Artificial intelligence has crossed a threshold of visibility. Systems that once belonged to specialist laboratories now write software, interpret images, summarize documents, translate speech and answer questions for hundreds of millions of people. Their fluency can feel like understanding, and their speed can make expertise appear instantly available.
Yet the decisive frontier is no longer fluency. It is reliability.
A model can produce a persuasive answer without possessing evidence for it. It can solve a difficult problem and then fail a simpler variation. It can follow a plan for several steps before quietly losing the original objective. When connected to tools, the same weaknesses can move beyond words: an agent may send a message, change a file, purchase a service or modify a technical system.
The central research and institutional challenge is therefore not simply to make artificial intelligence more capable. It is to make capability measurable, corrigible and trustworthy under the conditions in which it will actually be used.
This distinction separates a remarkable demonstration from dependable infrastructure.
What a language model is—and what it is not
A large language model is trained to predict and generate sequences of tokens: pieces of text that may correspond to words, punctuation or fragments of words. During training, the system adjusts a vast number of numerical parameters so that it becomes increasingly effective at representing patterns in language and other data.
This apparently simple objective can produce complex abilities. Language contains descriptions of mathematics, science, human behaviour, software and the physical world. Learning its structure can enable a model to translate, classify, explain, infer relationships and combine ideas in ways that were not individually specified by its developers.
But the training objective does not automatically guarantee truth. A model learns which continuation is plausible in context, not which claim has been experimentally verified. It does not carry a perfectly indexed database of facts. Its internal representations are distributed, approximate and shaped by the quality, frequency and contradictions of the material used in training.
This is why a model may be eloquent and wrong. The sentence can be statistically coherent while its subject, date, quotation or causal claim is false.
Calling such errors “hallucinations” is convenient but imperfect. The term may suggest that a machine has an inner experience comparable to a human perception. More precisely, the system is generating unsupported or false content with no built-in guarantee that the content corresponds to reality.
The problem is not solved merely by increasing model size. Larger models often acquire broader capabilities, but reliability depends on training incentives, data, evaluation, access to evidence, task design and the system surrounding the model.
Why models guess when they should abstain
Many evaluations reward an answer only when it matches the expected response. An unanswered question receives no credit. Under that scoring rule, guessing can be rational: a model that always attempts an answer may score higher than one that honestly says it is uncertain.
A 2026 Nature paper by Adam Tauman Kalai and colleagues formalized this incentive problem. The researchers argued that conventional accuracy-based evaluations can encourage confident guessing and therefore hallucination. Their work does not imply that every false output has one cause, but it identifies an important structural pressure: systems are frequently rewarded for producing an answer rather than for knowing when evidence is insufficient.
Reliability requires a different objective. A useful system should be able to distinguish among at least four states:
- It has strong evidence and can answer directly.
- It can answer after consulting an external source or tool.
- It can offer a hypothesis but must label the uncertainty.
- It lacks sufficient evidence and should abstain.
Abstention is not a weakness. In medicine, engineering, law and scientific research, calibrated uncertainty is part of competence. A system that refuses five questions and answers the remaining ninety-five correctly may be more valuable than one that answers all one hundred while hiding five fabricated claims.
This changes how evaluation should work. Accuracy alone is insufficient. We also need calibration, citation fidelity, error severity, consistency across reformulations, performance under distribution shift and the cost of false confidence.
Does producing more steps mean reasoning better?
The emergence of reasoning-oriented models has improved performance on mathematics, programming and other tasks that benefit from intermediate computation. These systems may spend additional inference time exploring solutions, checking candidates or using tools before returning an answer.
That is genuine technical progress. It is not proof that the model reasons exactly as a person does, nor that visible step-by-step language reveals its complete internal process.
Three distinctions matter.
First, a correct answer is not evidence of a robust method. A model may exploit a familiar pattern in a benchmark without learning a transferable principle. Small changes in wording, irrelevant details or task structure can reveal brittleness.
Second, a plausible explanation is not necessarily the cause of the answer. Generated reasoning may rationalize a conclusion after the model has already been influenced by other statistical cues. Explanations should therefore be evaluated against behaviour, not accepted merely because they sound coherent.
Third, more computation does not guarantee monotonic improvement. Additional attempts can uncover errors, but they can also introduce new ones. Verification is valuable only if the verifier is sufficiently independent and capable of distinguishing correction from confident revision.
Research presented at ACL 2026 through the ReasonIF benchmark found substantial failures when open reasoning models were tested on whether they followed instructions throughout their reasoning processes. The result is a warning against treating “reasoning” as a single solved ability. Logical competence, instruction adherence, factual grounding, planning and uncertainty calibration are related but separable properties.
