A plateau for artificial intelligence? (I)

The question of whether artificial intelligence (AI) is approaching a “hard ceiling” — a point beyond which further improvement becomes extremely costly, marginal, or even impossible — is at once speculative and urgent. The pace of AI development in recent years, especially in large models and generative systems, invites both optimism and scepticism. Drawing on scholarship in AI, philosophy, cognitive science, and recent research, we will argue that while we are probably not at the absolute “limit” yet, several structural and conceptual constraints suggest that we may soon face diminishing returns, or at least encounter qualitatively different bottlenecks than in past decades. The question is not whether AI willhit limits, but when and in what form.

The nature of progress in AI: scaling, architecture, and paradigm shifts

To assess whether AI might be nearing a plateau, it helps to reflect on how we have arrived where we are, and whether those trajectories have unbounded potential.

1. Historical drivers: scale, compute and data

In the last decade particularly, the dominant driver of progress in many branches of AI (especially large language models and vision models) has been scale — more data, more computational algorithms, and more parameters. This scaling paradigm has proven remarkably effective: models trained with ever more resources tend to produce qualitatively better performance, sometimes in surprising generality (emergent behaviors). Many researchers expect further gains from continued scaling, albeit with higher costs (computational capacity, energy and environmental). Indeed pretraining large foundation models is already extremely expensive, and that cost curve is steep.

However, as Ilya Sutskever and others have observed, data is a limited resource: the internet is finite, and the marginal gain from more raw text data is diminishing.  If models exhaust the “low-hanging fruit” in data, and returns on additional parameters or compute slow down, then a scaling-only approach may yield diminishing returns.

2. Architectural innovation vs incremental improvement

Beyond scale, further breakthroughs may require new architectures, learning paradigms, or principles of cognition (e.g. incorporating symbolic reasoning, causal inference, world models, or embodied interaction). But such shifts are hard, and there may be conceptual ceilings to how far one paradigm can be pushed before requiring replacement.

In this sense, even if today’s paradigm has room to improve, that room may be bounded in practice by the difficulty of paradigm change. Some researchers think we may be in a transitional period — where scaling is reaching its limits, and the next leap requires deeper conceptual insight (e.g. integrating reasoning, common sense, planning).

If such deeper insight is hard or even unachievable, we may reach a practical plateau: improvements become ever more marginal, requiring exponentially more effort.

Fundamental and conceptual limitations

Beyond engineering and cost constraints, there are deeper limitations arising from the nature of intelligence, cognition, and representation. These may impose ceilings on what AI in principle could do, or at least what it can do in forms we currently envisage.

1. Lack of true understanding, common sense and world models

A classic critique is that current AI systems are essentially pattern matchers: they learn statistical correlations from large corpora, but lack a grounded understanding of the world. They don’t experience or model the world in the way humans do, and they struggle with common sense reasoning, causality, or conceptual understanding beyond surface patterns. 

For example, in solving word problems combining mathematics and real-world semantics, current systems still fail reliably. As Davis (2023) argues, AI systems have difficulty bridging the gap between symbol manipulation and semantic understanding. 

If this gap is not just a matter of scaling but a difference of kind, then merely refining current models may not surmount it.

2. Reasoning collapse in high complexity tasks

Recent empirical work suggests that advanced AI systems may suffer fundamental breakdowns when confronted with complex reasoning tasks. For instance, Apple researchers found a “complete accuracy collapse” in so-called large reasoning models (LRMs) when task complexity rises beyond a threshold: the models’ reasoning degrades sharply, worse than more naïve or simpler approaches. 

This suggests that the architecture and learning mechanisms of LLMs and reasoning models might be brittle: they do well on mid-level complexity tasks, but hit a hard wall when extrapolating to higher-order reasoning. If that generalises, it indicates a structural ceiling to performance in reasoning-intensive domains.

3. Interpretability, alignment, and trust

As AI models grow in complexity, their opacity becomes more acute. Understanding why a system outputs a given response is increasingly hard, giving rise to issues of explainability and trust. This “black box” nature constrains adoption in high-stakes domains (medicine, law, policy). 

Moreover, alignment with human values (making AI behave in ethically acceptable, controllable ways) remains a deep unsolved problem. Models may optimise proxies that diverge from human intent. The “alignment problem” is in part a ceiling on how capable AI can be while still being safe and trustworthy. (See for example Brian Christian’s The Alignment Problem.) 

Thus even if technically higher performance is possible, deploying it safely may become prohibitively difficult, effectively acting as an upper bound in practice.

4. Symbolic grounding, embodiment, and the frame problem

AI systems generally lack embodiment (sensory-motor interaction with the physical world) and thus struggle with grounding symbols and concepts in physical reality. Some branches of AI argue that true intelligence requires embodied interaction, not just disembodied text and image processing. Without grounding, the system cannot learn truly novel concepts outside its training distribution.

