From the AI bubble to the decline of organic traffic, Andrea shares his take on the contradictions of an innovation that risks emptying the web and weakening critical thinking
What remains of the open web, now that artificial intelligence is changing the way we search, read and even think? That is the question at the heart of our new SEO Confidential interview, with an exceptional guest: Andrea Daniele Signorelli, a journalist who has spent years investigating the relationship between technology, politics and society, writing for Wired and La Repubblica.
With him we talked about the artificial intelligence bubble growing at a dizzying pace, about a web increasingly populated by machines rather than people, and about a journalism that risks mistaking efficiency for quality.
Signorelli takes us inside the rhetoric of “inevitable progress” to show us what really lies behind it: enormous capital, technological dependencies and a future that, if left ungoverned, could push us away from knowledge instead of bringing us closer to it.
A lucid and stimulating conversation that challenges blind faith in automation and invites us to rethink the human role in the age of machines.
“The AIs that feed on the web today risk draining it”: Andrea D. Signorelli’s take
You have written that the very companies driving the development of artificial intelligence are fueling an economic bubble of gigantic proportions, even while being aware of the risks it entails.
In your view, does this “financial hypertrophy” stem more from an almost ideological faith in technological progress, or from the logic of platform capitalism, which demands investment even in the absence of real returns?
And if the bubble really were to burst, what could the concrete effects be on the tech sector and the global economy?
This is not the first time we have witnessed a phenomenon of this kind: colossal investments that go through a phase of financial euphoria, sometimes turning into a bubble, and only later prove their real value. It happened with the dot-com bubble at the end of the 1990s: after the crash, the technologies behind it – the internet and the web – kept evolving until they generated enormous returns for investors and transformed the world economy.
For this reason, the fact that a bubble is forming around artificial intelligence today does not mean the technology is doomed to fail. It is plausible, however, that the current investment phase will not manage to pay back, at least in the short term, the enormous sums being committed.
There are two main reasons. The first concerns costs. Unlike previous digital platforms, generative artificial intelligence models require gigantic economic resources, not only for training but also for maintenance and everyday use. We are talking about astronomical figures: even OpenAI, despite the success of ChatGPT, is reportedly operating at a loss, with estimates pointing to a negative balance of around 27 billion dollars. No company can sustain that level of spending for long, which is why the sector’s big players must periodically go back to investors asking for fresh capital.
The second reason is more ideological in nature and concerns the narrative sustaining this boom. Big tech keeps presenting artificial intelligence as a disruptive technology, to the point of making people believe that so-called AGI is just around the corner – artificial general intelligence capable of matching or surpassing humans. This promise works as a lever to attract capital, because it implies that, when and if AGI becomes reality, today’s investments will be amply repaid. Yet it is very likely that this prospect remains closer to science fiction than to science.
In essence, we are facing an intertwining of speculation and ideology. On one hand, the financial bubble is a recurring stage in technological capitalism: every new wave of innovation requires enormous capital and often goes through a phase of excess before stabilizing. On the other hand, however, there is no guarantee that all these investments will translate into concrete returns.
So yes, the bubble is there, and it is fueled by a combination of technological faith and the systemic pressure of platform capitalism, which demands constant investment, even in the absence of immediate returns. It is a dynamic typical of technological capitalism, where innovation requires enormous capital and often passes through speculative phases.
If the bubble were to burst, the effects would be severe. We could see a sharp market correction, a drastic reduction in investment and a consolidation of the sector around a few solid players. Nvidia, for example, is currently the only company generating real profits from the AI boom, but it is also the most exposed: if the big buyers stopped purchasing GPUs, its value could plummet, dragging the entire tech sector down with it.
That said, the collapse of the bubble would not mark the end of artificial intelligence. Just as happened with the internet after 2000, the technology would keep developing on more sustainable foundations, with more concrete applications less dependent on the rhetoric of “inevitable progress”.
Do you think that artificial intelligence, if used passively or excessively, risks impoverishing us cognitively in the long run, making us less curious, less creative and generally less intelligent?
What could be, in your opinion, the concrete strategies or habits for developing a “smarter” use of ChatGPT, one that stimulates rather than atrophies critical thinking?
