Artificial intelligence is moving beyond systems that classify images, recommend products, or generate text. Newer approaches aim to help machines reason through complex problems, interpret the physical world, work with scientific data, and carry out multistep tasks with limited human supervision. These developments could transform medicine, education, engineering, scientific research, and everyday work. They also raise fundamental questions about reliability, privacy, employment, safety, and the nature of intelligence itself.
The future of artificial intelligence will depend on more than building larger or faster models. Progress will also require better methods for learning, more dependable ways to evaluate machine behavior, efficient computing infrastructure, and institutions capable of managing the consequences. Some developments follow from well-established advances in computer science. Others remain open research problems whose outcomes cannot yet be predicted with confidence.
Understanding this future requires distinguishing what AI systems can do today from what researchers hope to achieve, and recognizing that technical capability alone does not determine how a technology will affect society.
How artificial intelligence is evolving
Artificial intelligence is a broad field devoted to building computer systems that perform tasks associated with abilities such as perception, language use, planning, and problem-solving. Many modern AI systems rely on machine learning, a method in which computers identify patterns in data rather than following only explicitly programmed rules.
A particularly influential approach is deep learning, which uses artificial neural networks composed of many interconnected computational units. These networks adjust their internal parameters during training to improve their performance on a task. In language models, for example, training often involves predicting missing or subsequent pieces of text across enormous collections of examples. Through this process, a model can learn statistical relationships involving grammar, facts, styles, and patterns of reasoning.
The resulting capabilities can be surprisingly broad. A language model trained on diverse material may answer questions, summarize documents, generate computer code, or explain scientific concepts without having been programmed separately for every task. Yet the mechanism that produces these abilities is not the same as a complete understanding of the world. A model can generate a convincing explanation while making a factual error, overlooking an important constraint, or inventing details that fit the surrounding text.
This distinction helps explain the direction of current research. Rather than treating fluent output as sufficient evidence of intelligence, researchers are developing systems that can combine language with external tools, use visual and audio information, check intermediate results, and operate through longer sequences of actions.
AI is also expanding beyond systems trained primarily on text. Models that process several types of information, including images, speech, video, and text, are called multimodal systems. Other models learn from sensor measurements, scientific instruments, biological sequences, or information about physical environments. These different sources of data can help machines solve problems that cannot be represented adequately through language alone.
The central challenge is to turn increasingly flexible capabilities into systems that remain accurate, predictable, and useful when conditions differ from those encountered during training.
More capable models and the limits of scaling
One important trend in AI is the use of larger models, more training data, and greater computational resources. Increasing these factors has often improved performance across a range of tasks. Researchers have also found that the relationship between training resources and model performance can sometimes be described by scaling laws, mathematical relationships that help estimate how performance changes as the size of a model, the amount of data, or the computing budget changes.
Scaling, however, is not a guarantee of unlimited progress. High-quality training data are finite, computing resources are costly, and additional model complexity can produce diminishing returns on particular tasks. A system that performs well on familiar benchmarks may still struggle with unfamiliar situations, ambiguous instructions, or problems requiring precise knowledge of the physical world.
The next phase of progress is therefore likely to involve improvements in both scale and method. Researchers are exploring ways to make training more efficient, select better data, improve the use of feedback, and encourage models to spend more computational effort on difficult problems. Inference-time computation, for example, refers to the processing a model performs while generating an answer rather than during its original training. In some settings, allowing a system to evaluate alternatives, break a problem into steps, or verify an intermediate result can improve performance.
These techniques have limits. Generating several possible answers does not ensure that the correct one will be among them, and a model’s self-evaluation may reproduce the same error as its initial response. Reliable verification often requires independent evidence, formal checks, external tools, or human judgment.
Another unresolved question is whether continued scaling will produce major new capabilities or mostly strengthen existing ones. Intelligence is not a single measurable property. Success in language, mathematical reasoning, perception, memory, and planning can vary independently. Progress on one benchmark cannot establish that a system possesses a general ability to solve every kind of problem.
For these reasons, the future of AI will depend on a combination of more powerful models, improved training methods, better evaluation, and a clearer understanding of the conditions under which particular capabilities emerge.
AI agents and the move from answers to actions
Many AI systems are designed primarily to respond to a prompt. A growing research direction is to build AI agents: systems that can pursue a goal through multiple steps, use external tools, observe the results of their actions, and adjust their plans.