Relatedly, the “frame problem” in AI and philosophy — the difficulty of determining which aspects of a situation to attend to without exhaustive enumeration — has long challenged AI. In dealing with open-ended, dynamic environments, deciding which variables are relevant is nontrivial. These problems may impose cognitive boundaries upon what AI can reliably do, especially in unstructured, real-world settings.

5. The compressor / Kolmogorov complexity barrier

On a more theoretical level, one might consider that intelligence is partly about compressing and modeling patterns. But there are limits in data compressibility, and in the trade-off between expressive power and overfitting. If learning more elaborate representations yields diminishing compression gains (because the “signal” is mostly noise beyond a point), then further gains are constrained.

In other words, there is an information-theoretic bound: not all patterns are learnable from finite data, and the more intricate the pattern, the more data or stronger inductive biases you need. Past some frontier, further learning may require stronger prior assumptions or structural inductive assumptions which are hard to discover.

Practical and socio-technical ceilings

Even if the “in principle” ceilings are high, real-world constraints may enforce earlier plateaus.

1. Computational, energy, and environmental cost

The search for ever-larger models is increasingly expensive in computation, energy and carbon footprints. At some point, the marginal benefit of extra computation power may fail to justify its cost. If training and inference become too resource intensive, further scaling may be infeasible for most actors.

2. Data limitations and diminishing returns

As already mentioned, data is finite; the “frontier” of new, diverse, clean data is narrowing. Moreover compounding data quality problems (bias, mislabelling, duplication, distribution shift) further erode returns. Building models that generalise robustly across contexts demands more than raw scale; it demands better data and domain adaptation methods, which are challenging. The quantity of human input to ensure accurate AI output becomes exponentially higher.

3. Institutional, regulatory, and safety constraints

Deploying AI in real-world systems (healthcare, transportation, governance) must contend with regulation, safety certification, liability and public trust. These extra layers may slow or bar further deployment of the most capable systems. Even if we can build better models, we might choose not to deploy them because of legal or ethical risk.

4. Overhype, hype cycles, and misallocation

Some of the promise of AI is speculative or exaggerated, which leads to misallocation of talent and capital. If resources are poured into incremental improvements rather than paradigm-changing ideas, progress may stall. The peer-review and publication system is already strained by exponential AI research outputs, raising concerns about quality and reproducibility. The fact is that humans can no longer assess whether materials are produced by AI models, and they lack the capacity to do so or to evaluate whether AI models are producing accurate information.

5. Human-in-the-loop and augmentation ceilings

In many applications, AI is integrated into human workflows as an assistant or amplifier. There may be diminishing returns in how far AI can augment human capacities before humans become the bottleneck (in understanding, oversight, decision-making). In other words even if AI capability grows, human cognition, culture and organisations might limit how far it is fruitful to push it.

Are we “just about there”?

Given these theoretical, empirical and institutional constraints, what is the state of the art today — and is it plausible that we are nearing the “peak” of AI’s capacity?

1. Empirical signs of flattening or collapse

The recent “accuracy collapse” in reasoning models is a red flag: as AI systems scale in purported reasoning ability, they may actually perform worse on high-complexity tasks beyond some threshold. 

Surveys of AI researchers (e.g, the Thousands of AI Authors on the Future of AI) show a wide spread of forecasts, with many expecting significant improvements in the next decade, but also a nontrivial minority expecting stagnation or ceiling effects. 

Delphi-style meta-studies of AI’s future (Alon et al. 2025) review multiple forecasts and point out that many predict breakthroughs or “domino effects” from new paradigms — but also warn of bottlenecks in alignment, computational capacity and societal acceptability. 

Thus the community is not unanimous: many still expect further growth, but many also see credible risks of plateauing.

2. Some domains already look saturated

In narrow domains — e.g. image classification, speech recognition, basic language tasks — progress is nearing saturation: gains are incremental and ever more expensive. The low-hanging fruit is largely picked. In many practical tasks, the gap between AI and human-level performance is closing, leaving harder tasks (reasoning, abstraction, common sense) still open.

Also, performance in benchmarks can mask brittleness: models may exhibit adversarial failures, domain fragility, or poor out-of-distribution generalisation. A good example is AI filtering of job applicants' curriculum vitae: an AI model using rules to parse text in a curriculum vitae may miss the quality of a curriculum vitae, because it is not set out in a format that the AI algorithm's rules recognise.

3. Qualitative change may be needed

It is plausible that further progress will look qualitatively different: not a smooth continuation of scaling, but sudden jumps tied to new paradigms (causal learning, hybrid symbolic–neural systems, embodied agents, interactive learning). If those paradigms are hard to discover, progress may slow until breakthroughs occur.

