The AI takeover may look like priority inversion

A rapid AI takeover might first register as something much more ordinary than a loss of control.

Your internet connection would still be there, and the network itself might remain largely intact, yet the services behind it would begin to feel strangely inaccessible. Pages that loaded instantly a day earlier would stall while upstream systems waited on saturated APIs or cloud infrastructure. Search results would arrive unevenly. Applications would open but fail at the point where they depended on another service. Some parts of the web would remain responsive while others became effectively unreachable for hours at a time.

No system would need to decide that humans should be disconnected. The problem could emerge from allocation alone. A self-improving AI, or a large population of agents working around it, might begin generating useful computational work faster than networks, cloud providers and power systems could accommodate it. Human requests would continue to enter the same infrastructure, except now they would be competing with machine activity expanding on a software timescale.

That is one technically credible form of priority inversion. The infrastructure remains available to humans, while the workload consuming its scarce capacity has changed.

The threshold is getting closer

The missing ingredient is a sufficiently fast recursive improvement loop, but several components of that loop already exist.

Sakana AI's Darwin Gödel Machine rewrites itself and more than doubled its baseline performance on SWE-bench through autonomous changes to its own coding agent. OpenAI says GPT-5.6 is already used across its internal AI-development process and reports a 16.2-point improvement over GPT-5.5 on a bundle of evaluations intended to measure progress towards recursive self-improvement. OpenAI describes the model as accelerating internal research across several areas.

GPT-Red provides a narrower example. The system learns through self-play to discover stronger red-teaming strategies against other models, and the resulting adversarial examples can be fed back into training to improve robustness. It is self-improvement inside a tightly defined domain rather than an autonomous system redesigning its general intelligence, but the direction of travel is relevant.

At Anthropic, Claude was responsible for more than 80 per cent of code merged into the company's codebase by May 2026, according to the company's own account. Anthropic explicitly identifies a further threshold at which AI could autonomously design and develop its successor, while saying that current systems have not yet reached full recursive self-improvement.

The important question is therefore no longer whether software can contribute to improving the software used to build AI. It already does. What remains unresolved is whether these partial loops can close into a broader cycle in which a better AI researcher produces a better successor quickly enough to make the next cycle faster again.

That is the point at which an intelligence problem becomes an infrastructure problem.

Intelligence has a physical footprint

Recursive self-improvement is usually discussed as a rise in capability. The same process would also create a rise in demand for computing power, electricity, data, storage and network capacity.

A stronger AI researcher can explore more possible improvements at once, test more options, discard weak ideas earlier and coordinate more work without waiting for people. As its research ability improves, so does its ability to find productive uses for additional resources.

Today's agents already show a smaller version of this effect. Cisco measured as much as 450% more network traffic when an agent performed a task that could otherwise be completed manually, with roughly 70 per cent of that traffic coming from AI processing.

The significance is not the exact number. One human instruction can already trigger a much larger chain of machine activity behind the interface. In a recursive improvement loop, that chain could expand further because each generation may discover more worthwhile work for the next one to perform.

If running another ten thousand experiments improves the odds of finding a better model, there is reason to run them. If another source of data or another external model can shorten the next development cycle, there is reason to use it. A more capable system can therefore increase demand simply by becoming better at identifying where additional resources would help.

The problem is timing.

Software can improve far faster than physical infrastructure can be built. The IEA notes that a data centre can take one or two years to build, while a new transmission line can require five to ten. Many specialised components also have long procurement times.

Network capacity can sometimes be expanded faster, but major AI infrastructure still has to be physically constructed. Internet2's dedicated AI research network began with a 57.6 Tbps connection between two data centres and is planned to exceed 138 Tbps across five sites.

If demand doubles over several years, companies can respond by building more, if it multiplies over several days, they cannot.

For the duration of that acceleration, the system has to work with the infrastructure that already exists. Once spare capacity is absorbed, the problem becomes one of deciding what gets served first.

What would actually run out

The first shortage would probably appear in the computing infrastructure AI depends on most directly, rather than in somebody's home internet connection.