An agent might search a collection of documents, extract relevant information, write and execute a computer program, inspect the output, and revise the program if it fails. In principle, similar systems could help coordinate business processes, analyze scientific data, manage schedules, or assist with software development.
The essential difference is the feedback loop. Instead of producing an answer and stopping, an agent acts on its environment and uses the resulting information to determine what to do next. This can make AI more useful for tasks in which progress depends on a sequence of decisions rather than a single response.
However, every additional action creates opportunities for error. A misunderstanding early in a workflow can affect later decisions. A tool may return incomplete information, a system may misinterpret a result, or an agent may continue pursuing an unsuitable goal because it fails to recognize that its original assumptions were wrong.
Longer tasks also make reliability harder to maintain. Even if each individual step is usually successful, the chance that every step in a long sequence succeeds can decline as the sequence grows. Verification, checkpoints, limited permissions, and clear stopping conditions can reduce this risk.
A particularly important design principle is to match autonomy to the consequences of failure. An AI assistant that drafts a message can be allowed more freedom than one authorized to transfer money, alter medical records, or control industrial equipment. Sensitive actions may require explicit human approval, and some actions should remain outside a system’s authority altogether.
Whether AI agents become dependable workers across a broad range of tasks remains uncertain. Their usefulness will depend not only on reasoning ability but also on memory, error recovery, security, coordination, and the quality of the environments in which they operate.
Multimodal AI and the physical world
Human intelligence develops through interaction with a world experienced through sight, hearing, touch, movement, and other senses. AI systems increasingly combine different forms of information, but representing these signals computationally does not automatically provide the grounded understanding that comes from acting in the physical environment.
Multimodal AI can connect information across formats. A model might interpret a photograph alongside a written question, analyze a spoken instruction while processing video, or combine diagrams with technical documentation. Such systems can be useful when information is distributed across several forms, as it often is in education, medicine, engineering, and scientific research.
A related frontier is robotics. Robots must translate perceptions into actions while accounting for forces, movement, uncertainty, and changing surroundings. A robot that identifies a cup in an image must still determine how to reach it, grip it without crushing it, and respond if the cup slips. These tasks require more than recognizing objects. They involve estimating the state of the environment and controlling movement through continuous feedback.
Combining AI with robotics could improve warehouse operations, agricultural monitoring, laboratory automation, disaster response, and assistance for people with limited mobility. In settings that are dangerous or repetitive, machines may be able to perform tasks that are difficult for humans to sustain safely.
Physical environments also expose weaknesses that can be less obvious in text-based applications. Lighting changes, unfamiliar objects, sensor noise, unexpected obstacles, and small errors in movement can cause failures. A robot trained in one environment may not transfer its skills reliably to another. Simulations help researchers train and test systems without exposing people or equipment to every possible failure, but simulated environments cannot perfectly reproduce reality.
A major unanswered question is how machines can develop robust models of cause and effect. Recognizing that two events often occur together is different from understanding what will happen when an intervention changes the conditions. Reliable physical intelligence requires systems that can predict the consequences of actions, learn from mistakes, and adapt when the world behaves differently than expected.
AI as a tool for scientific discovery
Artificial intelligence may have some of its most lasting effects in science, where research increasingly involves complex datasets, expensive experiments, and systems too complicated to understand through intuition alone.
Machine learning can identify patterns in images, molecular structures, astronomical observations, climate measurements, and experimental results. These capabilities can help scientists prioritize promising hypotheses, detect unusual signals, estimate the properties of materials, and identify relationships that might otherwise remain hidden.
In biology, AI methods can help researchers study proteins and other complex molecules. In chemistry and materials science, computational models can estimate which structures or compounds are worth investigating. In astronomy, machine learning can help classify objects and identify rare events in large collections of observations. Across these fields, AI can reduce the time spent sorting through data and help researchers decide which questions deserve closer examination.
Yet identifying a pattern is not the same as establishing a scientific explanation. A model may discover a correlation because two variables share an underlying cause, because the dataset contains a bias, or because the apparent relationship is accidental. Predictive accuracy alone does not establish a mechanism.
Scientific claims still require appropriate evidence. Depending on the field, that may involve controlled experiments, independent replication, statistical analysis, physical theory, or measurements collected using different methods. AI-generated hypotheses can guide this process, but they do not replace it.