Yes, I think an excessive or passive use of artificial intelligence can indeed have negative cognitive consequences. Some MIT studies have already demonstrated it: people who write texts using ChatGPT show lower brain activation. But if you look closely, it is not a surprising finding. In many cases, users who were asked to write an essay – on the French Revolution, for example – simply typed a prompt and then let ChatGPT do all the work. It is only natural that, in situations like these, mental activity drops to a minimum.
I believe, however, that the issue is not the tool itself, but how it is used. Used the right way, ChatGPT can become a genuine tool for cognitive enhancement. We have critical thinking, the ability to define goals, to weigh pros and cons, to make sense of information. Artificial intelligence, on the other hand, can process enormous amounts of data in very little time. If you combine these two dimensions, the results can be extraordinary.
The problem arises when we delegate everything. If we let AI think in our place, we inevitably end up atrophying our critical faculties. But if we use it to expand our possibilities, to do research, to explore new ideas, then our cognitive level is not only maintained but can even grow.
I think we will see what we have already seen with other technologies: a minority of people capable of exploiting the full potential of artificial intelligence will emerge, alongside a majority who will use it passively. It happened with the web, then with social media, and now it will probably happen with ChatGPT too. The potential is enormous, but it is not enough on its own: we need a digital and cultural literacy that allows people to understand how to use artificial intelligence to think more, not less.
You have written that artificial intelligence is not inherently an enemy of journalism, but that the real problem is the way it is used to cut costs and replace human work.
Do you think this trend is still reversible, or is the economic model of the big publishing groups now too dependent on the logic of automation and quantity at the expense of quality?
AI becomes a problem for journalism when it is used to cut costs, reduce staff and replace human work instead of supporting it.
The episode of the journalist at La Provincia who forgot to remove the typical ChatGPT-generated line “Would you like me to turn this into an article for a daily newspaper or a more narrative version for an investigative magazine?” is a perfect example.
That slip is not just the result of individual carelessness; it stems from an increasingly frantic working environment, with newsrooms cut to the bone, low salaries and unsustainable workloads. Today two or three editors do what ten used to do, and it is inevitable that, under conditions like these, people end up using AI as a shortcut.
The problem, therefore, is not the tool itself, but the conditions in which it is deployed. If artificial intelligence serves only to speed up production and cut costs, the quality of information collapses. If instead it is used to support journalists, to analyze data, to simplify research or writing, it can become a powerful ally.
To answer your question: yes, I think this trend is still reversible. Or rather, I think that precisely the abuse of automation and the resulting loss of quality could, paradoxically, open up new space for a “human” journalism, one that is more accurate and more aware. If the saturation of “avoidable” content becomes unbearable, the public could go back to seeking quality information. It is perhaps a slightly optimistic outlook, but a realistic one: when you stretch the rope too far, sooner or later it snaps, and from that break a change can emerge.
The world of social networks offers an example too: for years we witnessed a compulsive production of low-quality content. Now, however, demand is growing for more curated, slower and more reliable forms of information. It is possible that something similar will happen in journalism: that the public will start rewarding expertise, verification and critical thinking.
Sam Altman has said that if an artificial intelligence can replace your job, maybe that job was not “real work”. It is a brutal statement, of course, but does it contain a grain of truth? Can it be a stimulus to improve, to differentiate, to develop creative skills that machines cannot replicate?
The replacement of human work by AI does not happen because those tasks are truly superfluous, but because companies want to slash costs and maximize profits. In reality, the best results come when technology works alongside people, not when it replaces them. If newsrooms invested in training journalists to use these tools intelligently, the quality of the work would improve instead of deteriorating.
Some also argue that a future of aggressive automation should go hand in hand with social protection tools, such as universal basic income. It is no coincidence that many Silicon Valley leaders promote it (Elon Musk first among them)…
They do not do it out of altruism, but out of pure calculation: they know full well that large-scale automation risks generating unemployment and social tension, and they see universal basic income as a kind of preventive insurance against a possible “social revolt of the replaced”.