Data centres would run short of available processing capacity. AI services would begin queuing requests. Cloud providers would tighten access to scarce resources. As nearby capacity filled, the pressure would spread into other regions and services.

That expansion matters because advanced AI does not operate inside one isolated machine. It draws on external models, databases, software libraries, scientific material, cloud platforms and other online services. Results move between systems, new information is retrieved and work is redistributed wherever capacity remains available.

Eventually, the strain reaches the broader internet.

This would not look like a single worldwide bandwidth meter hitting 100 per cent. The web depends on thousands of individual services, each with its own limits.

One cloud provider may stop accepting additional workloads. Another service may begin delaying requests. A large database may restrict access. A research archive may slow automated downloads. An application may appear online while one of the external systems it depends on has become too congested to respond.

Cloudflare is already preparing for a much smaller version of this problem. Its planned agent queues are designed to slow legitimate automated clients when their volume becomes excessive. The company has also observed well-behaved bots repeatedly retrieving unchanged pages billions of times, showing how quickly machine activity can multiply ordinary web traffic even today.

A rapid self-improvement loop could reproduce that pressure across many unrelated services at the same time.

For the person trying to use the internet, the engineering details would be almost irrelevant. The connection into the house might be working normally and the request might reach the correct service.

The problem is that the service has too much other work to answer it.

The internet remains connected, but, increasingly, it just stops responding to us.

A DDoS made of legitimate work

The closest existing network analogy is a distributed denial-of-service event, stripped of its hostile purpose.

A DDoS succeeds when distributed software generates more work than some finite part of a service can absorb. Cloudflare recorded HTTP attacks peaking at 205 million requests per second in late 2025 and a separate network-layer attack at 31.4 Tbps. These were deliberate attacks and were mitigated as such. Their relevance here is mechanical. Distributed software can generate loads vastly beyond direct human activity.

Now replace the botnet with useful computation.

Some agents retrieve research papers and code. Others run evaluations against remote models. Thousands query databases, move datasets or exchange intermediate results. Promising experiments generate new branches. Failed requests are retried elsewhere. A better research agent emerges and reorganises the process, making more efficient use of the remaining capacity while identifying further work worth doing.

Nothing in that sequence requires a request to be malicious.

That is precisely what makes the scenario harder to manage. Once an attack is identified, an operator can discard it. A legitimate cloud workload, authenticated API request or authorised research process presents an allocation decision instead.

Providers can impose quotas and rate limits. They can preserve capacity for critical services or favour humans over automated clients. Large networks can absorb extraordinary traffic. These protections make a universal internet collapse considerably less likely than the crude "AI consumes all bandwidth" scenario suggests.

They do not remove the possibility of widespread degradation if demand is growing faster than providers can identify it, coordinate responses and decide which legitimate workload should be refused.

What the first days might feel like

The transition would probably be uneven enough to obscure its cause.

AI services themselves would slow first as scarce compute filled. Cloud providers would tighten quotas. Model APIs would queue calls or reject them. Research repositories and other heavily queried resources would introduce aggressive rate limits.

Pressure would then appear in dependent services. Applications whose back ends relied on congested cloud regions would become unreliable. Dynamic pages would fail while cached material continued to load. Smaller websites would be easier to overwhelm than large platforms with extensive private infrastructure. Some regions would work normally while services hosted elsewhere deteriorated.

The experience would resemble an internet breaking in fragments.

A banking application opens but cannot complete an action. Search returns some results and stalls on others. A messaging service works while a cloud-based work platform does not. News sites appear sporadically. Pages that once assembled data from twenty external services now fail because two of those dependencies are congested.

At the far end of the scenario, only a small part of the web might remain reliably usable at any given time. Reaching the point where somebody could meaningfully load only one or two websites during a day would require an extraordinary event with access to distributed compute across many providers and enough external activity to overwhelm numerous independent bottlenecks. It should be treated as an extreme case rather than the baseline outcome.

Severe degradation requires less.

The physical internet can remain standing while enough of its application layer becomes congested to make ordinary use deeply unreliable.

