When machines improve themselves: Who controls our AI future?

From hype trap to real value creation: How AI is truly changing the world of work

AI builds AI: The uncanny leap in productivity in secret labs

Artificial intelligence has definitively arrived at the heart of global economic and political discourse. The days when public debate revolved solely around hallucinating chatbots or the looming threat of job losses are over. Today, a fundamental shift is at stake: Who determines the pace of technological development? While the US is pumping unprecedented billions into scaling and is already helping cutting-edge AIs develop their own successors, Europe is searching for the delicate balance between necessary regulation—for example, in child protection—and maintaining its own competitiveness. But even within companies themselves, the initial hype is giving way to a harsh reality: AI may reduce production costs, but it is far from guaranteeing added value. The following article examines the crucial aspects of this structural transformation and demonstrates why humans must retain control in the future—not merely as correctors, but as the strategic architects of the entire system.

Europe regulates, America scales – and the machines begin to improve themselves

The societal debate on artificial intelligence is entering a new phase. Until now, the focus has primarily been on the capabilities of generative models, their susceptibility to errors, and the question of which tasks could be automated. Now, the perspective is shifting. The crucial question is no longer just what AI can achieve, but who controls its development, who reaps the economic benefits, who bears the risks, and which political regulations determine where innovation takes place.

Three recent developments illustrate this transition. In marketing, the realization is gaining ground that while AI accelerates production and analysis, it cannot replace a viable brand strategy. In European politics, the protection of minors from manipulative platform mechanisms is becoming increasingly important, while at the same time there is growing concern that additional regulations could burden smaller European providers more than global technology companies. And in leading AI labs, models are already automating parts of the research and development work that will give rise to the next generation of models.

These developments are economically interconnected. They describe three levels of the same power shift: the application of AI in companies, the government's shaping of digital markets, and the automation of AI production itself. Those who focus solely on individual tools or laws underestimate the structural change. AI is becoming a foundational technology that simultaneously transforms productivity, competition, capital requirements, work organization, and regulatory sovereignty.

From AI hype to the question of value creation

The debate in marketing and advertising is a good starting point because tensions become apparent there particularly early on. Generative AI can produce texts, images, videos, variations, target group approaches, and analyses at a speed that is hardly achievable with traditional processes. The economic benefit lies initially in decreasing marginal costs: once a system is set up, the hundredth variation of an ad costs only a fraction of the first. This logic favors scaling, personalization, and faster testing.

However, lower production costs do not automatically generate higher added value. If all market participants use similar models, data sources, and optimization logics, the quantity of available content increases faster than its quality. The result can be an inflation of interchangeable communication. Technical efficiency increases, while customer attention becomes scarcer and more expensive. Companies then save on the creation of individual pieces of content but have to invest more just to be noticed at all.

This is precisely the central AI paradox of marketing. Automation increases operational efficiency, but can simultaneously weaken strategic differentiation. A model can apply brand guidelines, mimic tones, and analyze existing data. However, it does not decide, based on its own economic responsibility, which market position a company should defend in the long term, which customer group is deliberately excluded, or which promises must be kept even under cost pressure.

The statement that humans guide AI is therefore less a reassuring formula than a demanding management task. Human control doesn't mean manually editing every result. It means defining goals, boundaries, quality criteria, and responsibilities in such a way that automation generates not just more output, but better business results. The more powerful the systems become, the less sufficient it is to simply optimize individual prompts. Robust operating models for data, approvals, brand management, measurement, and liability are required.

Brand value is not created in the model

In economically uncertain times, trust, quality, and reliability gain in importance. For brands, this is a crucial counter-movement to the pressure of technological acceleration. Customers don't buy solely based on a perfectly crafted advertising message. They evaluate whether product performance, service, price, delivery capability, and communication create a consistent overall picture. AI can efficiently convey this picture, but it cannot replace the underlying performance.

Authenticity should not be understood romantically. It's not about rejecting machine-generated content outright or making all communication appear visibly handcrafted. Rather, what's economically relevant is the alignment between claimed and actually delivered benefits. A company appears authentic when its messages are verifiable, its decisions align with its stated values, and its customers have comparable experiences at all touchpoints.

Generative AI increases the risk that companies will overpromise in their communications and deliver more than their organizations can realistically provide. Images can make products appear higher quality, supply chains more sustainable, and services more personal than they actually are. In the short term, this can boost awareness and conversions. In the long term, however, reputational damage, returns, and customer retention costs increase. The easier it becomes to create perfect presentations, the more valuable verifiable evidence, real customer experiences, and consistent performance become.