The most meaningful question is not whether a model produces a long chain of thought. It is whether its conclusions remain correct when the problem changes, whether intermediate claims can be checked and whether the system can detect when its method is failing.
Grounding intelligence in evidence
One route to greater reliability is to connect models to external evidence. Retrieval-augmented generation allows a system to search a controlled collection or the web and incorporate relevant documents into its response. Tool use can provide calculators, code interpreters, databases, sensors or specialized scientific software.
These methods reduce some errors, but they do not create an automatic truth machine. Retrieval introduces a new chain of possible failure:
- The search query may omit a critical concept.
- The system may retrieve an outdated or low-quality source.
- A relevant document may be misread.
- The model may cite a source that does not support the claim.
- Several individually correct facts may be combined into an invalid conclusion.
- A malicious page may attempt to manipulate the agent through hidden instructions.
Reliable grounding therefore requires provenance. Readers and downstream systems need to know where a claim came from, when the source was published, whether it is primary or secondary, and which part of the source supports the statement.
For high-stakes work, the design should separate evidence collection, inference and final communication. Independent checks can compare claims against sources. Structured tools can perform arithmetic rather than asking a language model to approximate it. Access should be limited to the minimum data and actions required by the task.
The goal is not to eliminate the model's generative ability, but to place it inside an epistemic architecture: a system that treats evidence, uncertainty and correction as first-class components.
From assistants to agents
An assistant responds. An agent acts toward a goal.
The difference may involve memory, planning, tool selection and repeated interaction with an environment. An agent asked to organize a journey might search transport options, compare schedules, reserve accommodation and update a calendar. A software agent might inspect a repository, modify code, run tests and open a review request.
Agency can multiply usefulness because many valuable tasks are not completed in one response. It also multiplies risk. Each step depends on the previous state, and small errors accumulate. If a system succeeds at an independent action 95 percent of the time, the probability of completing twenty such actions without error is only about 36 percent. Real tasks are not fully independent, but the calculation illustrates why long-horizon reliability is difficult.
Research benchmarks increasingly expose this gap. IBM's ITBench evaluated agents on realistic scenarios in site reliability engineering, security and compliance operations, and financial operations. In its initial study, state-of-the-art agents resolved only 13.8 percent of the SRE scenarios, 25.2 percent of the security and compliance scenarios, and none of the financial operations scenarios. The exact figures will age as systems improve, but the structural lesson will remain: success on isolated questions does not translate automatically into dependable work across tools and time.
Safe agents need more than stronger models. They need transaction boundaries, permission controls, previews, logs, reversible actions and escalation rules. A well-designed system should distinguish between reading a document and deleting it, between drafting an email and sending it, between recommending a purchase and spending money.
Human oversight must also be meaningful. Requiring a person to approve hundreds of opaque decisions is not effective control; it is fatigue disguised as governance. The system should present the evidence, uncertainty and consequences needed for a human to make a real decision.
Multimodality is not a single unified understanding
Artificial intelligence increasingly works across text, images, audio and video. Multimodal models can interpret a diagram, answer questions about a photograph, transcribe speech or generate a video from a description. This broadens the interface between humans and machines and creates valuable applications in accessibility, education, design and scientific analysis.
Yet combining modalities also combines their failure modes.
An image may be cropped, synthetic or taken out of context. Speech may contain noise, accents or several speakers. A video can suggest temporal and causal relationships that are not actually present. Text embedded in an image may be misread. When a model integrates all of these signals, it may produce an answer that feels more convincing precisely because several modalities are involved—even if one crucial observation is wrong.
Multimodal reliability must therefore test more than recognition accuracy. Does the system know which part of an image supports its conclusion? Can it distinguish absence of evidence from evidence of absence? Does it remain stable when irrelevant visual elements are added? Can it recognize manipulated media and disclose uncertainty?
These questions become more urgent as multimodal models connect to robots, medical imaging, vehicles and surveillance systems. Perception errors become action errors when the model is embodied or operationalized.
Small and local models have a different advantage
The public narrative often treats progress as a race toward ever-larger models and data centres. But many applications do not require the largest available system.
Smaller models can operate on telephones, personal computers, vehicles or industrial equipment. Local execution can reduce latency, preserve sensitive data, continue without an internet connection and lower the cost of repeated inference. A specialized model trained for one well-defined domain may outperform a general model while using far fewer resources.
This does not mean local AI is automatically private or safe. An application may still transmit telemetry; a compromised device may expose local data; and a compressed model can retain harmful capabilities. But edge deployment creates architectural options that centralized services cannot provide.
The future AI ecosystem is likely to be distributed. Large frontier models may handle open-ended or difficult tasks. Smaller models may classify, monitor or personalize locally. Specialized scientific systems may work with domain-specific representations. A routing layer may select the cheapest and safest model that can meet the required standard.