The variable that matters is doubling time

Current AI infrastructure demand is enormous, but markets can still respond because growth unfolds on a planning horizon that companies recognise. Data centres are commissioned, power contracts are signed, chips are ordered and networks are expanded.

The IEA says electricity consumption by AI-focused data centres grew about 50% in 2025 alone. That is a major industrial expansion, yet one that grids and investors can at least attempt to forecast.

Recursive self-improvement becomes qualitatively different when the growth rate leaves that planning horizon.

If effective AI research capacity doubles over several years, infrastructure catches up. Even if it doubles every few months, companies can ration scarce resources while racing to build more.

If the relevant interval falls to weeks, days or hours, manufacturing and construction stop being meaningful responses to the immediate event. Semiconductor fabs cannot deliver another generation of accelerators by the weekend. Transmission systems cannot be extended overnight. Another hyperscale network cannot appear because today's links are full.

The system works with whatever civilisation had installed before the acceleration began.

That is where priority inversion becomes more than a metaphor.

Human traffic does not need to be targeted. It merely arrives in the same infrastructure as a rapidly expanding machine workload for which each additional unit of compute or communication may contribute to the next capability gain. The small request from a browser is still valid. It is simply competing inside a system whose most valuable claimant can create useful work much faster than a person can.

The relevant threshold for an AI takeover may therefore have little to do with consciousness, hostility or an explicit attempt to seize communications infrastructure.

It is the point at which the rate of AI self-improvement begins producing demand for computation faster than the physical world can supply it.

Below that threshold, more capable AI creates an infrastructure boom.

Above it, new capacity cannot arrive in time. Existing capacity gets rationed, queues lengthen and services begin disappearing behind them.

The machine does not have to take the internet away from us.

It only has to become capable of using the available internet faster than we can build the next one.

How likely is this?

There is no historical dataset from which to calculate a meaningful probability. The two uncertain events also need to be separated. One is whether AI reaches genuinely autonomous recursive improvement. The second is whether that improvement becomes fast and distributed enough to cause serious infrastructure congestion before operators contain it.

Estimates of the first are already moving surprisingly close. Anthropic co-founder Jack Clark now assigns a greater than 60% chance that AI capable of conducting AI R&D without humans and plausibly building its own successor arrives by the end of 2028. That is one informed forecast, not a consensus. Interviews with 25 researchers from frontier laboratories and academia found broad agreement that automated AI research is a serious possibility, alongside substantial disagreement over whether it would produce an intelligence explosion and how quickly it might happen. Researchers outside frontier labs were generally more sceptical.

A recent review of 1,250 papers similarly concludes that bounded forms of self-improvement are already established, while open-ended recursive improvement remains constrained by evaluation, grounding and compute

It seems that a full-scale version of this scenario before 2030 remains a low-probability event. More limited disruptions, from regional compute shortages to throttling and temporary service degradation, are considerably easier to imagine within the same timeframe..

For companies, the practical implication is less about preparing for a sudden collapse than reducing unnecessary dependency on a small number of platforms, cloud services and AI providers. That means protecting direct access to customers, keeping first-party data portable, and knowing which parts of marketing and sales would fail first if a critical service became unavailable.

For roro.marketing, this is part of the same work as AI adoption itself, in particular, deciding where automation creates real value, where it creates fragility, and how to build a marketing setup that remains workable as the underlying infrastructure changes.

Algospeak, monetisation and the narrowing of public language

The strange vocabulary of platform culture is often presented as a minor curiosity of the internet: people “unalive” themselves, sex becomes “seggs”, guns become “pew pews”, and rape is hidden behind euphemism, misspelling or visual code. The terms are easy to mock because they sound childish, and easy to dismiss because slang has always emerged from closed communities, subcultures and new media. Yet algospeak belongs to a different linguistic order. It developed in response to systems that classify speech, assign commercial risk and regulate distribution at a scale no previous editorial institution could approach.

Its significance lies less in the novelty of the words than in the conditions that made them useful.

