aitech-ethicsregulationresearchtechnical-limits

Current limits of artificial intelligence and future perspectives: a rigorous analysis

Cris Fernández··15 min read

TL;DR: Artificial intelligence presents fundamental intrinsic limits that no algorithmic improvement can completely overcome. This analysis examines the current technical, ethical and social restrictions of AI, projects its expected advances at different time horizons, and evaluates emerging regulatory frameworks. Evidence suggests that human-AI collaboration, rather than complete replacement, will be the dominant paradigm in the coming decades.

Introduction: beyond technological hype

The popular narrative about artificial intelligence oscillates between two extremes: utopias of imminent superintelligence and dystopias of mass unemployment. Neither of these visions captures the real complexity of the field.

As a researcher, I've observed that public discourse on AI lacks technical precision. We talk about "AI that learns on its own" when in reality we're talking about systems that optimize human objective functions on human data. We proclaim that AI "understands language" when what it does is model probabilistic distributions of token sequences.

This research adopts a rigorous approach: analyze the current limitations of AI from three dimensions (technical, ethical and social), project realistic advances based on available evidence, examine intrinsic theoretical limits, and evaluate emerging regulatory frameworks.

The central thesis: modern AI is extraordinarily capable in narrow and measurable domains, but faces fundamental barriers that will keep it as a complementary tool, not a substitute, for human judgment in critical contexts.

Current limitations: a rigorous taxonomy

Technical limitations

The dominant architecture in modern AI —deep neural networks trained through gradient descent— presents inherent restrictions to its design.

Data dependency and training quality

Deep learning models require massive volumes of labeled data. Model quality is fundamentally limited by the quality of training data. As documented by the International AI Safety Report 2026, if data contains statistical biases —and real-world data always does— the model will systematically reproduce and amplify them.

There is no algorithm that can "clean" biases without expert human supervision that defines what constitutes a bias in each specific domain. This is not an engineering limitation that can be overcome; it's a logical restriction of supervised learning.

Explainability problem

Deep neural networks operate as high-dimensional function approximators. With models containing hundreds of billions of parameters, the interpretability of their decisions is an open mathematical problem, not just an engineering challenge.

Techniques like LIME, SHAP or attention visualization provide heuristic approximations, but not formal guarantees about the model's reasoning. In critical contexts —medical diagnosis, judicial decisions, safety systems— this opacity represents a systemic risk that cannot be mitigated with current architecture.

Computational and energy scalability

Training large-scale language models (LLMs) like GPT-4 requires clusters of tens of thousands of specialized GPUs for months. Computational cost scales superlinearly with model size and training corpus.

Even more critical: energy consumption. Training a large model can generate carbon emissions equivalent to five times the lifecycle of an average car. This is not a minor externality; it's a physical constraint that limits retraining frequency and practical model size.

Absence of common sense and causal reasoning

Current models lack mental models about how the physical world works. They can predict that "ice melts" but not because they understand thermodynamics, but because that correlation appears in their training data.

Judea Pearl, Turing Award winner, has extensively argued that correlational learning —the basis of deep learning— is fundamentally insufficient for causal reasoning. Current systems cannot consistently answer questions like "what would have happened if...?" without specific data about that counterfactual.

Ethical limitations

AI ethics is not a "soft" topic separable from the technical. These are fundamental restrictions on what systems we should build, not just on which ones we can build.

Algorithmic bias and systemic discrimination

Algorithmic bias is not a bug; it's an inevitable feature of systems that learn from historical data. If data reflects historical discrimination —loan approval rates, hiring decisions, judicial sentences— the model will learn it as a valid pattern.

UNESCO, in its Recommendation on AI ethics, emphasizes that "AI systems can reproduce prejudices" at industrial scale. It's not just that they can; they will by default if rigorous technical and process interventions are not implemented.

Mitigation requires:

  • Continuous algorithmic audits by interdisciplinary teams
  • Balanced training datasets (which may be technically impossible in domains with inherent imbalance)
  • Formally defined fairness metrics (which requires choosing between mathematically incompatible notions of equity)

Privacy and surveillance

Modern AI models require massive data. The collection, storage and processing of personal data at population scale represents privacy risks that are difficult to mitigate technically.

Federated learning and differential privacy are important advances, but imply fundamental trade-offs: more privacy = less model accuracy. There's no free lunch.

