The conversation that no longer disappears

On 9 September, Apple presented Live Rewind and Siri Recap for the new Apple Watch Series 12 and Ultra 4. The first feature can recover the previous fifteen seconds of a conversation as text. The second generates titles, summaries and key points so that what happened can be consulted later. Apple says both will arrive in beta towards the end of 2026, initially in English, and will not initially be available in the European Union.

The features are opt-in. Apple also says that Audio Intelligence does not identify speakers or attribute speech to specific people. Siri Recap can be configured to activate according to schedules or locations.

Apple has built an explicit privacy architecture around these functions. The S11 chip includes Secure Exclave, an isolated hardware compartment that handles raw audio. Live Rewind carries out transcription on the watch itself. Siri Recap also does not store the audio. After local processing, the text needed to produce the summary is sent to Private Cloud Compute rather than the original recording. Results saved in the Siri app are synchronised using end-to-end encryption.

When Live Rewind is activated, the watch plays a sound — even when the device is muted — displays a full-screen animation and shows that the microphone is in use. The result is deliberately short-lived. It disappears roughly thirty seconds after the screen dims.

Siri Recap works on a different scale of memory. Its summaries are automatically deleted after seven days unless they are explicitly saved. During that window they can be consulted and retained.

This is a sophisticated answer to technical privacy. It leaves another question open.

A conversation belongs to a relationship, while the decision to turn it into computational memory belongs to the person wearing the watch.

An asymmetry of remembering appears here. After a conversation, one person can retain a structured and retrievable representation of what happened while everyone else relies on biological memory. There need not be a recording. What remains can be something more operational in many settings, a summary that can be searched, consulted and reused.

There is also a striking difference in how the two functions are signalled. Apple explicitly documents audible and visual cues for Live Rewind. For Siri Recap, which can be scheduled by time or location, the public documentation available so far does not describe an equivalent cue directed at other people taking part in the conversation. That does not prove that no such cue exists. It does leave a relevant question open. Someone may have no comparable indication that the conversation is being turned into a retrievable summary.

Live Rewind is probably the smaller break with existing social practice. Fifteen seconds still resembles “sorry, what did you say?”. Siri Recap creates something different, a persistent and selective memory of the conversation.

That memory is not neutral either. Apple warns that generated text and summaries can be incomplete or incorrect and recommends checking important information with the people involved. Human memory also fails, but its contradictions are often negotiated between people. A system-generated summary introduces a third element into that negotiation, a representation that can appear to be a record rather than an interpretation simply because it has a technical form.

Human memory is imperfect, but that imperfection is not merely a defect. It allows diplomacy, forgiveness and the renegotiation of lived experience. When a machine summarises a conversation, it does not simply remember it differently. It introduces another kind of authority into remembering. That authority is not distributed symmetrically among those who took part.

The keynote also supplied the conceptual frame. Apple’s new CEO, John Ternus, described the iPhone as an “intelligent personal hub” and explained that this intelligence can work with calendars, relationships, routines and some of the most private details of daily life. The watch now extends that project to what happens around the body — what is said, heard and answered. Conversation stops being only an exchange and becomes an input to the system as well.

The initial exclusion of these functions from the European Union also creates an uneven regulatory geography. In the documentation cited here, Apple does not publicly attribute the decision to a specific rule, so assigning a particular legal cause would be premature. The difference nevertheless exposes a broader issue.

Consent is still treated as individual even when the data comes from a relationship.

References

Teaching the craft the machine is trying to learn

For years, a significant part of the invisible labour that made AI training possible involved classifying images, transcribing audio, moderating content or assigning labels. Rest of World now documents another phase in China.

Architecture, law, engineering, teaching, musical composition, journalism and accounting are entering an economy of specialised annotation.

Alibaba recruits expert profiles through Siriser, a specialised training platform. ByteDance says Xpert has brought in more than 50,000 experts. Other platforms are looking for a wide range of professional profiles. Economic slowdown and worsening labour conditions are pushing some of these people to use such tasks as an additional source of income.

What matters is what they are being asked to do.

They design realistic work situations, provide documents drawn from professional activity and describe their own processes so that the model can learn from them. They also create problems the system cannot yet solve and, in some cases, evaluate its answers. This is not only training. Companies are also buying the human ability to measure where machine competence ends.

Something deeper is happening than a simple transfer of information. Through years of practice, people develop knowledge that is not fully contained in manuals or documents. It includes judgement, shortcuts, intuitions about what works and what does not, and a sense of when a rule should bend. That tacit knowledge is difficult to transfer. These platforms are asking for precisely that transfer. What someone does without fully articulating it must become a statement that a machine can process.

Tasks can take several hours and, according to the people interviewed, pay between 100 and 500 yuan. The familiar risk structure of platform work remains intact even when the work requires highly specialised knowledge.

An architect in Shenzhen explains that she builds tasks from her actual work and teaches the model how to prepare construction proposals or profitability analyses. A professional with a master’s degree in law asks models to produce legal advice or draft a judgment and then evaluates the answers.

The data is no longer simply an image, a sentence or a document. It is the effort required to make explicit what years of practice had turned into tacit knowledge.

The data is the craft.