In other words, we may be entering a “valley of diminishing returns” until new ideas emerge. If paradigmatic shifts are rare or hard, then for a period AI will exhibit plateau-like behavior, even if deeper ceilings remain far away.

Arguments against a near-term ceiling

It is worth airing counterarguments. The notion that AI is close to its limit is pessimistic; there are reasons to believe we still have a long horizon.

1. Unforeseen innovation: Past history shows that breakthroughs sometimes come from unexpected directions (e.g. deep learning’s resurgence). The next paradigm might come from neuroscience, quantum computing, self-supervised world modelling, or architectures we have not yet conceived.

   
   

2. Modular systems and ensembles: Intelligence might emerge from modular composition of specialised systems (vision, planning, theory of mind, simulation), rather than monolithic models. Scaling and integrating such modules may still yield large gains.

   
   

3. Active learning and interaction: Embodied agents interacting with environments (e.g. robotics, reinforcement learning) could gain self-supervised experience not limited by static datasets. That suggests a new axis of growth beyond the textual web.

   
   

4. Transfer, meta-learning, and open endedness: Systems that learn to learn, or explore latent structures, may better generalise, reducing the need for ever more data. That could break current bottlenecks.

   
   

Therefore a ceiling is not inevitable in the near term, especially if new ideas arrive. But these arguments depend on the success of paradigm shifts, which are uncertain.

Conclusion: approaching the “practical ceiling,” not the absolute one

Putting this together, the most defensible position is a cautious middle ground:

• We are not (yet) at the absolute, principle-level limit of artificial intelligence: there likely remain unexplored architectures, integration strategies, and paradigms that could push performance further.

  
  

• But we may be close to practical ceilings in many domains, especially those dominated by scaling and data. Gains still exist, but they are exponentially more expensive, riskier, or incremental.

  
  

• In reasoning, abstraction, common-sense understanding, safety alignment, and real-world deployment, the bottlenecks are already acute, and further progress may require qualitative innovation, not just brute force.

  
  

• Thus, rather than expecting continuous exponential growth, it is plausible that the next decade will see slower progress, selective breakthroughs, and plateaus in many areas, until new conceptual leaps occur.

In short, artificial intelligence may not be “about to reach the peak” in any absolute sense, but we may be entering a phase where diminishing returns intensify, and where the constraints (computational, theoretical, societal) become as decisive as algorithmic ingenuity. The frontier is not gone yet — but its nature is shifting, and only fresh ideas may break through future ceilings.

---

Bibliography

Baars, B.J. (1988). A Cognitive Theory of Consciousness. Cambridge: Cambridge University Press.

Christian, B. (2020). The Alignment Problem: Machine Learning and Human Values. New York: W. W. Norton & Company.

Davis, E. (2023). “Commonsense Reasoning and Large Language Models.” arXiv preprint arXiv:2301.09723. Available at: https://arxiv.org/abs/2301.09723

Dehaene, S. (2014). Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. London: Viking.

Friston, K. (2010). “The Free-Energy Principle: A Unified Brain Theory?” Nature Reviews Neuroscience, 11(2), pp. 127–138.

LeCun, Y. et al. (2022). “A Path Towards Autonomous Machine Intelligence.” Open Review draft, Meta AI. Available at: https://openreview.net/forum?id=BZ5a1r-kVsf

Sutskever, I. (2024). Remarks in “AI with Reasoning Power Will Be Less Predictable,” Reuters, 14 December. Available at: https://www.reuters.com/technology/artificial-intelligence/ai-with-reasoning-power-will-be-less-predictable-ilya-sutskever-says-2024-12-14/

Tononi, G. (2004). “An Information Integration Theory of Consciousness.” BMC Neuroscience, 5(1), p. 42.

Alon, N., Kemp, C., Tegmark, M. et al. (2025). “Forecasting the Capabilities and Limits of Artificial Intelligence: A Delphi Study.” Futures, 156, Article 102319. Available at: https://doi.org/10.1016/j.futures.2024.102319

AI Impacts Team (2023). Thousands of AI Authors on the Future of AI: Survey Synthesis. AI Impacts Report. Available at: https://aiimpacts.org/wp-content/uploads/2023/04/Thousands_of_AI_authors_on_the_future_of_AI.pdf

AAAI Presidential Panel (2025). Artificial Intelligence 2025: Directions, Debates, and Dependencies. AAAI Report. Available at: https://aaai.org/wp-content/uploads/2025/03/AAAI-2025-PresPanel-Report_FINAL.pdf

Apple Machine Intelligence Research Team (2025). “On the Collapse of Reasoning Accuracy in Large Reasoning Models.” Unpublished working paper, referenced in The Guardian, 9 June 2025. Available at: https://www.theguardian.com/technology/2025/jun/09/apple-artificial-intelligence-ai-study-collapse