This results in a changed role for the marketing department. It must not only produce content but also manage the company's credibility. This includes verifiable product data, clear origin information, solid performance promises, coordinated approval processes, and measurement that goes beyond clicks or reach. The decisive productivity gains arise when AI improves the connection between market research, customer knowledge, creation, sales, and service. If it is merely used as a cheap content engine, its benefits remain limited and the risk of interchangeability is high.

Human control becomes the operating system

The idea that humans must individually review every AI result is not scalable as usage increases. Large organizations generate thousands of texts, images, recommendations, forecasts, and automated decisions daily. Purely manual review would negate any efficiency gains and still overlook many errors. Therefore, human oversight must evolve from individual review to system design.

A sound control model begins with risk classification. An internal summary or a non-binding list of ideas requires less control than a public product claim, a personalized pricing decision, or communication to minors. Companies should therefore not make a blanket distinction between human and automated processes, but rather consider potential harm, reach, reversibility, and legal implications.

Based on this, tiered approval processes can be established. Low-risk applications can run largely automatically if data sources and quality thresholds are defined. Medium-risk applications require sampling, documented responsibilities, and clear escalation paths. High-risk applications still require binding human approvals, independent audits, and auditable protocols. Crucially, humans must not only tick a box at the end but also define the objective and metrics.

This also changes the required qualifications of employees. Effectively managing AI requires expertise, data literacy, critical judgment, and the ability to assess business consequences. Purely operational skills are becoming less valuable as user interfaces become simpler and models more powerful. What remains valuable is the ability to select relevant problems, place results in an economic context, and recognize inconsistencies. Humans do not automatically retain leadership. They only retain it if organizations institutionally secure their role and strategically develop their competencies.

Efficiency gains without the illusion of productivity

The use of AI is often justified by the time savings it offers. While these savings are real, they shouldn't be confused with overall economic productivity. If a task takes only one hour instead of four, this initially creates a technological advantage. Economic value only arises when the three hours freed up are actually used for higher-value work, when the same output is achieved with fewer resources, or when additional demand can be profitably met.

In many companies, this second step remains unclear. Employees produce more variations, analyses, and presentations without making decisions faster or better. The organization receives more information, but not necessarily more guidance. This can lead to new costs for review, coordination, and administration. AI then reduces the costs of individual work steps, but increases the volume of work and thus the coordination effort.

For a realistic cost-benefit analysis, all costs must therefore be considered. These include licenses, computing power, integration, data preparation, security, training, monitoring, potential errors, and dependence on external providers. Opportunity costs also arise when teams invest significant time in pilot projects that are never transitioned into stable processes. The relevant metric is not the number of tools implemented, but rather the change in throughput time, error rate, revenue, customer retention, or total costs.

The distribution of profits is particularly important. Automation can relieve the burden on employees, increase margins, or lower prices. The actual effect depends on competition, bargaining power, and corporate strategy. In concentrated markets, platform providers can capture a large portion of productivity gains through pricing and dependencies. In highly competitive markets, savings are more likely to be passed on to customers. For Europe, therefore, it is crucial not only to implement AI but also to gain economic control over a larger share of the underlying infrastructure, models, and platforms.

Child protection meets industrial policy

The debate surrounding a European Kids Act demonstrates the close connection between societal goals and location policy. Age limits, parental controls, daily usage restrictions, and bans on manipulative design elements address real problems. Endless scrolling, autoplay, reward loops, and highly personalized recommendations are not neutral product features. They are tools for capturing attention and thus part of a business model that translates user time into advertising revenue, data, and market power.

The logic of protection is particularly plausible when it comes to minors. Depending on their stage of development, children and adolescents do not yet possess the same ability as adults to assess persuasive mechanisms, emotional appeals, or long-term data privacy consequences. AI chatbots exacerbate the problem because they not only display content but can also engage in dialogue, build trust, and obtain personal information. Digital companions can be useful, but they also carry the risks of emotional dependency, manipulative advice, and inappropriate interactions.

The economic policy challenge begins with implementation. Age verification, separate youth accounts, functional restrictions, usage time controls, and security checks all incur fixed costs. Large platforms can spread these costs across hundreds of millions of users. Small European providers have significantly fewer economies of scale. A rule that is formally the same for everyone can therefore have unequal economic effects and even increase market concentration.

This creates a classic regulatory dilemma. Insufficient requirements can cause societal harm to families, schools, healthcare systems, and those affected. Overly complex requirements can stifle innovation, prevent market entry, and strengthen the very corporations whose behavior the regulation is intended to limit. The real question, therefore, is not whether to regulate, but how to simultaneously balance protective effect, feasibility, and competitive dynamics.