That creates a paradox. The growing precarity or declining value of some professions can turn professional knowledge into a supplementary source of income at precisely the moment when technology companies need that knowledge to build systems capable of taking on parts of the same work.

A less visible issue also appears. Rest of World reports that contributors use documents derived from real work, and one engineer expresses concern about the possible misuse of the files she submits. When professional experience becomes part of a dataset, the boundary between expertise, workplace documentation and confidentiality can become difficult to maintain.

The phenomenon is not exclusively Chinese. The investigation notes that US platforms such as Mercor, Surge AI and Handshake also recruit specialised profiles. What appears distinctive in China is the combination of scale, lower pay and a labour market supplying a growing pool of available expertise.

Advanced automation does not magically arrive after a machine has learned a profession. It first needs someone who knows the work to sit in front of the system and build the exam.

And the most familiar rule of platform labour still applies. The platform judging the quality of the task may also be the platform deciding whether it gets paid.

Reference

When finding becomes power too

More than 200 representatives from over 65 countries took part in the first Sommet Lumière in Saint-Paul-de-Vence on 7 September, co-chaired by France and South Korea. The meeting addressed the future of cinema and the moving image in a context shaped by platforms, artificial intelligence and changing economic models across the sector.

The Déclaration Lumière begins by doing something that may seem unusual for a technology summit. It tries to define the value of an image.

Audiovisual works are described as vehicles for emotion, memory and knowledge, capable of producing collective transformation and shared culture. The declaration also refuses to reduce images to the attention they capture.

That definition is not neutral. If an image were simply a unit of content competing for attention, the market and algorithms could be left to decide which works survive. Treating images as part of cultural memory allows a different kind of intervention to be justified.

One of the document’s most interesting concepts follows from that move — discoverability.

The sixth principle argues that, in an environment of abundance, ensuring that works can be found, circulated and connected with audiences has become as important as creating them. It explicitly mentions agentic artificial intelligence and the possibility that access to works will increasingly be mediated by automated agents.

The question is therefore no longer only whether AI can make a film.

It also matters which films it allows people to find.

For decades, a large part of cultural power involved financing, producing, distributing and exhibiting works. Another layer is now emerging. Recommendation systems, search engines and agents can place themselves between a potentially immense cultural supply and the people trying to reach it.

Not necessarily by banning works.

Simply by failing to surface them.

The general declaration takes a non-prohibitionist position on artificial intelligence. It recognises that AI can contribute to creativity, innovation and the circulation of works, while keeping human creation at the centre and treating respect for intellectual property as fundamental.

This should be distinguished from the more specific demands highlighted by the SACD around the declaration devoted specifically to artificial intelligence. The organisation stresses fair remuneration for rights holders whose works are used for training, alongside transparency and traceability mechanisms that would make it possible to identify the materials used and content generated or assisted by AI.

The Sommet Lumière also went beyond a declaration of principles. France and South Korea announced the Partenariat Lumière, a €1 billion fund running from 2027 to 2031, shared equally by both countries. France will mobilise €500 million through Bpifrance, with South Korea making an equivalent commitment. The summit also launched IRIS — International Resilience Initiative for Independent Screens —, an international initiative intended to strengthen independent cinemas and the circulation of works.

That slightly changes the question. The issue is no longer only whether there is political will to intervene. It is whether this architecture will have meaningful capacity against private platforms and agents that control the layer of access.

The cultural politics of the future may not be decided only by who is able to produce a work.

It may be decided by who controls the path leading to it.

References

Remembering, knowing, finding

The three stories place pressure on the same point.

The unit of consent, compensation or decision does not match the unit in which value is produced.

With Apple, an individual decision turns a relational good — conversation — into computational memory. Consent is given by the person wearing the watch, while the data comes from a relationship between people.

In expert work, a task paid for over a few hours contributes to knowledge that remains aggregated in a system long after the labour relationship ends. Compensation is micro and transactional. The value produced is macro and systemic.

At Lumière, states can set cultural principles and finance support mechanisms, while much of the effective power to make a work visible sits inside private layers of recommendation, search and, increasingly, personal agents. Creation is cultural and collective. Control over access is infrastructural and opaque.

For a long time, we thought about artificial intelligence as a machine learning to do things. These three signals make it possible to look from somewhere else.

Before doing them, it needs to make the human practices that enabled them legible.

The conversation negotiated through shared memory must become retrievable text. Knowledge accumulated through years of practice must become a processable statement. Culture circulating through multiple paths must become a recommendable vector.

That changes the question too.

It is not only whether the machine will remember a conversation accurately, learn a profession correctly or find a good film. It also matters who can be held accountable when the memory, the craft or the catalogue fails.

The answer is not symmetrical. With the watch, there is at least a product, a company and regulatory frameworks to address. In expert work, the platform judging the quality of a task may be the same one deciding whether it gets paid. In cultural discoverability, a work can stop appearing without any visible act of exclusion and without leaving an easily identifiable responsible party.

Those who supplied the raw material — the conversation, the craft, the work — rarely retain full control over the outcome.

Artificial intelligence is not only a machine that generates. Before that, it is machinery for making things legible.

And once legibility has been produced, it can move beyond the control of those who made it possible.