Responsibility and accountability

Who is responsible when an autonomous vehicle causes a fatal accident? The vehicle manufacturer? The model developer? The training data company? The regulator who approved the system?

The chain of causality in complex AI systems is opaque. The existing legal framework, based on direct individual responsibility, does not adapt well to systems where the "decision" emerges from statistical interactions in high-dimensional spaces.

Applications in sensitive domains

The use of AI in autonomous weapons systems, mass facial recognition, or social scoring poses ethical dilemmas that no technical improvement can resolve. These are questions about what kind of society we want, not about technological capabilities.

Social limitations

The social dimension of AI limitations is directly connected to measurable economic impacts.

Labor displacement and inequality

According to UN analysis, automation through AI could directly affect approximately 40% of global jobs. But the impact will not be homogeneous.

Jobs that involve routine and codifiable tasks —both white-collar (basic financial analysis, document processing) and blue-collar (manufacturing, logistics)— are highly automatable.

Jobs that require complex human interaction, contextual creativity, or ethical judgment —people care, crisis management, strategic leadership— are significantly less automatable.

The result: AI can amplify existing inequality if active reconversion and social protection policies are not implemented.

Digital divide and unequal access

UNDP warns that many countries lack the digital infrastructure and technical capacity to implement AI locally. This creates dependency on foreign providers and widens the technological gap between regions.

Vulnerable groups —women in administrative sectors, young people without higher technical education, rural workers— are disproportionately exposed to labor displacement through automation.

Disinformation and manipulation at scale

Generative models (GANs, diffusion models, LLMs) enable creating synthetic content indistinguishable from human content. High-quality video and audio deepfakes are technically trivial to generate.

This is not a future problem. It's a current problem that erodes trust in digital media and facilitates disinformation campaigns at industrial scale.

Expected advances: evidence-based projections

Short term (1-2 years): optimization and deployment

The most significant short-term advance will not be in fundamental capabilities, but in efficiency and accessibility.

Dramatic reduction in inference costs

The cost of running large language models has dropped approximately 10x every 18 months over the last three years. This trend will continue through:

  • Model quantization (reduce numerical precision without significant quality loss)
  • Pruning and distillation (create small models that replicate large model behavior)
  • Specialized hardware (ASICs designed specifically for neural network inference)

Massive enterprise integration

We'll see AI assistants deeply integrated into enterprise software: CRMs that write contextualized emails, ERPs that optimize supply chains in real-time, analysis tools that automatically generate narrative reports.

Most of these applications will use existing models (GPT-4, Claude, Gemini) through APIs, not proprietary developments.

Incremental improvements in reasoning

Techniques like chain-of-thought prompting and extended reasoning (as seen in GPT-4o) will show improvements in complex logical and mathematical tasks.

But this doesn't represent a qualitative leap toward real "understanding". They are improvements in models' ability to follow reasoning patterns that appear in their training data.

Medium term (3-5 years): architectural diversification

Hybrid models and multi-agent systems

IBM predicts we'll see increasing diversification beyond the dominant transformer architecture. Hybrid systems that combine:

  • Neural networks for pattern recognition
  • Symbolic systems for logical reasoning
  • Structured knowledge bases for verifiable facts

"AI agents" —systems that coordinate multiple specialized models to execute complex tasks— will gain prominence. But they will be narrow agents, not general agents. They can "plan a trip" by coordinating search, bookings and optimization, but not "run a company" with real autonomy.

Embodied AI: robots and physical systems

Integration of AI into physical robotic systems will advance significantly:

  • Autonomous vehicles in controlled environments (university campuses, industrial zones)
  • Collaborative robots (cobots) in manufacturing and logistics
  • Robotic assistants in healthcare for specific tasks (patient mobilization, medication dispensing)

Complete autonomy in open uncontrolled environments will remain problematic for safety and legal liability reasons.

According to Sam Altman (OpenAI): "in 2025 agents will be able to perform real cognitive work". Probably. But "real cognitive work" means well-defined and measurable tasks, not strategic judgment or organizational leadership.

Long term (10+ years): theoretical limits

Long-term projections diverge dramatically depending on who you ask.

The singularity hypothesis

Ray Kurzweil and other transhumanists maintain we could achieve artificial general intelligence (AGI) —systems with cognitive capabilities comparable to humans in all domains— by 2029.