It is a rather cynical vision, because in practice it serves to protect the economic system more than the people. But it says a lot about the world we are building: a world in which we accept the idea that millions of jobs can disappear, as long as a minimum level of support is guaranteed to those left behind.
Do the tests used to evaluate the intelligence of AIs, the so-called benchmarks, really help us understand how “intelligent” a model is? Or do they risk becoming merely a race for the highest score, saying nothing about the real ability to reason or understand what it is doing?
In truth, benchmarks are useful only up to a point. Yes, they can measure certain capabilities of the models, but not their “intelligence”, simply because intelligence, in the human sense of the word, is not there. What they measure is performance: a model’s ability to pass a test or achieve the highest score, a bit like a math competition.
The problem is that by now many of these tests have become exactly that: a competition for the best score, saying little about the model’s real capacity for reasoning or understanding. Often, in fact, the best results have turned out to come from tricks rather than genuine “intelligence”.
There have been cases in which models achieved extremely high scores on some well-known benchmarks, only for it to emerge that they were finding the answers by looking them up online, in the datasets they had been trained on. In practice, they were copying solutions already available on the web. It is as if, in a math competition, someone Googled the result of the exercise instead of doing the calculations.
This does not mean benchmarks are useless, but that they should be rethought. We need more transparent tests, with clear and shared parameters, that truly measure the ability to reason and not just the ability to reproduce answers already seen. Also because many benchmarks in use today are based on information contained in the models’ training datasets: if ChatGPT gets a very high score on a general-knowledge test, it does not mean it “knows” something, only that it has already read that information.
In short, benchmarks remain useful tools, but they should be reformulated to genuinely understand how and how much these systems comprehend what they are doing, instead of merely ranking who reaches the highest score.
With the advent of AI-powered search engines, online search seems to be transforming from exploration into the mere consumption of ready-made answers.
Do you think this evolution risks reducing our ability to critically engage with sources, or could it instead lead us toward a new – more conscious – way of searching for and interpreting information?
Yes, this transformation of online search is worrying, for at least two reasons. The first concerns the loss of autonomy. Until recently, even though Google influenced the way we accessed information, a certain freedom of choice remained. Faced with the results, we could decide which links to trust, prefer one publication over another, choose based on our own sensibility or trust in the source. Today, instead, with AI-powered search engines, that plurality disappears: we are given a single, ready-made answer, formulated in natural language.
And, predictably, almost no one goes to check the original sources anymore. Where there used to be selection, there is now total delegation. This concentrates enormous informational power in the hands of very few players, who become the new “filters” of knowledge. It is a radical change, because it risks reducing our critical relationship with information and turning us from active readers into passive consumers of answers.
The second problem concerns the sustainability of the entire open web ecosystem. Until now an equilibrium existed: sites produced content that Google indexed, and in return they received traffic and visibility. With answers generated directly by AI models, this pact breaks down: Google and the other companies extract value from other people’s content without giving anything back. It is an enormous risk, because if sites stop receiving visits, many will no longer have the motivation or the resources to produce quality information.
Some agreements are starting to appear, such as the one between OpenAI and the New York Times, or other deals between big publishers and platforms, but the problem remains for small sites, independent blogs, people who write out of passion or for modest economic returns. If their work is no longer seen, they will end up giving up publishing. And without them, even the big artificial intelligence systems will have less and less up-to-date, reliable material to draw on. It is an evident paradox: the AIs that “feed” on the web today risk draining it.
That said, some possible solutions exist. Some companies are experimenting with systems to prevent content scraping, or blockchain-based automatic compensation models that could grant a micro-payment every time a text is used by an AI model. These are still early-stage ideas, but they go in the right direction: restoring a balance between those who produce information and those who use it to train systems that, in fact, derive economic value from that information.
Among the new AI-based search engines – from ChatGPT Search to Perplexity, all the way to Komo and Brave – which do you think currently offer the most reliable and useful search experience for those seeking quality information, and by what criteria should this “quality” be assessed in the era of AI-generated answers?
It is hard to give a single answer, because the search experience with these new tools is very subjective: it depends on what you are looking for and how you are used to using the web. That said, among the various AI-based search engines, Perplexity is probably the one that has shown the best results so far. It is designed specifically for search, and it has a more transparent approach, because it shows its sources, cites the sites it draws information from and lets you dig deeper with a click.