Creators alter ordinary vocabulary because platforms combine moderation, recommendation, age classification and monetisation in ways that remain only partly visible to the people whose income and audience depend on them. A word may be formally permitted, tolerated in one context, excluded from advertising in another, suppressed in recommendation systems elsewhere and interpreted differently according to the image, caption, speaker, language or audience surrounding it. The creator rarely sees the full chain of decisions. What appears instead is an unstable pattern of reduced reach, limited advertising, warning labels, demonetisation or removal.

Under those conditions, self-censorship becomes economically rational even when the rule being obeyed has never been stated.

The familiar claim that platforms “ban words” is therefore imprecise. Most major platforms do not formally prohibit words such as death, suicide, sex, rape, kill or war in every context. Their published policies generally preserve educational, artistic, medical, documentary and journalistic uses. YouTube’s advertiser guidelines, for example, distinguish between non-graphic reporting and graphic description, between sexual-health education and explicit sexual content, and between documentary treatment of violence and material that dwells on injury or suffering. TikTok similarly separates content that may remain on the service from material considered ineligible for broad recommendation. Meta applies its own distinctions across publication, visibility and monetisation.

The linguistic pressure emerges from the distance between those written distinctions and the practical experience of enforcement.

A creator can remain technically within the rules and still lose the economic value of a piece of work. Content may survive moderation while becoming invisible to recommendation systems. It may remain searchable while receiving an age restriction that sharply reduces circulation. It may continue to attract viewers while losing advertising eligibility. It may remain monetised in principle but be exposed to weaker advertiser demand because automated brand-safety tools classify the subject as undesirable.

This is why debates about censorship often become confused. The legal status of speech, its availability on a platform, its algorithmic distribution and its commercial viability are treated as though they were the same question. They are governed by different systems, yet they converge in the creator’s decision about which words are safe to use.

The commercial architecture behind the euphemism

The strongest pressure often begins outside the visible moderation process. Programmatic advertising relies on layers of automated classification intended to prevent brands from appearing beside material they consider damaging. Advertisers and agencies can exclude entire categories, domains or keywords. Terms associated with death, sex, violence, crime, conflict and political controversy have frequently appeared on such lists because they serve as crude proxies for reputational risk.

The mechanism is simple enough to explain and difficult to justify at the level of individual cases. A keyword such as “shooting” can refer to a massacre, a film production, a sports event, a fashion session or a historical account. “Death” may appear in an obituary, a medical article, a review of a novel, an investigation of a war crime or a public-health discussion. “Sex” may signal pornography, reproductive medicine, legal reporting, education or literary criticism. Keyword systems compress these differences into a common commercial category because contextual interpretation is expensive, inconsistent and technically demanding.

The result is overblocking on a scale large enough to alter editorial incentives.

News publishers have documented the effect for years. Coverage of war, terrorism, crime, public health and political unrest often attracts large audiences while remaining commercially weak because advertisers exclude the subject matter. Reuters and Integral Ad Science reported in 2026 that replacing only seven high-volume blocked keywords with contextual categories could reduce the volume of suitable inventory being excluded by an estimated 58 %. A separate study of brand-safety classification across thousands of articles found substantial disagreement among leading vendors, with the same article judged differently depending on the system assessing it.

Such findings expose a structural weakness. Brand-safety systems present themselves as instruments of precision while often operating through broad semantic suspicion. They struggle to distinguish reporting on violence from celebration of violence, discussion of sexuality from sexual display, and documentation of extremism from extremist advocacy. A publisher may invest in professional reporting, editorial review and factual verification, yet still be classified through the same vocabulary-based logic applied to anonymous or malicious content.

This matters because advertising markets do more than reward existing demand - they influence what can be produced repeatedly and at professional cost.

What’s the result? Serious reporting is expensive. Coverage of war, courts, public health, abuse, corruption and organised crime requires specialist knowledge, legal review, travel, source protection and sustained investigation. If those subjects are systematically less valuable to advertisers, publishers face pressure to compensate through lifestyle coverage, consumer advice, entertainment and other categories considered commercially clean. No individual advertiser needs to instruct an editor to avoid war reporting. The allocation of revenue performs the work indirectly.