Duty of care without data traps

Age verification sounds simple, but it is technically and legally complex. A platform must reliably distinguish whether a user is below or above a certain age limit. At the same time, the verification process must not require everyone to submit identification documents, biometric data, or comprehensive identity information for ordinary online services. A poorly designed protection system could, in fact, create new data sets that are attractive for surveillance, profiling, or identity theft.

Economically, this is a question of proportionality. The accuracy of age verification can usually only be increased by accessing additional data. However, greater accuracy leads to higher costs and greater data privacy risks. Conversely, particularly data-efficient methods may be less precise and could falsely admit minors or exclude adults. A robust system must therefore use different levels of verification depending on the risk.

For lower risks, age-appropriate default settings, restricted functions, and local technical signals may suffice. For higher risks, stronger verification methods are acceptable, provided that only the necessary age information is transmitted and not the complete identity. Crucially, the architecture should ensure that the service provider only learns whether an age limit has been met. Such data-minimizing verification methods should be interoperable so that each platform does not have to build its own identity system.

Enforcement deserves more attention than the formulation of new obligations. When authorities enact numerous detailed rules but lack the necessary personnel, technical expertise, and swift procedures, an asymmetric burden arises. Smaller, compliant companies bear high implementation costs, while large players can drag out proceedings for years. Good regulation therefore requires clear standards, a shared technical infrastructure, realistic transition periods, sandboxes for smaller providers, and sanctions that are swift enough to prevent distortions of competition.

Industry focus areas: B2B, digitalization (from AI to XR), mechanical engineering, logistics, renewable energies and industry

More information here:

• Expert Business Hub

A thematic hub offering insights and expertise:

• Knowledge platform covering global and regional economies, innovation and industry-specific trends
• A collection of analyses, insights, and background information from our key areas of focus
• A place for expertise and information on current developments in business and technology
• A hub for companies seeking information on markets, digitalization, and industry innovations

Europe's AI gap is primarily a scaling gap

The concern about a transatlantic AI divide is not unfounded. The United States is home to vast amounts of private capital, leading cloud platforms, semiconductor expertise, research talent, and global distribution channels. By 2025, private AI investments in the US had reached approximately $285.9 billion, compared to roughly $20.9 billion for Europe. The concentration was even more pronounced in generative AI: the US accounted for approximately $163.6 billion, while Europe contributed around $3.2 billion.

These figures are not directly comparable to public European investment announcements. Private equity financing, data center investments, subsidy programs, and mobilized capital measure different economic processes. Nevertheless, they demonstrate that American companies can achieve very large financing rounds much more easily. This is particularly crucial for basic models, data centers, and chips, because high initial investments, short innovation cycles, and uncertain returns converge in these sectors.

Europe certainly possesses research, industrial data, skilled professionals, strong user industries, and a large domestic market. The deficit often arises during the transition from prototype to international scaling. Young companies find early funding but reach their limits when dealing with hundreds of millions of euros, global distribution, and long-term computing capacity. As a result, they become dependent on non-European capital, relocate operations, or are acquired.

Regulation is therefore only part of the problem. Even a complete dismantling of European AI regulations would not close the gap in capital markets, cloud infrastructure, energy supply, procurement, and scaling. Conversely, it would be wrong to ignore regulatory costs. If approvals, legal interpretation, and national implementations are slow or contradictory, they exacerbate existing disadvantages. Europe thus does not need blanket deregulation, but rather a combination of clear rules, rapid application, common infrastructure, and decisive demand-side policies.

When AI builds on AI

The developments at Anthropic mark a new economic order of magnitude. According to the company, in August 2026, Claude carried out the majority of a task, from initial instruction to final result, in 26 percent of the measured research and development work, while humans retained oversight. In February, this figure was still below one percent. In more than 90 percent of the recorded work, AI was involved at least as a close collaborator.

These figures do not mean that Claude autonomously designs, trains, and releases a successor. Anthropic categorizes the activities on a scale. At the collaborative level, AI completes larger work packages under close human guidance. At the leadership level, it handles the majority of a task independently but is still supervised and does not make the final decision alone. Complete autonomy without human intervention was not observed in any of the measured areas.

Despite this limitation, the dynamics are economically significant. When AI writes research code, prepares experiments, analyzes errors, generates reports, and solves technical problems, the time between hypothesis and result decreases. Researchers can conduct more experiments in parallel and work with a larger search space. This not only increases the productivity of existing teams but can also accelerate the pace of innovation itself.