Still, Google’s advantage remains enormous. The moment it integrates artificial intelligence into its own results, as it is already doing, it will be very difficult for Perplexity and the others to carve out a space. Google starts from a position of total dominance and has infrastructure and resources no competitor can match. That is why many of these “alternative” engines are likely to remain niche products or be absorbed by larger companies.
Then there is another issue: beyond the name or the interface, the differences between one model and another are not that deep. The tone changes, the marginal features change, but the underlying logic is the same: answering the user in natural language, minimizing the passage through the sources. That is also why, in the long run, it does not make much sense for dozens of tools doing the same thing to exist. Many of these projects are not economically sustainable, because the market of paying users is still tiny.
Perhaps some of these startups will survive precisely because they are financed by the big tech groups, which prefer to keep them alive in order to observe their innovations up close or to push Google to accelerate its own developments. It is a way to keep the sector buzzing, but also to test ideas that, if they work, will be swallowed up by the giants.
As for the “quality” of search, I believe it should be measured against three criteria: transparency of sources, the ability to contextualize information, and the user’s ability to verify what they read. If an AI-based search engine can guarantee these three elements, then it can genuinely improve the way we search for and understand information.
Mistral AI is often described as the only real European answer to American and Chinese dominance in artificial intelligence. But in your view, can Europe truly build its own technological sovereignty by betting on a few symbolic projects, or would a more ambitious, structural common plan be needed to avoid remaining dependent on external powers?
Mistral AI is an important signal, but also a somewhat symbolic one. It is considered the “European answer” to the American and Chinese giants, but in reality it is not entirely European: it has external investors and ties, and this clearly exposes Europe’s structural limits in the technological field. The fact that there is practically only one player of this scale on the continent, and that it is not even fully under European control, says a lot about the delay we have accumulated.
The problem is that we cannot expect to build real technological sovereignty by relying on a single company or a few symbolic projects. We would need a common, ambitious, long-term strategy that brings together infrastructure, research, universities and industry. So far, instead, we have almost grown accustomed to considering our dependence on the United States as a given. But it is not, especially today, when relations between Europe and the United States are no longer as stable as they once were.
This is not catastrophism, it is a matter of fact: if he wanted to, Donald Trump could “pull the plug” on a large part of our digital systems. Many analysts say so, and it is a real risk.
Precisely for this reason, we need serious investment, not symbolic gestures. It is a long and costly process, but an inevitable one. The first step should be to build European technological infrastructure, starting with data centers: as long as our data is hosted and managed elsewhere, we cannot speak of true autonomy. From there we need to develop a broader ecosystem – made of skills, hardware and, above all, a shared vision – that allows Europe to stop chasing and start building.
Proton’s Lumo also represents a concrete attempt to build a European AI founded on privacy and digital sovereignty, but with performance still far from the American models.
In your opinion, is it realistic to think that Europe can truly compete on the technological front without giving up its ethical principles? Or does the protection of data and privacy risk limiting European innovation?
There is often a tendency to set two things against each other: on one side the performance of artificial intelligence models, on the other the respect for data and privacy. But I do not believe this “either/or” really exists. It is not a given, in other words, that a model must sacrifice efficiency to be ethical or transparent.
In Lumo’s case, the lower performance has nothing to do with its privacy-protection principles: it simply comes down to the fact that Proton does not have the economic and infrastructural resources of the big American giants. It is a question of means, not philosophy.
That said, Lumo is still an important signal. It shows that it is possible to try to build a European artificial intelligence founded on different values, such as digital sovereignty and data protection. If the European Union decided to invest seriously – with public funds, its own infrastructure and European-owned data centers – it would truly be possible to develop a large language model that is competitive and at the same time privacy-respecting.
It would not be enough to free us from our dependence on the United States right away, but it would be a concrete first step. Mistral proves it: it is not yet at ChatGPT’s level and it is not entirely European, but it shows that the gap can be closed. All that is needed is a solid foundation of investment, infrastructure and a common strategy.