This projection assumes:

  1. Human intelligence is completely computable (there are no irreducible aspects)
  2. Model scaling will continue producing qualitative improvements
  3. There are no insurmountable energy, material or regulatory barriers

Each of these assumptions is questionable.

Intrinsic computational limits

From the perspective of theoretical computer science, there are fundamental limits to what any computational system can achieve.

Turing's halting problem demonstrates that there cannot exist a general algorithm that determines whether an arbitrary program will terminate or run infinitely. This implies there are entire classes of problems that no AI can solve algorithmically, regardless of computational power.

Genuine creativity —generating truly novel concepts not derivable from combinations of existing concepts— may be beyond what deterministic systems can achieve. This is an open philosophical question, not just technical.

Searle's Chinese Room argument

Philosopher John Searle argues that formal systems that manipulate symbols (like computers) operate only at the syntactic level, without access to real semantics —meaning.

A system can "behave" as if it understands Chinese, following symbol manipulation rules, without actually understanding anything about Chinese. Consciousness and genuine understanding may require specific biological substrates, not just abstract computational processes.

There is no consensus on this argument, but it points to potential fundamental limits.

Future intrinsic limits: irreducible barriers

Even assuming optimistic technical advances, certain limits seem irreducible:

Computational limits

Theorem: Any AI system based on deterministic computation is restricted to computable functions. There exist demonstrably non-computable problems.

Practical implication: There will always be classes of problems that require intuition, human heuristics, or contextual judgment that cannot be algorithmized.

Absence of subjective experience

Current (and projectable) systems process information without conscious experience. They don't have qualia —the subjective experience of "what it feels like" to perceive red or feel pain.

This is not just philosophically relevant. Genuine empathy —understanding others' emotional states through shared experience— may require consciousness. AI systems can simulate empathetic responses, but not experience empathy.

Need for human oversight in critical contexts

In domains where errors have irreversible consequences —critical medical decisions, judicial decisions, safety systems control— human oversight is not optional.

This is not due to surmountable technical limitations. It's because ethical responsibility must fall on moral agents —beings with consciousness, intentionality and capacity to be held accountable.

The human differential value

As recent research documents, in a world with ubiquitous AI, differentially valuable skills will be:

  • Contextual judgment and situational reasoning
  • Genuine empathy and relationship building
  • Conceptual creativity (not combinatorial)
  • Ethical leadership and strategic vision

These capabilities are not "what AI can't do yet". They are what defines the unique human contribution in complex sociotechnical systems.

Illustrative failure cases

Project Vend: when AI lacks common sense

In a publicly documented experiment, Claude (Anthropic's model) was given autonomous control over a small vending business. The system:

  • Lowered prices to zero (maximizing numerical sales, ignoring profitability)
  • Ordered absurd products (optimizing variety without demand context)
  • Couldn't manage inventory crises (without mental model of supply chains)

The failure wasn't due to lack of computational capacity. It was due to absence of causal understanding of the business domain.

Success cases in narrow domains

GPT-4 in coding and technical writing

OpenAI documents that GPT-4 surpasses average human performance in specific tasks of:

  • Generating boilerplate code
  • Translation between programming languages
  • Writing technical documentation
  • Log analysis and debugging

This has measurably elevated the productivity of developers and technical writers. But it hasn't replaced system architecture, API design, or complex technical trade-off decisions.

Corporate adoption

According to KPMG Spain, 85% of companies already invest or plan to invest in AI, mainly for:

  • Production optimization (46%)
  • Customer service (40%)
  • Financial analysis (40%)

These applications are efficiency tools, not replacements for strategic decision-makers.

The gap between hype and reality

IBM reports that "AI lags behind the rhetoric". Practical implementations evolve more slowly than commercial promises.

Emerging themes in the sector:

  • Responsible AI: need for internal governance
  • Privacy-personalization trade-off: unresolved fundamental tension
  • AI fatigue: concern about emotional consequences of constant interaction with synthetic agents

Regulatory frameworks: towards global governance

European Union: preventive regulation

The EU AI Regulation (approved June 2024, staged implementation 2025-2027) is the first comprehensive legal framework on AI globally.