This creates a feedback loop. Better models improve research into even better models. These, in turn, automate a larger proportion of the development work. As long as humans set goals, review results, and grant approvals, this is not autonomous, recursive self-improvement. However, even partial feedback can have significant economic consequences because development cycles shorten and the lead of leading laboratories increases.

The cumulative advantage of top laboratories

AI-driven AI research favors companies that already possess robust models, computing power, data, talent, and internal development platforms. A leading lab leverages its own models to develop new ones more quickly, thereby improving the very tool it uses for research. This cumulative advantage resembles a learning curve, but is potentially steeper because the means of production itself becomes more intelligent.

For competition, this means that a gap that exists today may not only persist tomorrow but could widen rapidly. A smaller provider may be able to read the same scientific publications or use open models. However, they may lack the internal data on millions of development decisions, the specialized infrastructure, and the financial reserves for large-scale experiments. Furthermore, large laboratories benefit from a cycle of better products, more users, higher revenues, greater computing power, and faster research.

At the same time, concentration is not inevitable. Decreasing inference costs, powerful open models, and specialized applications can open up new markets. Many companies don't need to train their own base model to create economic value. They can combine industry-specific data, process knowledge, customer access, and regulatory expertise. Europe has a particularly strong starting position in this regard, especially in industry, logistics, healthcare, energy, and public administration.

The strategic danger, therefore, lies less in Europe needing to copy every American basic model. The problem would be complete dependence at all levels: chips, cloud computing, models, agent platforms, and application ecosystems. Sovereignty does not mean autarky. It means maintaining critical choices, being able to switch providers, enforcing one's own standards, and building competitive alternatives in selected areas.

Transparency must not be self-promotion

Anthropic's Automation Index provides unusually concrete data. The approach analyzes approximately 15,000 granularly described tasks, categorizes them hierarchically, and weights them according to the time spent on each task. The degree of automation is then assessed for each category. This offers more insight than general statements like "AI increases productivity" or "AI supports research.".

However, the methodology has limitations. Part of the data collection and evaluation is again carried out using Claude. The model and human evaluators do not agree in every borderline case, particularly when distinguishing between close collaboration and largely independent management. Furthermore, the task basket is based on a specific time period. If automation creates new tasks, people take on more demanding responsibilities, or the research organization changes, a fixed index can only partially reflect these shifts.

This metric also doesn't measure employment impact. A 26 percent share of AI-driven tasks doesn't mean that 26 percent of employees have been replaced, nor that 26 percent of all research costs have been eliminated. Humans define problems, monitor processes, evaluate results, and take responsibility. At the same time, automated tasks can generate additional demand for experiments, infrastructure, monitoring, and safety research.

Nevertheless, publication is an important step. Governments and the public need key performance indicators (KPIs) that not only demonstrate modeling capabilities but also reveal the production process of advanced AI. These include the proportion of AI-driven research, the scope of agent-based systems, monitoring coverage, the number of blocked actions, the response time to anomalies, and the allocation of resources to security. In the long term, such data should be collected according to common standards and independently verified. Otherwise, transparency will remain dependent on the voluntary self-disclosure of individual companies.

Security as a production factor

Anthropic reported that at times, approximately 30,000 agents were simultaneously performing research and development tasks on the main internal platform. More than one billion decisions were processed through automated controls in a single month. About 0.002 percent of actions were blocked, which corresponds to roughly one in 47,000 decisions. This low rate can be interpreted as an indication of rare anomalies, but it also highlights the scaling problem: With billions of operations, even very rare events become practically significant.

Security, therefore, cannot be understood as a retrospective check. It must be built into the technical architecture. Real-time controls should stop irreversible or immediately dangerous actions. Subsequent analyses can detect slowly emerging patterns. Identities for individual agents, traceable communication channels, limited permissions, and complete logs facilitate the attribution of actions.

From an economic perspective, such controls are not simply a burden. They reduce anticipated damage costs, increase the reliability of automated processes, and ultimately make further delegation possible. A company will only delegate critical research, financial decisions, or industrial control to agents if errors can be detected, interventions triggered, and responsibilities clarified. Security thus becomes a prerequisite for scaling.

At the same time, there is a conflict of objectives regarding resources. In one week studied, according to company data, around six percent of the computing power used for AI research was allocated to security work; for AI-driven AI research, it was about twelve percent. Such figures are difficult to compare because security and capability research are intertwined, and computing power does not reflect all personnel costs. Nevertheless, this metric opens up an important debate: Those who increase the pace of development must also demonstrate that monitoring, interpretability, testing, and responsiveness can keep pace.