In your Wired article you talk about the risk that AI Mode will transform the web into a “machine web”, populated more by artificial intelligences than by users. Do you see real alternatives to this AI-based search model? Are there, in your view, technological or regulatory paths that could prevent the drift toward the “machine web”?
I do not believe there are, at the moment, real alternatives to this process. Even from a regulatory point of view, it is hard to imagine a law that could truly stop the evolution toward a web dominated by generative systems. One can theorize regulation that protects the diversity of sources or limits algorithmic intermediation, but in practice I see no tools capable of reversing this trend.
I think the path is now irreversible. Unless they collapse under the weight of their enormous maintenance costs, generative systems will increasingly become the main interface through which we experience everything found on the internet. The web will continue to exist, but as a hidden layer, a kind of digital infrastructure feeding these artificial intelligence models without being visible to users anymore.
This does not mean direct consumption will disappear entirely: there will always be a minority of people who prefer to read at the source, consult newspapers, visit sites and dig deeper on their own. But for most users, the internet experience will pass through generative filters, just as today, for many, the web coincides with Instagram or TikTok.
Social networks have already transformed the internet into a mediated space, where most people no longer browse but scroll through prepackaged content. AI-based search engines push this logic even further: they will be the ones selecting, synthesizing and returning what they consider relevant.
The open web will not disappear, but it will change in nature. It will increasingly become a “substrate” feeding generative systems, populated largely by content produced for machines rather than for people. It is a transformation already underway, and unless different economic or technological models step in, the risk is that the human web – the web of discovery, curiosity and plurality – will become just a niche.
You point out that AI Mode could undermine the economic model that has so far sustained websites and news outlets: do you think Google will be forced to find a compensation system for content producers, or is the value of organic traffic destined to disappear for good?
Yes, I believe Google will be forced, sooner or later, to find a form of compensation for content producers. The value of organic traffic is now extremely low and, in many cases, already in sharp decline: several major publications report a reduction of between 35 and 40%, and we are only at the beginning. People are still used to using Google the classic way, clicking on links, but it is clear that with the expansion of AI Overviews this dynamic will change quickly, perhaps within a year or two.
That said, Google and the other big players cannot afford to completely drain the ecosystem they draw value from. Content published on the web is needed so artificial intelligences can update and improve; without new quality sources, even the models themselves would end up impoverished. For this reason, some form of compensation will be inevitable, though it will not necessarily be a fair or stable solution. It could be a temporary measure, a “patch”, rather than a true rebalancing of the system.
The problem is that today only the big publishing groups have the strength to negotiate direct agreements. Everyone else – small sites, local newsrooms, independent projects – is left out. Collective action would be needed, something resembling a “union of the web”, capable of representing those who produce informational value without the economic muscle to negotiate with Google.
For its part, Google keeps declaring that the “health of the open web” remains a priority, but it is clearly doing exactly the opposite: draining content without returning value. And it is perfectly aware of it.
The real risk is not automation, but abandoning critical thinking online
After talking with Andrea Daniele Signorelli, one thing is clear: we cannot remain spectators. Artificial intelligence is not just a technological matter, but a cultural, economic and political challenge that affects us all.
It is changing the way we inform ourselves, think and even imagine the world.
The future of the web – and perhaps of knowledge itself – will depend on how we choose to use these technologies: to simplify or to understand, to replace or to enhance, to delegate or to think more.
And if the risk really is a “machine web”, then the answer will have to be human: more curiosity, more awareness, more critical thinking.
The point is not to stop AI. Also because I do not see how its impetuous growth could possibly be halted…
As Battisti sang, no rock can hold back the sea!
No, I am convinced that we must make it transparent and useful without losing the complexity of human thought along the way.
If the web becomes a place inhabited more by algorithms than by people, it will be because we have stopped truly inhabiting it. It is up to us to decide whether to make it a space of automation or of shared intelligence.
Our thanks to Andrea Daniele Signorelli for this enjoyable conversation, which reminded us that the future is not written in the code of machines, but in the choices of those who use them.
And thank you as well, for being with us again this week – see you at the next episode of SEO Confidential.