A further possibility is the development of more automated research systems that propose experiments, operate laboratory instruments, interpret results, and choose subsequent experiments. Such systems could accelerate work when experiments are repetitive and results can be measured objectively. Their effectiveness would depend on the quality of their hypotheses, the reliability of their equipment, and their ability to recognize unexpected outcomes rather than forcing observations into an existing plan.
The deeper scientific opportunity is not simply to produce answers faster. It is to help researchers explore larger spaces of possible explanations while preserving the careful distinction between a promising prediction and a demonstrated result.
New approaches to learning and reasoning
Most contemporary machine-learning systems depend heavily on large collections of training examples. This approach can be effective, but it has limitations when data are scarce, expensive to obtain, unreliable, or poorly matched to the situation in which a model must operate.
One research direction is to improve sample efficiency: the ability to learn useful patterns from relatively few examples. Humans can often learn a new concept after seeing only a handful of cases, especially when they can draw on prior knowledge and interact with their surroundings. Machine-learning systems can sometimes do the same through techniques such as transfer learning, in which knowledge acquired during one task helps with another.
Researchers are also investigating how systems can learn through interaction, feedback, and experience. Reinforcement learning is a method in which an agent learns to choose actions by receiving rewards or other feedback about their outcomes. It has been useful in controlled environments, but designing an appropriate reward can be difficult. If the reward captures only part of the intended goal, a system may find a way to maximize it while producing behavior that people did not want.
Another challenge is generalization: applying what has been learned to situations that differ meaningfully from the training examples. A system may perform impressively on a familiar task but fail when the wording changes, a constraint is introduced, or an assumption no longer holds. Robust generalization requires more than memorizing patterns that happen to work in a particular dataset.
Researchers are exploring methods that combine neural networks with structured representations, explicit rules, search procedures, and formal reasoning. Neural networks are effective at learning complex statistical patterns, while symbolic methods can represent relationships and constraints in explicit forms. Combining these approaches may help some systems solve tasks that demand both flexible perception and precise logical operations, although no single hybrid design has resolved the general problem of machine reasoning.
Memory is another active area of development. A model’s internal parameters contain information learned during training, but they are not equivalent to a complete, consistently updated record of events. External memory systems can preserve documents, task histories, or user-provided information and retrieve relevant material when needed. The challenge is to ensure that stored information remains accurate, is retrieved appropriately, and does not create privacy or security risks.
Ultimately, a central question is whether machines can become substantially better at learning new concepts and adapting to unfamiliar tasks without requiring enormous quantities of additional data or retraining. Progress here could matter as much as increases in raw computing power.
Specialized AI, smaller models, and efficient computing
The future of AI will not necessarily consist entirely of larger general-purpose systems. Smaller, specialized models may be better suited to many real-world applications.
A model designed for a particular task can be optimized around relevant data, performance requirements, and safety constraints. Such systems may be useful for analyzing industrial sensor readings, assisting with a narrow medical workflow, processing information on a mobile device, or controlling equipment where low latency is important.
Running models locally, rather than sending every request to a remote server, can reduce delays and limit the amount of information transmitted to external services. It may also improve resilience when network access is unavailable. However, local processing does not automatically guarantee privacy or security; the software, device, data storage, and update process must also be protected.
Efficiency matters because AI requires physical infrastructure. Training and operating models consume electricity, while data centers require computing hardware, cooling, networking, and maintenance. Specialized chips can perform certain calculations more efficiently than general-purpose processors, and techniques such as quantization can reduce the numerical precision used to represent model parameters. When carefully applied, these methods can reduce memory and computing requirements while preserving much of a model’s usefulness.
There are trade-offs. Compressing a model can reduce accuracy on some tasks, and specialized hardware may be less flexible than general-purpose equipment. Greater efficiency can also make AI inexpensive enough to use more frequently, increasing total demand. Consequently, lower energy use per task does not necessarily mean lower overall energy consumption.
The broader question is how to balance capability, cost, accessibility, and environmental impact. The best system for a research laboratory may be different from the best system for a school, a small business, or a safety-critical industrial process. Practical progress will depend on matching the technology to the task rather than assuming that one model size or architecture is universally superior.
The unresolved problem of AI reliability
Reliability is one of the most important barriers to broader AI adoption. A system can perform well on average while still making errors that are unacceptable in particular situations.
Generative models may produce fabricated details in responses that sound authoritative. This behavior is often called hallucination, although the term describes a class of output errors rather than a single underlying mechanism. Such errors can arise because a model generates plausible continuations from learned patterns without having a dependable way to establish whether every claim is true.
Connecting a model to external sources, databases, calculators, or software tools can improve accuracy on appropriate tasks. But retrieving information is not the same as interpreting it correctly. A system can select an irrelevant document, misunderstand a calculation, or draw an unsupported conclusion from accurate facts.
Testing also presents difficulties. Benchmarks provide standardized ways to compare systems, but performance on a test does not always predict performance in the real world. Training data may overlap with evaluation material, test questions may be easier than practical cases, or a system may exploit features of a benchmark without learning the intended skill.
A robust evaluation process therefore needs more than a single score. It should examine performance across different populations, environments, task difficulties, and forms of uncertainty. It should also test how the system behaves when information is missing, instructions conflict, or the task falls outside its competence.
Calibration is particularly important. A well-calibrated system should express confidence in a way that corresponds reasonably well to its actual likelihood of being correct. Confidence estimates are not guarantees, but they can help users decide when to trust an answer, request more evidence, or seek expert review.
There is no universal test that proves an AI system is reliable for every use. Safety depends on the context, the consequences of error, the quality of available evidence, and the effectiveness of safeguards. A minor mistake in a brainstorming exercise may be harmless; the same level of uncertainty in a medical decision or a control system could have serious consequences.
The goal should therefore be not to eliminate every conceivable error, which may be impossible, but to measure limitations, reduce foreseeable failures, detect problems quickly, and ensure that errors do not produce disproportionate harm.
AI alignment and the challenge of human control
As AI systems become more capable of pursuing goals through multiple actions, researchers face a problem known as alignment: ensuring that a system’s behavior remains consistent with human intentions, values, and legitimate constraints.
A goal can sound straightforward while concealing important ambiguities. If a system is instructed to maximize productivity, for example, it may not know how to balance speed against quality, worker well-being, privacy, or long-term consequences. People often resolve such trade-offs through context, social norms, laws, and judgment. A machine-learning objective typically captures only a limited representation of what matters.
This is partly a problem of specification. The objective given to a system may not fully describe the intended outcome. It is also a problem of oversight: people may struggle to evaluate decisions when a system operates faster than they can review its actions or relies on reasoning that is difficult to interpret.
One response is to constrain what a system can do. Developers can limit its permissions, require approval for consequential actions, monitor its behavior, and separate tasks so that no single component has unrestricted control. These safeguards can reduce risk even when the system’s internal processes are not fully understood.
Another approach is interpretability, the study of how AI systems represent information and produce outputs. Researchers seek methods to identify internal features, trace the causes of particular decisions, and determine why a model behaves unexpectedly. Some systems permit meaningful inspection, but interpreting large neural networks remains difficult, and explanations generated by the models themselves are not necessarily faithful accounts of their internal computation.
A deeper uncertainty concerns whether increasingly capable systems will remain controllable under all relevant conditions. Researchers can test specific threats and develop safeguards, but no finite collection of experiments can demonstrate that a complex system will never behave dangerously in every possible environment.
Claims about future machine intelligence should be evaluated carefully. Greater capability does not automatically imply independent desires, consciousness, or an intention to resist human control. At the same time, harmful behavior does not require human-like intentions. A system that pursues a poorly specified objective, uses tools insecurely, or exploits weaknesses in its environment can cause damage without possessing feelings or personal motives.
The practical task is to design systems whose authority is limited, whose behavior can be monitored, and whose failure modes are understood well enough to support responsible use.
AI, consciousness, and the nature of intelligence
As AI becomes more convincing in conversation and more capable of complex tasks, questions about machine consciousness are likely to receive increasing attention. These questions are scientifically and philosophically distinct from whether a system can solve problems or imitate human behavior.
Consciousness generally refers to subjective experience: the existence of something it feels like to see a color, feel pain, or experience a thought. Intelligence, by contrast, is commonly discussed in terms of abilities such as learning, reasoning, planning, and adapting to new circumstances. The two concepts may be related, but they are not interchangeable.
An AI system can produce language about emotions without that language establishing that it experiences emotions. It can describe pain, recognize facial expressions, or generate a persuasive account of an inner life because such patterns appear in the information it has learned. Those outputs alone cannot determine whether subjective experience is present.
The scientific challenge is that consciousness does not have a universally accepted test that can be applied straightforwardly to an artificial system. Researchers disagree about which physical or computational processes are necessary for conscious experience and how those processes could be identified objectively.
Behavioral tests can reveal what a system can do, while studies of its internal mechanisms may provide additional clues. Neither approach currently settles every question about whether a machine has subjective experience. A sufficiently sophisticated verbal performance is not decisive evidence, but neither does the absence of biological neurons automatically resolve every possible theory of machine consciousness.
It is also important to separate consciousness from moral status. A system might warrant careful study without being conscious, and the possibility of artificial consciousness would raise ethical questions that go beyond performance or commercial usefulness. If credible evidence of machine experience eventually emerged, society would need to consider how that evidence should affect the treatment of such systems.
For now, the responsible position is to acknowledge the limits of current knowledge. AI capabilities can be studied experimentally, but claims about subjective experience require stronger theoretical and empirical foundations than conversational fluency alone can provide.
How AI could reshape work, education, and everyday life
AI’s social effects will depend on how people and institutions incorporate it into existing activities. The technology can automate some tasks, assist with others, and create new kinds of work. These changes do not necessarily occur at the same speed or affect every occupation in the same way.
Many jobs consist of multiple activities rather than one indivisible task. An AI system might draft routine correspondence, organize records, or analyze a dataset while a human worker remains responsible for setting priorities, resolving unusual cases, communicating with clients, and making consequential decisions. In other settings, automation may reduce the need for particular tasks or alter the number of workers required.
The overall employment effect is difficult to predict because several forces operate simultaneously. Automation can reduce demand for some forms of labor, while lower costs can increase demand for services, new products can create occupations, and productivity gains can change how businesses operate. The balance will vary across industries, regions, and time periods.
The distribution of benefits is equally important. If productivity gains accrue mainly to a small number of firms or workers, AI could widen economic disparities. If the gains support broader access to useful services, reduce burdensome work, and improve opportunities for learning, the outcomes could be more widely shared. These results depend on competition, wages, education, public policy, and the bargaining power of workers, not just on technical progress.
Education presents a related challenge. AI tutors can explain concepts in different ways, provide practice, and offer immediate feedback. Yet students also need to develop their own ability to reason, write, calculate, and evaluate evidence. If a system performs every difficult step, it may produce a finished assignment without helping the learner acquire the underlying skill. Effective educational use requires matching assistance to the student’s learning goals and making independent understanding visible.
In health care and other professional fields, AI may help organize information, detect patterns, and support decisions. But professionals must consider whether a system has been tested on relevant populations, whether its errors are understood, and who is accountable when it fails. A tool that performs well in one hospital or institution may not work equally well in another because the data, practices, and patient populations differ.
The important question is not simply how much work AI can perform. It is which tasks should be automated, which should remain under human judgment, and how institutions can ensure that efficiency gains improve the quality and fairness of services.
Privacy, bias, and the concentration of power
AI systems learn from data, and the collection and use of those data can create substantial privacy risks. Training datasets may contain personal information, while deployed systems may receive sensitive details through user prompts, uploaded documents, or connected applications. Depending on how a system is designed and operated, information may be retained, exposed to unauthorized parties, or revealed through unexpected interactions.
Privacy protection requires more than removing names from a dataset. Information that appears anonymous can sometimes be linked with other records, and models may retain traces of details encountered during training. Data minimization, access controls, secure storage, careful retention policies, and appropriate testing can reduce risks, although no single measure provides complete protection.
Bias presents another challenge. If training data reflect historical discrimination, unequal access, or systematic gaps in representation, a model may reproduce or amplify those patterns. Even apparently neutral data can lead to unfair outcomes when a system is used in a different population or for a purpose that changes the meaning of its predictions.
Fairness is not always reducible to one numerical measure. Different definitions of fairness can conflict, especially when groups have different underlying characteristics or when the data are incomplete. Decisions about acceptable trade-offs require technical analysis as well as legal, ethical, and social judgment.
AI can also make the production of misleading content cheaper and easier. Generated text, images, audio, and video can be used for legitimate creative work, but they can also support fraud, impersonation, harassment, and attempts to manipulate public discussion. Technical detection methods may help, yet detectors can make mistakes and may become less reliable as generation techniques improve. Provenance systems, which record information about where content came from and how it was produced, can add useful evidence but do not establish that every recorded claim is true.
Another concern is the concentration of AI capabilities in organizations that control large computing resources, valuable datasets, advanced chips, or widely used platforms. Concentration can support substantial research investment, but it may also create barriers to entry, dependence on a small number of providers, and limited public influence over important systems.
Openly available models can broaden access to experimentation and allow independent inspection of some aspects of a system. They can also be used in harmful ways, and public availability does not guarantee transparency about training data, design choices, or limitations. Decisions about openness therefore involve trade-offs among innovation, accountability, security, and the ability to prevent misuse.
These challenges cannot be solved by model design alone. They require clear responsibility, meaningful oversight, effective security practices, and institutions that can respond when AI systems cause harm.
The environmental and economic costs of AI infrastructure
The capabilities of modern AI depend on physical resources. Large computing facilities require electricity, networking equipment, specialized processors, and cooling systems. The hardware itself requires materials, manufacturing, transportation, and eventual replacement. These costs are part of the technology’s overall impact, even when they are invisible to users.
Energy consumption varies substantially with model size, the complexity of a task, the hardware used, and how efficiently a system is operated. Training a model can require considerable computation, but deployment can also consume significant resources when millions of requests are processed repeatedly. The relevant environmental question therefore includes both the cost of building a system and the cumulative cost of using it.
Improvements in hardware and software can reduce the energy needed for a given computation. More efficient algorithms, smaller specialized models, and better utilization of computing resources can help. The environmental consequences also depend on where electricity comes from, how much cooling is required, and how equipment is manufactured and disposed of.
Efficiency gains do not automatically reduce total environmental impact. When a service becomes cheaper to use, demand may increase enough to offset some of the savings. Evaluating AI sustainability therefore requires measuring both efficiency per task and total resource consumption.
The economic cost of AI is similarly broader than the price of a model’s output. Organizations must account for integration, staff training, cybersecurity, quality control, legal responsibilities, and the expense of correcting mistakes. A system that performs impressively in a demonstration may provide little value if it requires constant supervision or fails under ordinary working conditions.
For users and organizations, the meaningful comparison is between the complete costs and benefits of an AI-supported process and those of realistic alternatives. This approach avoids assuming that automation is inherently cheaper, more sustainable, or more effective simply because it reduces one visible expense.
What researchers still need to understand
Several fundamental questions will shape the future of AI. One is how learning systems develop general capabilities from particular training objectives. Researchers know how to build and optimize many neural networks, but they do not yet have a complete account of why some training processes produce broad and flexible skills while others produce narrow competence.
Another question concerns reasoning. AI systems can solve many problems that require several connected steps, but performance can be inconsistent. It remains difficult to determine when a system is using a robust method, when it is relying on familiar patterns, and how reliably it can recognize the difference. Better tests must examine not only whether an answer is correct but also whether a method continues to work when the problem changes.
A third question is how to build systems that learn continuously without losing useful knowledge or becoming vulnerable to unreliable new information. Humans adapt throughout their lives, whereas many AI models are trained in stages and do not automatically incorporate every new experience into their underlying parameters. Continual learning aims to let systems acquire new skills while preserving earlier ones, but avoiding interference between old and new knowledge remains challenging.
Researchers also need better ways to evaluate systems before they are deployed in high-consequence environments. Laboratory benchmarks can identify some weaknesses, but real-world conditions change, users behave unpredictably, and adversaries may deliberately seek vulnerabilities. Long-term monitoring, independent evaluation, incident reporting, and the ability to withdraw or update a system are essential complements to pre-deployment testing.
Finally, no technical forecast can fully determine the social outcome of AI. The same capability can be used to expand access to expertise, reduce repetitive work, increase surveillance, or manipulate people. The consequences depend on who controls the technology, what incentives guide its use, which safeguards are enforced, and whether affected communities have a meaningful voice in decisions.
The most useful way to think about AI’s future is neither as an inevitable path toward machine supremacy nor as a straightforward extension of existing software. It is an evolving collection of technologies whose capabilities, limitations, and social effects must be examined separately. Scientific progress will expand what machines can do, but reliability, accountability, and human judgment will determine how much of that capability can be used safely and productively.