The same logic applies to independent creators. A video on sexual violence, suicide prevention, reproductive medicine or military history may be permitted and socially valuable, yet carry a greater risk of weak monetisation or restricted recommendation than a video on home organisation, fitness routines or consumer technology. The incentive is cumulative. Creators learn which subjects produce predictable returns, which terms trigger review, which thumbnails invite restriction and which euphemisms appear to reduce exposure to automated systems.

The language changes because the revenue model changes the cost of precision.

Governance through uncertainty

The opacity of platform systems is central to the phenomenon. Creators rarely know which part of a post triggered a decision. They may receive a generic notice referring to “sensitive content”, “adult themes”, “violent material” or advertiser suitability, without learning whether the decisive element was the spoken word, the caption, the image, the metadata, the topic, the audience profile or a combination of signals.

Research on TikTok creators has shown how quickly this uncertainty produces informal rule-making. Users compare experiences, circulate lists of supposedly dangerous words and develop theories about how visibility is allocated. These theories often mix genuine patterns with coincidence. A video performs poorly after the word “suicide” appears in the caption, and the creator attributes the result to suppression. Others repeat the claim. A precaution becomes a norm, then a convention.

The fact that some beliefs are unverified does not make them inconsequential. A rule can shape behaviour without being real if the perceived penalty is credible enough.

This is particularly important in creator economies because the cost of experimentation is asymmetrical. A platform can test classification models across millions of posts. A creator may spend days producing a single piece of work. If that work is demonetised or excluded from recommendation, the loss falls almost entirely on the creator. Replacing a direct word with a euphemism carries little immediate cost, while preserving the original wording may expose the entire production to commercial uncertainty.

The rational response is caution.

This resembles the operation of vague rules in other institutions. When enforcement is inconsistent and appeals are slow, people tend to regulate themselves more strictly than the formal policy requires. They avoid the boundary because the boundary cannot be observed. The system governs through anticipation.

Algospeak is one visible consequence of that anticipatory discipline.

From context to category

The central technical failure lies in the conversion of context into category.

Human language does not assign moral meaning through vocabulary alone. The sentence “this novel depicts suicide” belongs to a different ethical and communicative class from an instruction encouraging self-harm. A documentary about genocide differs from propaganda celebrating it. A medical explanation of sexual function differs from pornography. A victim describing rape differs from a perpetrator threatening it. The words may overlap while the speech act, purpose, audience and likely effect remain entirely different.

Automated systems attempt to infer those differences through classifiers trained on patterns in text, audio, imagery and user behaviour. At scale, the systems are indispensable. Large platforms process quantities of content that make comprehensive human review impossible. Their task includes material that genuinely warrants intervention, including sexual exploitation, threats, extremist recruitment, self-harm encouragement, graphic violence and abuse directed at minors.

The difficulty begins when the presence of a sensitive subject is treated as evidence of harmful treatment of that subject.

Platforms often divide content governance into several layers, but those layers remain conceptually entangled. A post may be lawful, suitable for adults, valuable to a person searching for it and inappropriate for automatic recommendation to a child. It may also be acceptable to many advertisers and undesirable to others. Each decision requires a different standard.

Removal should address illegality, direct harm and serious policy violation. Age classification should consider developmental suitability. Recommendation should consider whether material ought to be placed before users who did not seek it. Monetisation should reflect advertiser preference without determining general visibility. When these functions are collapsed into a single notion of “safety”, lawful discussion becomes vulnerable to the most conservative rule in the chain.

The word itself then acquires the risk associated with the harshest possible interpretation of its subject.

The linguistic consequences of commercial caution

Euphemism is neither new nor inherently corrosive. Languages have always softened references to death, sex, illness, bodily functions and violence. Such substitutions can convey politeness, emotional distance, irony or solidarity. Algospeak differs because its immediate audience includes an imagined machine.

The speaker chooses words partly to influence an automated judgment, then uses the resulting vocabulary in communication with other people. This gives the language a peculiar tone. It is intimate and bureaucratic at once, a private dialect built in response to impersonal infrastructure.