Efficiency should be treated as a capability. If two systems achieve comparable reliability, the one requiring less computation, energy and latency may be the more advanced technology.
AI for science: prediction is not discovery
Scientific applications reveal what artificial intelligence can contribute when it is connected to disciplined evidence.
Foundation models can learn patterns in proteins, molecules, materials, weather and Earth-system data. In 2025, the Aurora foundation model was reported in Nature after training on more than one million hours of geophysical data. It was adapted to tasks involving air quality, ocean waves, tropical cyclone tracks and high-resolution weather, outperforming selected operational forecasts while using far less computation for inference.
This is important, but the language of “AI discovery” requires precision. A model may generate a candidate molecule, predict a structure or identify a statistical relationship. Scientific discovery also requires experimental validation, causal interpretation, reproducibility and integration with existing knowledge. A predicted material is not a manufactured material; a proposed drug is not a safe medicine; a forecast improvement in a benchmark is not universal superiority under every weather condition.
The strongest scientific systems form a loop:
- Models propose predictions or candidates.
- Simulations and experiments test them.
- Measurements expose error and uncertainty.
- Researchers revise hypotheses and models.
- Independent groups attempt replication.
Artificial intelligence can accelerate several parts of this loop. It cannot replace the loop itself.
Scientific AI also demonstrates why domain expertise remains essential. A model may find correlations that a person would miss, while a scientist identifies confounding variables, impossible assumptions or measurements that do not answer the intended question. The productive relationship is neither blind automation nor rejection of machine assistance. It is a division of labour in which computation expands the space of possibilities and scientific method disciplines the conclusions.
Intelligence has an energy and material substrate
AI is often discussed as if it existed in an abstract cloud. In reality, it depends on semiconductor fabrication, servers, networks, cooling systems, electricity grids, water and physical sites.
The International Energy Agency estimated that data centres consumed about 415 terawatt-hours of electricity in 2024, approximately 1.5 percent of global electricity use. Its 2025 base-case projection placed consumption near 945 terawatt-hours by 2030, with AI as the most important source of growth alongside other digital services. The IEA reported that data-centre electricity demand rose 17 percent in 2025 and that consumption at AI-focused centres grew faster still.
These are projections and aggregated estimates, not a precise meter attached to “AI” as a single category. Future demand depends on adoption, chip efficiency, model architecture, utilization and the kinds of tasks people choose to run. Simple text generation and long-running agentic or video workloads can differ enormously in energy use.
Efficiency improvements are also real. Better hardware, quantization, model compression, caching and specialized architectures can reduce the energy required for each task. But lower cost may increase total usage—the rebound effect. An efficient system used a billion times can consume more resources than an inefficient system used a million times.
Responsible accounting should therefore report both efficiency per task and total consumption. It should consider local grid constraints, carbon intensity, water use, equipment lifetime and the environmental cost of manufacturing. The social value of the application matters as well: computation used to improve weather warnings cannot be evaluated in the same way as endless generation of disposable content.
Regulation is becoming operational
Governance is moving from general principles toward enforceable obligations.
In the European Union, requirements for providers of general-purpose AI models began applying in August 2025. From 2 August 2026, the European Commission gained enforcement powers over those obligations. The framework includes transparency and copyright responsibilities for general-purpose models, with additional evaluation, security and systemic-risk duties for the most advanced systems.
The EU AI Act is not proof that every technical problem has been solved. Standards, measurement practices and interpretations will continue to develop. Its importance lies in recognizing that general-purpose models participate in a wider value chain. A developer integrating a model into hiring, medicine or critical infrastructure needs information about capabilities, limits and risk controls.
In the United States, the National Institute of Standards and Technology's AI Risk Management Framework provides a voluntary structure organized around governing, mapping, measuring and managing risk. Its generative-AI profile identifies risks such as confabulation, privacy, information integrity, cybersecurity and harmful bias, while emphasizing testing and documentation across the system lifecycle.
Good governance should be proportional to context. A spelling assistant and a system influencing medical treatment should not face identical requirements. At the same time, calling a model “general purpose” does not remove responsibility when its deployment predictably affects rights, safety or access to essential services.
Regulation works best when it reinforces technical discipline: documented data and evaluation, incident reporting, independent scrutiny, secure development and clear responsibility. Paper compliance without meaningful testing creates the appearance of safety rather than safety itself.
Intelligence, generality and consciousness are different questions
Debates about artificial general intelligence frequently combine three separate issues.
Capability concerns what a system can do. Can it solve novel problems, learn unfamiliar tasks and operate across domains?
Generality concerns the breadth and transferability of those abilities. Does competence in one setting survive changes in context, tools and objectives?