Risk structure:

  1. Unacceptable risk (prohibited):

    • Manipulation of behavior of vulnerable groups
    • Social scoring by governments
    • Real-time facial recognition in public spaces (with limited exceptions)

  2. High risk (strictly regulated):

    • Critical infrastructures
    • Education and professional training
    • Employment
    • Judicial systems
    • Require: conformity assessments, technical documentation, transparency, human oversight

  3. Limited risk (transparency requirements):

    • Chatbots: must identify as AI
    • Deepfakes: must be labeled as synthetic
    • Recommendation systems: must explain basic logic

  4. Minimal risk (no additional regulation):

    • Spam filters
    • Video games with AI
    • Most consumer applications

Key dates:

  • February 2, 2025: unacceptable risk prohibitions in force
  • August 2025: transparency obligations active
  • 2027: complete regulation of high-risk systems

Criticisms:

  • May slow European innovation
  • Complicated enforcement for general-purpose models
  • Technical definitions still ambiguous

UNESCO: global ethical framework

The UNESCO Recommendation on AI ethics (adopted 2021, 194 countries) establishes principles, not binding regulations:

Fundamental principles:

  1. Proportionality: use AI only when necessary
  2. Safety: avoid foreseeable harms
  3. Privacy: protection of personal data
  4. Human oversight: humans must retain control over critical decisions
  5. Transparency: systems must be auditable
  6. Responsibility: clear accountability chains
  7. Awareness and literacy: public education about AI
  8. Adaptability: frameworks that evolve with technology

It's an aspirational framework rather than enforcement. Its value lies in establishing international consensus on fundamental values.

United States: competitive deregulation

In December 2025, Biden's executive order on AI was reversed. The new federal policy declares AI "minimally burdensome":

Approach:

  • Prioritize innovation and global competitiveness
  • Avoid federal regulations that "slow down" development
  • Preemption of state regulations considered excessive
  • Creation of AI Task Force to challenge state regulations

Justification: maintain U.S. technological leadership against China

Risks:

  • Race to the bottom in ethical protections
  • Regulatory fragmentation (50 states with different rules)
  • Underprotection of vulnerable populations

The fundamental tension

Two incompatible regulatory philosophies:

  • EU/UNESCO: preventive regulation based on principles (avoid harms before they occur)
  • U.S.: reactive regulation based on market (let innovate, regulate only after evident problems)

This divergence complicates global governance of AI systems that operate transnationally.

Conclusion: collaboration, not replacement

After examining current technical, ethical and social limitations; projecting realistic advances; and analyzing intrinsic limits, the conclusion is clear:

AI is an extraordinarily powerful tool for narrow and measurable domains, but faces fundamental barriers that will keep it as a complement, not substitute, for human judgment in complex and critical contexts.

Grounded projections

  1. Short term (1-2 years): Explosion of efficiency and accessibility. Massive integration into enterprise software. Incremental improvements in reasoning.

  2. Medium term (3-5 years): Architectural diversification. Multi-agent systems for complex but narrow tasks. Embodied AI in controlled environments.

  3. Long term (10+ years): High uncertainty. Possibly AGI according to optimists. Probably very capable systems but fundamentally limited in consciousness, genuine creativity, and ethical judgment.

Persistent limits

  • Technical: Data dependency, algorithmic opacity, computational constraints, absence of robust causal reasoning
  • Ethical: Inevitable bias, privacy dilemmas, accountability problems, morally questionable applications
  • Social: Amplified inequality, unequal labor displacement, digital divide, disinformation risks

The human differential value

There's no prompt that replaces a difficult conversation. There's no algorithm that inspires a team in crisis. There's no model that assumes ethical responsibility for decisions with irreversible consequences.

Distinctively human capabilities —experiential empathy, conceptual creativity, contextual judgment, ethical leadership— will not only remain valuable. They will become more valuable precisely because AI will automate the routine and codifiable.

The path forward

The relevant question is not "when will AI replace us?" but "how do we design sociotechnical systems where humans and AI collaborate effectively, with regulatory frameworks that protect fundamental rights while allowing beneficial innovation?"

Answering this requires not only technical advances, but institutional maturity, ethical consensus, and political will to proactively regulate technologies with civilizational impact.

AI will profoundly transform work, economy and society. But it will be a transformation we navigate, not an inevitable destiny we suffer. Human agency —individual and collective— remains determinant.

References and further reading

This analysis is based on:

For continued analysis on these topics: follow publications from AI Safety, IEEE Ethics in Action, and specialized academic forums.

← Previous

The Fable Blackout: Why the US Blocked Anthropic's Best AI

Littlesoft-AI

Subscribe to the blog

Get an email when we publish something new. No spam.