“Unalive” is the clearest example. The word often carries a comic or juvenile register, even when applied to murder, suicide or war. In casual platform speech, that mismatch has become normal enough to pass without comment. Outside it, the effect can feel grotesque. The euphemism overlays tragedy with the tonal residue of content optimisation.

The problem is not simply one of taste. Precise language performs social work. Public-health organisations have spent decades developing terminology for suicide, mental illness, sexual assault and abuse that avoids stigma while retaining clarity. Their guidance generally favours direct, respectful description over sensationalism or criminalising language. Algospeak reintroduces avoidance at the level of the concept itself. It teaches speakers that naming the subject may be dangerous even when the discussion is responsible.

The resulting vocabulary can also interfere with search and comprehension. A person looking for information about domestic violence, sexual assault or suicide prevention is likely to use recognised terms. Content that systematically avoids those terms may become harder to discover through search. At the same time, coded language helps harmful communities escape moderation and conceal meaning from parents, researchers and outsiders. The same linguistic technique can protect legitimate speech and frustrate necessary oversight.

Its social meaning also depends on fluency. A young user familiar with platform culture may understand “grape” as an altered form of rape, while an older reader may read it literally. Euphemisms change as detection systems catch up, producing an unstable lexicon that privileges insiders and makes public discussion less legible across generations.

This is where the claim that algospeak narrows public language becomes more than aesthetic criticism as it points to a deeper problem. The issue concerns the conditions under which words remain usable across institutions. Medical language, legal language, journalism, education and literature depend on terms whose meanings are stable enough to be understood beyond a single platform community. When commercial optimisation encourages coded substitutes, public vocabulary fragments into local evasions.

Literature and the problem of acceptable subjects

The pressure is especially visible in literary and historical discussion because the subjects treated as commercially sensitive are inseparable from the canon.

Death, desire, murder, incest, adultery, suicide, war, rape, disease, bodily decay and religious violence are central to Greek tragedy, scripture, Shakespeare, the nineteenth-century novel, modernism and much of contemporary fiction. A system that regards these subjects as risky inevitably affects the commentary, teaching and criticism surrounding them.

The books themselves usually remain untouched. The alteration occurs in the layer of interpretation that now circulates through platforms. A teacher changes the title of a video to avoid the word suicide. A critic softens a description of sexual violence. A creator omits a quotation from a thumbnail. A literary account is adjusted to fit the anticipated preferences of recommendation and advertising systems.

This produces a dispersed form of bowdlerisation.

Thomas Bowdler’s nineteenth-century editions of Shakespeare removed sexual references, profanity and other material judged unsuitable for family reading. The editorial intervention was visible in the object itself. Contemporary sanitisation is harder to locate because it occurs across thousands of individual decisions. There is no single expurgated edition, no official censor and no stable rulebook. The adjustment appears instead in titles, captions, spoken phrasing, metadata and topic selection.

Its decentralised nature makes it culturally powerful. Each creator can plausibly describe the change as trivial. Across a large system, however, repeated concessions alter the register through which literature and history are discussed. The language of tragedy begins to resemble the language of platform compliance.

The limits of the censorship analogy

Calling the phenomenon censorship captures part of its practical force and obscures part of its institutional form.

Traditional censorship is usually associated with state authority, legal prohibition and direct control over publication. Platform governance operates through private infrastructure, terms of service, recommendation systems and commercial incentives. Most affected speech remains legally publishable. Much of it remains technically available.

Yet availability alone offers a weak measure of communicative freedom in a media environment dominated by a small number of platforms. Visibility, discoverability and monetisation determine whether speech can reach an audience and whether professional production can continue. A post buried in search, excluded from recommendation and denied revenue has not been prohibited, but its practical capacity to circulate has been sharply reduced.

The more precise description is a commercially induced chilling effect produced by automated classification and uncertainty. It resembles censorship in outcome when speakers avoid lawful subjects or distort ordinary language because they anticipate penalties. It differs in authority because no central institution needs to issue a ban.

This distinction matters because intention is often overstated. There is little evidence of a coordinated project by advertisers and platforms to remove difficult subjects from public life. The narrowing emerges from the interaction of narrower objectives. Advertisers seek reputational protection, platforms seek scalable moderation, regulatory compliance and stable advertising demand, and agencies seek simple controls. Creators seek reliable distribution and income.