Consciousness concerns subjective experience: whether there is something it is like to be the system.
Progress in capability does not by itself answer the question of consciousness. A model's use of first-person language is evidence that it can generate patterns found in human conversation, not proof of inner experience. Conversely, the absence of a settled theory of consciousness means categorical metaphysical claims should be made cautiously.
For practical governance, we do not need to solve consciousness before addressing present harms. Systems can affect employment, privacy, information and public decisions regardless of whether they experience anything. Accountability belongs to the people and institutions that design, deploy and control them.
The pursuit of general intelligence should also not distract from uneven competence. A system may exceed most people on a particular examination while lacking stable memory, physical experience, social responsibility or the ability to recognize an unusual real-world exception. Intelligence is not a single number.
What reliable AI would require
No single technique will create trustworthy artificial intelligence. Reliability is a property of the complete socio-technical system.
The next frontier requires progress in at least eight areas:
- Calibrated uncertainty: systems should estimate when they are likely to be wrong and abstain appropriately.
- Evidence and provenance: important claims should be traceable to sources, measurements or reproducible computation.
- Robust evaluation: tests must include unfamiliar conditions, adversarial inputs, long tasks and consequential errors—not only benchmark averages.
- Independent verification: critical outputs should be checked by methods that do not merely repeat the same model's assumptions.
- Bounded agency: tools, permissions and spending authority should match the task and be reversible wherever possible.
- Lifecycle monitoring: performance after deployment must be measured because environments, data and user behaviour change.
- Human factors: interfaces must communicate uncertainty and enable meaningful intervention rather than ceremonial approval.
- Institutional accountability: responsibility for failures cannot disappear into a chain of model providers, integrators and users.
Some applications will meet these standards sooner than others. A model that proposes code which is automatically tested has a strong verification channel. A model offering emotional or medical guidance confronts ambiguity, privacy and human vulnerability. The appropriate level of autonomy must reflect the consequence of error.
The Aeternum Perspective
Artificial intelligence has made the production of convincing language abundant. That makes the discipline of knowledge more valuable, not less.
The question is no longer whether a machine can produce an answer. It is whether the answer is supported, whether uncertainty is visible, whether an action can be reversed and whether responsibility remains identifiable.
Human intelligence is also fallible. People misremember, rationalize and speak beyond their evidence. The proper comparison is therefore not between perfect humans and imperfect machines. It is between systems of inquiry. Science became reliable not because scientists stopped making mistakes, but because methods of criticism, replication, measurement and correction made error discoverable.
Artificial intelligence needs an equivalent culture of correction.
The decisive frontier will not be reached by the model that speaks most confidently or produces the longest reasoning trace. It will be reached by systems that can participate in the creation of knowledge without concealing the boundaries of what they know.
Fluency made AI visible. Reliability will determine whether it becomes worthy of trust.
---
Selected sources
- Kalai, A. T., Nachum, O., Vempala, S. S. et al. “Evaluating large language models for accuracy incentivizes hallucinations.” Nature 653, 1047–1051 (2026). https://doi.org/10.1038/s41586-026-10549-w
- Kwon, Y., Zhu, S., Bianchi, F., Zhou, K. & Zou, J. “ReasonIF: Large Reasoning Models Fail to Follow Instructions During Reasoning.” Findings of ACL 2026 (2026). https://aclanthology.org/2026.findings-acl.1456/
- Joshi, C. K. et al. “ITBench: Evaluating AI Agents across Diverse Real-World IT Automation Tasks.” ICML 2025. https://research.ibm.com/publications/itbench-evaluating-ai-agents-across-diverse-real-world-it-automation-tasks
- Bodnar, C. et al. “A foundation model for the Earth system.” Nature 641, 1180–1187 (2025). https://doi.org/10.1038/s41586-025-09005-y
- International Energy Agency. Energy and AI (2025). https://www.iea.org/reports/energy-and-ai
- International Energy Agency. Key Questions on Energy and AI (2026). https://www.iea.org/reports/key-questions-on-energy-and-ai
- Autio, C. et al. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1 (2024; updated 2026). https://doi.org/10.6028/NIST.AI.600-1
- Tabassi, E. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1 (2023). https://doi.org/10.6028/NIST.AI.100-1
- European Commission. “Guidelines on obligations for General-Purpose AI providers.” https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers
- European Commission. “The enforcement framework of the AI Act.” https://digital-strategy.ec.europa.eu/en/policies/enforcement-ai-act
---
Editorial note: This article distinguishes measured capability from marketing claims and separates current evidence from forecasts. Benchmark results describe particular models, tasks and evaluation conditions; they should not be generalized automatically to every system or deployment. Legal and regulatory information reflects the position reviewed on 21 August 2026 and may change.