19 May 2026 by Mondo2.0

Token poor

There’s a new digital asymmetry emerging in our daily lives. It’s not a lack of connectivity, nor technological illiteracy in the classic sense of the term. It’s something more subtle, and in some ways more revealing: running out of tokens .

But what exactly are tokens?

When communicating with a large language model—what we now call LLMs, or more simply “AIs”—text isn’t processed word by word, nor letter by letter. It’s broken down into smaller, statistically significant units called tokens.

A token can correspond to a whole word, a root word, a suffix, or a punctuation mark. In Italian, a word like “ragionevolmente” might be worth three or four tokens.

Every time you write something to the model, and every time the model responds to you, tokens are consumed. The longer and more complex a work session, the more tokens are consumed with each exchange, with a progression effect that can quickly become voracious.

Each subscription or usage plan today comes with a certain number of tokens available over a period of time. We’re still in the early stages of this technology, and the way we pay for it reflects its commercial immaturity. It’s not unlike what we experienced with cell phones: first, we paid for each SMS as if it were a telegram, then for minutes of talk time, then for megabytes of data, then for gigabytes in a monthly bundle. Each step seemed obvious until the next one came along and made it obsolete.

The same will likely happen with LLMs: tokens are the unit of measurement today, but it’s reasonable to expect them to evolve towards different models—flat subscriptions, usage-based pricing, differentiated institutional access, perhaps something we can’t yet imagine.

For now, however, when the tokens run out, the machine stops and no longer converses.

This mechanism is closely related to another distinction that is worth clarifying: that between so-called prompting and model training.

Training a model is a comprehensive, often expensive, process that takes place in the laboratories of large technology companies and requires working with enormous amounts of data. This significantly improves the quality and professionalism of responses in specific thematic areas.

Prompts, on the other hand, are what the user does; they’re based on conversation: it’s how you formulate a request, build a context, and guide the model toward the desired outcome. It doesn’t change the model, it doesn’t improve it structurally, and it doesn’t leave permanent traces. It’s conversation, not teaching. At the same time, while it’s a temporary conversation, it’s a real skill, requiring practice, linguistic sensitivity, and a certain understanding of how the machine thinks—or rather, how it simulates thinking.

Those who know how to prompt effectively achieve qualitatively different results than those who improvise. And it’s an activity that consumes tokens, obviously.

The metaphor of rationing is hard to avoid: when tokens run out, conversation stops, only to resume tomorrow, as if digital thinking had a closing time.

A new digital divide is emerging: on one side, those who master prompting, have premium plans, and have already integrated these tools into their daily workflow; on the other, those who haven’t yet fully grasped the added value these systems can offer, or simply lack the resources to use them consistently and consistently, are clearly at a disadvantage.

It’s not just a technical or functional difference, as with the traditional “digital divide.”

It is a deeper stratification, which affects the capacity to produce, elaborate and valorise knowledge and which risks reproducing, in a new form, already existing inequalities .

Public institutions face a very delicate challenge: ensuring that access to these tools does not become a new axis of inequality. Schools, universities, local authorities, libraries, and healthcare services must grapple with technologies that could greatly improve the quality of their daily work, but which risk becoming the exclusive prerogative of those who already have resources, skills, and digital acumen.

Making it easier for citizens means thinking about forms of public, subsidized, and informed access to this new generation of services.

Private companies that adopt these tools have the opportunity, and in many cases the cost-effectiveness, to provide their employees with adequate access, prompting training, and above all, an internal culture that treats AI as a work infrastructure.

Empowering industry professionals means investing in tools and real training.

The token issue, in short, isn’t technical. It’s political and social in the broadest sense of the word.

It concerns who can afford, and is even capable of, thinking with the help of machines and who cannot, who gets on the cognitive automation curve and who remains waiting with the concrete risk of not finding space in the job market, of being left far behind, of not being up to speed.

I wonder how many of you don’t know how to use an AI service, and have never used prompting.
I wonder how many of you have already used up your daily dose of tokens.

Mondoduepuntozero