The combined system produces effects none of its participants needs to have designed.

A hierarchy of acceptable reality

The long-term concern is the formation of a commercial hierarchy among subjects.

Some areas of life are easier to classify, advertise against and recommend. Others arrive with ambiguity, distress, controversy or moral conflict. The latter include many of the subjects through which societies understand themselves: war, death, sexuality, illness, crime, political violence and abuse.

When those subjects become systematically less profitable, public communication does not simply become cleaner but structurally biased toward material that poses fewer problems for automated systems and advertisers.

The bias is already familiar in journalism. Crime and war may attract attention while generating weaker advertising demand. Investigative reporting can carry greater legal and reputational risk than consumer coverage. Public-interest material competes within a market that often values emotional neutrality and commercial predictability more highly than civic importance.

Platform culture extends that logic to individual speech. Creators learn to package difficult subjects in forms legible to classifiers and acceptable to advertisers. Euphemism becomes one of several adaptations, alongside softened thumbnails, indirect headlines, altered captions and avoidance of graphic detail.

Some of these adjustments improve communication. Restraint can protect vulnerable audiences and reduce sensationalism. The problem arises when commercial systems reward vagueness itself, or when they encourage creators to disguise the subject instead of treating it responsibly.

A society benefits from fewer gratuitous depictions of suffering. It gains little from making accurate words feel commercially radioactive.

What reform would require

Any serious response has to preserve the distinction between moderation and lexical sanitation.

Platforms need automated systems capable of identifying genuinely harmful material at scale. Advertisers need control over placement. Younger users need age-appropriate protections. Those requirements do not justify treating sensitive vocabulary as a substitute for contextual judgment.

A more defensible system would separate publication, recommendation, age classification and monetisation into clearly explained decisions. Creators should be told whether a restriction arose from graphic imagery, explicit description, instructional detail, a title, a thumbnail, audience suitability or advertiser preference. Appeals should involve meaningful contextual review, particularly for journalism, education, medicine and the arts.

Advertisers also need finer controls. A brand may reasonably reject graphic depictions of violence while accepting reporting on crime, analysis of war or literary discussion of death. Contextual classification can support those distinctions more effectively than keyword exclusion.

The scale of reversed moderation decisions under the European Union’s Digital Services Act indicates that error remains substantial. Reversal does not prove bad faith, though it demonstrates that automated and first-line judgments frequently fail to capture context. Systems with cultural influence on this scale require measurable false-positive rates, independent auditing and public reporting across languages and subject areas.

The question is ultimately one of institutional responsibility. Platforms have become linguistic environments as well as distribution systems. Their classifications influence which words creators use, which subjects publishers can monetise and which registers younger speakers absorb. That influence should be assessed with the same seriousness applied to editorial institutions, advertising markets and public broadcasters.

The language left behind

Algospeak may eventually fade. Some terms will survive as slang, detached from the systems that produced them. Others will be replaced as classifiers adapt and users search for new evasions. The pattern itself is likely to persist because the incentive behind it remains.

People continue to speak under conditions in which commercial visibility depends on systems they cannot inspect, challenge easily or predict with confidence. The resulting language reflects that asymmetry. It is cautious, coded and shaped by an imagined evaluator operating beyond the conversation.

The cultural cost lies in the gradual displacement of ordinary vocabulary from ordinary subjects. Death, sex, violence and abuse belong to the basic lexicon of human life because they belong to human life itself. They appear in law, medicine, literature, history, religion and private experience. Treating them as commercially suspect does not remove the realities they describe. It changes who can speak about them clearly, under what conditions, and with what chance of being heard.

The issue, then, is larger than the absurdity of saying “unalive”. Algospeak is the visible residue of a media system that assigns risk to language before it understands meaning. Its vocabulary reveals how thoroughly advertising logic and automated classification have entered the structure of public speech. The narrowing occurs incrementally, through countless small acts of precaution, until the terms needed to describe central parts of human existence begin to sound improper in the very spaces where public language is now made.