Work doesn't disappear, it's redistributed

The growing involvement of AI in demanding research contradicts the simplistic notion that automation only concerns routine tasks. Tasks that are digitally available, produce quickly verifiable results, and offer extensive training data are particularly well-suited for automation. Programming, data analysis, technical documentation, and experimental design partially meet these criteria. In contrast, many physically, socially, or responsibility-intensive tasks remain more difficult to automate.

For the labor market, this doesn't mean a linear replacement of entire professions. Task bundles are broken down and reassembled. A researcher, developer, or marketing expert performs fewer initial drafts and standard analyses but spends more time on problem selection, monitoring, integration, and decision-making. In successful organizations, output per employee increases. In less well-managed organizations, only the quantity of mediocre results grows.

The pressure to adapt will be unevenly distributed. Top performers can use AI as a multiplier and take on greater responsibilities. Entry-level professionals may lose tasks where they previously gained fundamental experience. If simple analyses, initial code drafts, or standard texts are largely automated, companies will have to create new learning paths. Otherwise, a shortage of experienced specialists will arise in the long term because traditional entry-level positions will disappear.

Income distribution is also an open question. If owners of models, data, and computing infrastructure receive the lion's share of the profits, AI can exacerbate the concentration of income and wealth. However, if employees share in productivity gains, further training is funded, and new businesses are promoted, the technology can have a broader impact. Labor market policy should therefore not only compensate for potential job losses but also shape transitions, startups, competition, and the distribution of productivity gains.

The right European answer

Europe should not treat child protection, data privacy, and economic competitiveness as mutually exclusive. Good rules can build trust and foster a market for safe, verifiable AI. However, this will only succeed if regulations are technically feasible, consistent across Europe, and manageable for smaller providers. Therefore, every new obligation should be subject to a competition impact assessment: What fixed costs will arise, who can bear them, which market entries will be prevented, and what shared infrastructure could mitigate the burden?

A European child protection concept should be risk-based. Social networks, video platforms, games, and AI chatbots differ in their function and potential for harm. Blanket rules can lead to evasive maneuvers, over-blocking, and unnecessary data collection. More sensible are clear prohibitions on particularly manipulative designs, data-minimizing age verification, secure default settings, independent testing, and specific requirements for dialogic systems with emotional appeal.

In parallel, Europe must strengthen its supply side. This includes competitive electricity prices for data centers, faster permitting processes, long-term availability of computing capacity, larger growth funds, and joint procurement by public authorities. The announced expansion of AI factories and gigafactories can help, provided that the mobilized investment sums are actually translated into usable capacity, rapid access, and market-ready companies.

The public sector can also act as a demanding first customer. European AI providers need not only funding, but also reference projects, predictable demand, and the ability to scale solutions across borders. Common standards for administration, industry, healthcare, energy, and education could create a domestic market large enough to build international competitiveness.

Leadership decides the AI ​​economy

The three debates lead to a common conclusion: Humans do not retain control simply because AI remains fundamentally a tool. Control arises from institutions, technical limitations, ownership structures, competencies, and verifiable decisions. Without these structures, the formal presence of a human being can become a mere gesture of oversight, while systems, incentives, and platforms effectively determine the direction.

For companies, the central task is to transform AI from an isolated tool into a responsible operating model. Marketing needs less indiscriminate content production and more strategic differentiation. Research requires not only powerful agents but also measurable oversight and independent control. Leaders must translate productivity gains into real value creation and must not equate increased speed with better decision-making.

The challenge for policymakers lies in coordinating protection and innovation. Child protection is not a secondary economic issue, because societal harm incurs real costs. Competitiveness is also not a secondary issue, because regulations without domestic providers can increase dependence on foreign platforms. Europe must therefore become both more consistent and more pragmatic: clear protection goals, simple implementation, strong enforcement, shared infrastructure, and more capital for scaling.

The provocative assertion that Europe regulates while America scales up touches on a real weakness, but remains incomplete. The United States also regulates when societal and political pressure increases. Europe also invests and scales up, but so far more slowly and in a more fragmented way. The crucial difference lies less between regulation and freedom than between speed, capital depth, infrastructure, and institutional capacity.

The next phase of AI will not be decided solely by better models. It will be decided by the ability to transform technological acceleration into economic and societal benefits. Those who only slow down will lose their power to shape the future. Those who only accelerate will increase uncontrolled risks. Sustainable success will come to those who master both: scaling innovation and decisively defining its direction.

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

 

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here wolfenstein@xpert.digital:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

 

 

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs