Machine and hybrid intelligence represent two interconnected approaches to understanding, constructing, and extending intelligent capabilities through computation. Machine intelligence concerns the capacity of artificial systems to process information, learn from data or experience, reason over representations, solve problems, and pursue specified objectives. Hybrid intelligence concerns arrangements in which complementary forms of intelligence are integrated to achieve capabilities or outcomes that may be difficult for individual components to attain independently.
These concepts occupy a central position in artificial intelligence research because they connect computational algorithms with broader questions about cognition, learning, agency, cooperation, human judgment, and the relationship between biological and technological systems.
Machine intelligence can emerge from symbolic reasoning, statistical inference, machine learning, artificial neural networks, reinforcement learning, search algorithms, and combinations of these methods. Its capabilities depend on the architecture, training data, computational resources, environmental information, and objectives of the system.
Hybrid intelligence introduces an additional integration problem: how should different cognitive or computational capabilities interact, exchange information, allocate responsibilities, manage uncertainty, and jointly produce reliable outcomes?
In human–AI systems, this may involve combining human expertise, contextual understanding, and decision authority with artificial systems capable of large-scale information processing, prediction, pattern recognition, or automated reasoning.
In hybrid AI architectures, integration may instead occur entirely among computational approaches, such as neural networks and symbolic logic.
In brain–computer interfaces, biological neural activity and external computational systems are connected through signal-processing and control mechanisms.
These are related but technically distinct meanings of hybrid intelligence.
A machine that independently solves a difficult problem is not necessarily a hybrid intelligent system. Similarly, combining a human and an AI system does not automatically produce superior intelligence or performance.
The central scientific challenge is therefore not simply to maximize computational capability, but to understand how different forms of intelligence can be represented, coordinated, evaluated, and integrated while preserving reliability, human agency, and appropriate operational constraints.
1. DEEP DEFINITION
1.1. What Is Machine Intelligence?
Machine intelligence refers to the ability of an artificial computational system to perform activities involving information processing, learning, inference, problem-solving, planning, perception, or goal-directed decision-making in ways that satisfy defined performance criteria.
It is an operational concept rather than a universally agreed measure of machine consciousness or human-equivalent understanding.
Machine intelligence may be implemented through:
- Symbolic knowledge representation and logical inference.
- Search algorithms and automated planning.
- Statistical learning and probabilistic reasoning.
- Artificial neural networks and deep learning.
- Reinforcement learning and sequential decision-making.
- Computer vision and multimodal perception.
- Natural-language processing and generation.
- Knowledge retrieval and information synthesis.
- Adaptive control and autonomous action selection.
- Combinations of specialized computational architectures.
The common feature is not a particular algorithm but the system’s ability to produce task-relevant behaviour through computational processes.
A language model may generate a technically sophisticated explanation while possessing no independent authority to act on the external world.
A robotics controller may operate autonomously within a defined environment while lacking the ability to answer general questions.
A classification model may perform a narrow perceptual task with high accuracy without possessing broad reasoning capabilities.
Consequently, machine intelligence should be assessed in relation to specific tasks, environments, generalization requirements, and evaluation criteria [5].
1.2. A Formal Representation of Machine Intelligence
An artificial system can be modeled as a decision process that maps information about its environment to outputs or actions.
Let:
- X represent the input or observation space.
- Y represent the output space.
- A represent available actions.
- θ represent learned or configured parameters.
- G represent the task objective.
- πθ represent the decision policy.
For a prediction task:
ŷ = fθ(x)
Where:
- x is an input.
- fθ is the computational model.
- ŷ is the predicted output.
For a goal-directed sequential task:
aₜ ∼ πθ(· | hₜ, G)
Where:
- aₜ represents the selected action.
- hₜ represents available interaction history.
- G represents the applicable goal specification.
This distinction is important.
A predictive model generates an estimate or output. An autonomous agent may use predictions to select actions that alter an environment.
Machine intelligence can support both functions, but prediction and autonomous execution should not be treated as equivalent capabilities.
1.3. What Is Hybrid Intelligence?
In the human–AI research literature, hybrid intelligence commonly refers to systems that combine human and artificial intelligence to accomplish complex objectives through complementary contributions.
Dellermann and colleagues developed an influential formulation emphasizing the integration of human and artificial intelligence, superior joint performance, and opportunities for mutual learning [10].
Under this interpretation, a hybrid intelligent system is more than a conventional software application with a human operator.
Its architecture intentionally coordinates human and machine contributions to the same broader cognitive or operational process.
Examples include:
Medical decision support: An AI system analyzes medical images while a qualified clinician integrates clinical context and makes the appropriate professional decision.
Scientific research: Computational systems retrieve, process, and compare evidence while researchers evaluate hypotheses, methodology, and scientific significance.
Software engineering: AI tools generate candidate solutions while developers validate architecture, security, maintainability, and correctness.
Education: Adaptive systems recommend learning activities while educators establish instructional goals and respond to learners’ circumstances.
Engineering design: Computational optimization proposes candidate designs while human specialists evaluate requirements, feasibility, and consequences.
The intended benefit arises from the relationship between different capabilities.
However, genuine improvement must be empirically demonstrated rather than assumed.
1.4. Two Important Meanings of Hybrid Intelligence
The expression hybrid intelligence is used differently across research traditions.
Human–AI hybrid intelligence
This concerns sociotechnical systems combining biological human cognition with artificial computational capabilities.
Relevant mechanisms include task delegation, complementary expertise, interactive learning, oversight, shared decision-making, and joint performance evaluation.
Computational hybrid intelligence
This concerns AI architectures combining different algorithmic approaches.
Examples include:
- Neural-symbolic systems.
- Neuro-fuzzy systems.
- Probabilistic logic and neural learning.
- Neural networks integrated with search or planning.
- Evolutionary optimization combined with machine learning.
- Model-based and model-free reinforcement learning.
A computational hybrid system may contain no human participant during inference.
Conversely, a human–AI system may rely on a single machine-learning architecture rather than combining different computational paradigms.
Both forms involve integration, but the entities being integrated are different.
The distinction prevents confusion between a machine combining algorithms and an organizational system combining human and artificial cognition.
1.5. Machine Intelligence Versus Hybrid Intelligence
| Dimension | Machine intelligence | Human–AI hybrid intelligence | Computational hybrid AI |
|---|---|---|---|
| Main components | Artificial computational system | Humans and AI systems | Multiple AI methods or models |
| Primary objective | Perform defined cognitive or adaptive tasks | Coordinate complementary human and machine capabilities | Combine complementary computational methods |
| Learning | Algorithmic updates or programmed adaptation | Machine learning plus possible human learning and feedback | Learning across integrated algorithms |
| Decision authority | Determined by system permissions | Distributed according to explicit human–AI roles | Determined by the enclosing application |
| Examples | Image classifier, language model, planning agent | Clinician with AI decision support | Neural-symbolic reasoning system |
| Main challenge | Capability, reliability, and generalization | Coordination, trust, authority, and complementarity | Integration consistency and computational efficiency |
| Evaluation | Task-specific performance | Joint performance compared with standalone baselines | Combined system compared with component baselines |
These categories are not mutually exclusive.
A human–AI hybrid system may use several computationally hybrid models internally.
A multi-agent AI system may coordinate multiple machine intelligences without directly incorporating human cognitive contributions during execution.
1.6. Complementarity and the Meaning of Synergy
Complementarity describes circumstances in which different participants possess capabilities that can improve one another’s performance.
A human expert may recognize an unusual contextual factor that was poorly represented in an AI model’s training data.
An AI system may identify a complex statistical pattern that would be difficult for a human to detect unaided.
When these contributions are integrated effectively, the combined system may outperform either component.
However, two distinct outcomes must be separated.
Human augmentation occurs when the combined system outperforms the human working independently.
Human–AI synergy occurs when the combined system outperforms both the human and the AI working independently.
Let:
- J(H) represent human-alone performance.
- J(M) represent machine-alone performance.
- J(HM) represent hybrid-system performance.
For a metric in which larger values indicate better performance:
Augmentation = J(HM) − J(H)
Synergy = J(HM) − max[J(H), J(M)]
A positive augmentation value does not necessarily imply positive synergy.
For example:
- Human-alone accuracy: 82%.
- Machine-alone accuracy: 91%.
- Human–AI accuracy: 87%.
The hybrid system improves human performance by five percentage points.
However, it performs four percentage points worse than the machine alone.
Therefore, it demonstrates augmentation relative to the human baseline but not positive synergy.
This distinction is essential because improving a human user’s performance is not the same as developing a system that exceeds the strongest available standalone alternative.
1.7. The Biological and Computational Foundations of Hybrid Intelligence
Biological cognition and artificial computation share some functional characteristics, including information processing, adaptive behaviour, memory, and decision-making.
However, they differ in their implementation, development, constraints, and relationship to the environment.
Biological intelligence emerges through living nervous systems, bodily processes, development, evolution, and social experience.
Artificial intelligence is implemented through engineered computational representations, algorithms, hardware, training procedures, and software-controlled interfaces.
These differences can create opportunities for complementary performance.
Humans can contribute task interpretation, situational experience, social understanding, and responsibility for decisions.
Machines can contribute reproducible computation, rapid numerical operations, large-scale pattern analysis, and consistent application of specified algorithms.
Nevertheless, these are broad functional tendencies rather than absolute boundaries.
Humans can outperform machines in some computationally demanding situations, while AI systems may exceed human performance in particular classification, search, or optimization tasks.
Similarly, human decisions can be biased or inconsistent, while computational systems can exhibit systematic errors and misleading confidence.
The objective of hybrid intelligence is not to assume that either participant possesses universally superior capabilities.
It is to identify, measure, and coordinate the strengths and limitations of each.
1.8. Intelligence Is Not Equivalent to Consciousness
Machine intelligence is principally evaluated through capabilities and observable performance.
A system may learn patterns, solve problems, generate explanations, or select actions without those observations establishing subjective experience.
Likewise, connecting biological neural signals to computational devices does not necessarily produce a unified biological–machine consciousness.
The terms machine intelligence, hybrid intelligence, artificial general intelligence, and machine consciousness therefore refer to different concepts.
A highly capable computational system may remain narrowly specialized.
A hybrid system may improve task performance without possessing broad general intelligence.
A brain–computer interface may restore communication or control capabilities without creating a separate conscious agent.
Claims about consciousness require evidence and theoretical analysis beyond successful computational performance.
1.9. Historical Foundations of Machine and Hybrid Intelligence
The development of machine and hybrid intelligence reflects the convergence of mathematical logic, neuroscience, computer science, cybernetics, cognitive science, and human–computer interaction.
1943 — Mathematical neural computation
Warren McCulloch and Walter Pitts introduced a formal model of simplified neural activity using logical operations.
Their work contributed to foundational thinking about how networks of computational units might represent logical functions [1].
1950 — Machine intelligence and behavioural evaluation
Alan Turing’s Computing Machinery and Intelligence explored the question of machine intelligence through computational behaviour and the imitation game [2].
The work helped establish the scientific importance of examining machine capabilities through observable performance.
1960 — Human–computer symbiosis
J. C. R. Licklider proposed Man-Computer Symbiosis, envisioning close cooperation between humans and computers in complex intellectual activities [3].
His framework emphasized the possibility of combining human judgment with computational capabilities.
1962 — Augmentation of human intellect
Douglas Engelbart’s Augmenting Human Intellect: A Conceptual Framework proposed a systematic approach to improving human problem-solving through tools, representations, methods, and computer-supported interaction [4].
His work helped establish a foundational tradition in human–computer augmentation.
1986 — Learning internal representations
Rumelhart, Hinton, and Williams demonstrated how backpropagation could adjust neural-network parameters to learn useful internal representations [6].
This contributed to the development of modern connectionist machine learning.
2000 — Levels of human interaction with automation
Parasuraman, Sheridan, and Wickens developed a framework for differentiating automation across information acquisition, analysis, decision selection, and action implementation [12].
Their work remains relevant to the allocation of authority in hybrid systems.
2002–2012 — Brain–computer interfaces
Research demonstrated increasingly sophisticated mechanisms for converting neural activity into communication and control signals.
Clinical and experimental work, including neurally controlled robotic arms, established important examples of direct biological–computational integration [17][18].
2015 — Deep learning synthesis
LeCun, Bengio, and Hinton synthesized developments in multilayer neural networks and representation learning, documenting their importance in computer vision, speech processing, and other fields [7].
2017 — Transformer architectures
Vaswani and colleagues introduced the Transformer architecture, an influential foundation for subsequent large-scale language and multimodal models [8].
2019 — Contemporary hybrid intelligence frameworks
Dellermann and colleagues articulated hybrid intelligence as an integrated human–AI approach to solving complex problems [10].
The same year, Amershi and colleagues presented empirically informed guidelines for designing human–AI interactions [13].
2020–2021 — Selective prediction and complementary teams
Research investigated how machine-learning systems can defer decisions to human experts and how explanations affect human reliance on AI recommendations [14][15].
2024 — Empirical evidence on human–AI combinations
Vaccaro, Almaatouq, and Malone published a systematic review and meta-analysis examining when human–AI combinations improve performance [16].
Their findings challenged the assumption that combining human and artificial intelligence automatically creates superior outcomes.
These developments establish machine and hybrid intelligence as connected but distinct fields: one studies computationally implemented intelligent capabilities, while the other investigates how different capabilities can be integrated into effective systems.
2. TECHNICAL MECHANICS
2.1. Machine Intelligence as a Computational Architecture
A machine-intelligence system can be organized as a sequence of interacting computational functions.
A general architecture includes:
- Input acquisition.
- Data representation.
- Perception or feature extraction.
- Memory or learned parameter access.
- Inference and reasoning.
- Planning or decision selection.
- Output or action execution.
- Feedback and evaluation.
The exact implementation depends on the system.
A classification model may require only input processing and prediction.
A reasoning agent may combine retrieval, symbolic operations, language modeling, and tool execution.
An autonomous robot may integrate sensors, state estimation, planning, physical control, and safety mechanisms.
Therefore, a machine-intelligence architecture should be selected according to task requirements rather than the assumption that every intelligent system needs the same computational components.
2.2. Symbolic Machine Intelligence
Symbolic AI represents knowledge through explicit structures such as facts, rules, predicates, logical expressions, and graphs.
For example:
Human(x) → Mortal(x)
Human(Socrates)
From these premises, a suitable inference system can derive:
Mortal(Socrates)
Symbolic approaches are particularly useful when tasks involve explicit constraints, formal knowledge, theorem proving, or interpretable rules.
They can also support automated planning and verification.
However, symbolic systems may struggle when relevant concepts are difficult to formalize or when environmental information is incomplete and noisy.
A rule-based system does not automatically acquire robust perceptual capabilities merely because its inference procedure is logically valid.
This motivates integrating symbolic methods with statistical learning or neural representations.
2.3. Statistical Machine Learning
Statistical machine learning uses data to estimate predictive relationships or decision functions.
In supervised learning, a model receives examples containing inputs and target outputs.
Given a dataset:
D = {(x₁, y₁), (x₂, y₂), …, (xₙ, yₙ)}
The objective is to identify parameters θ that minimize a selected loss function.
θ = argminθ [(1/n)ΣᵢL(fθ(xᵢ), yᵢ) + λΩ(θ)]*
Where:
- fθ is the prediction model.
- L is a loss function.
- Ω(θ) is a regularization term.
- λ determines the regularization strength.
The model’s performance depends on the representativeness of its training data, the learning algorithm, the loss function, model capacity, and deployment conditions.
A low training error does not guarantee reliable performance on unfamiliar inputs.
This distinction is especially important when machine predictions are incorporated into human decisions.
2.4. Neural Networks and Representation Learning
Artificial neural networks transform inputs through layers of parameterized operations.
For a simplified neural layer:
h⁽ˡ⁺¹⁾ = σ(W⁽ˡ⁾h⁽ˡ⁾ + b⁽ˡ⁾)
Where:
- h⁽ˡ⁾ is the input representation at layer l.
- W⁽ˡ⁾ is a weight matrix.
- b⁽ˡ⁾ is a bias vector.
- σ is an activation function.
Multiple layers allow a model to learn increasingly complex representations.
Backpropagation provides an efficient method for computing gradients of differentiable loss functions through computational graphs [6].
A basic gradient-based parameter update is:
θₜ₊₁ = θₜ − η∇θL(θₜ)
Where η is the learning rate.
Deep learning systems can be highly effective for visual recognition, language processing, speech analysis, and other tasks [7].
However, their learned representations do not automatically guarantee faithful explanation, causal understanding, robustness, or compliance with external requirements.
2.5. Transformers and Foundation Models
Transformer architectures use attention mechanisms to model relationships among elements of an input sequence.
A standard scaled dot-product attention operation is:
Attention(Q, K, V) = softmax(QKᵀ/√dₖ)V
Where:
- Q represents queries.
- K represents keys.
- V represents values.
- dₖ is the key-vector dimension.
This mechanism allows representations to incorporate information from different input positions.
Transformers contributed significantly to the development of modern foundation models [8].
Such models can support a wide variety of language and multimodal tasks.
When integrated with retrieval systems, external tools, planning procedures, and permission controls, they may serve as components of larger intelligent agents.
Nevertheless, a foundation model alone does not constitute a complete, reliable autonomous system.
System-level intelligence also depends on information access, context management, external validation, execution controls, and recovery mechanisms.
2.6. Reinforcement Learning and Adaptive Decisions
Reinforcement learning studies how agents improve action-selection policies through interaction with an environment.
The objective is commonly to maximize expected cumulative reward.
J(π) = Eπ[Σₜ₌₀ᵀ γᵗrₜ]
Where:
- π is the policy.
- rₜ is the reward.
- γ is the discount factor.
- T is the horizon for a finite-horizon formulation.
A learned policy may be applied to navigation, resource allocation, robotics, or sequential decision-making.
However, optimizing a reward function does not guarantee alignment with every human requirement.
Hybrid systems must therefore distinguish algorithmic optimization objectives from legitimate human goals, safety constraints, and execution permissions [9].
2.7. Hybrid Intelligence as a Joint Decision System
A human–AI system contains at least two interacting decision-making components:
- A human participant or group.
- An artificial computational component.
The machine may generate a recommendation:
m = fθ(x)
The human may then produce a final decision:
y = g(x, m, c)
Where:
- x represents task information.
- m is the machine recommendation.
- c represents additional human context, experience, or information.
- g describes the joint decision process.
This formulation captures an important property of hybrid intelligence.
The combined result depends not only on machine accuracy but also on how the human interprets, trusts, challenges, or modifies the recommendation.
An accurate model can produce poor joint outcomes if users systematically reject its correct advice.
Conversely, users may perform worse if they accept incorrect recommendations too readily.
Therefore, optimizing machine accuracy alone is insufficient for optimizing hybrid-system performance.
2.8. Task Allocation Between Humans and Machines
A hybrid system must determine which participant should perform each task.
Let T represent the set of available tasks and P the set of participants.
An allocation function may be expressed as:
a: T → P
For a system with humans and machines, allocation decisions may depend on:
- Expected accuracy.
- Relevant expertise.
- Task complexity.
- Decision latency.
- Computational cost.
- Human workload.
- Uncertainty.
- Safety consequences.
- Required authority.
- Availability of appropriate evidence.
A conceptual allocation objective is:
minₐ Σₜ [E(C_error(t, a(t))) + λC_time(t, a(t)) + μC_resource(t, a(t))]
Subject to applicable authorization and safety constraints.
Here, λ and μ specify the relative importance of time and resources.
This is an illustrative optimization model rather than a universal allocation algorithm.
Some tasks require human approval regardless of whether the machine is expected to perform them more accurately.
Task delegation must therefore account for legitimate authority as well as predicted performance.
2.9. Learning to Defer
Selective prediction allows an AI system to produce an answer for some inputs while deferring others to a human expert.
Let:
- f(x) represent the model’s prediction.
- h(x) represent a human expert’s decision.
- r(x) represent a deferral decision.
Define:
r(x) = 0: Use the machine prediction.
r(x) = 1: Defer to the human expert.
The combined decision is:
ŷ = f(x), if r(x) = 0
ŷ = h(x), if r(x) = 1
A simplified expected-risk objective is:
R = E[(1 − r(x))L(f(x), y) + r(x)L(h(x), y)]
This can be extended to include deferral costs, time constraints, or other operational consequences.
Mozannar and Sontag investigated statistically consistent methods for learning when a classifier should defer to an expert [14].
Their work illustrates a fundamental design principle: hybrid intelligence can require learning not only what prediction to make but also when the machine should avoid making the final decision.
2.10. Confidence, Uncertainty, and Calibration
An AI model may output a probability or confidence estimate.
For a classifier:
p̂ = Pθ(Y = 1 | X = x)
If such estimates are well calibrated, predictions assigned approximately 80% confidence should be correct about 80% of the time over an appropriate collection of comparable predictions.
However, confidence values may be miscalibrated.
Model outputs can appear confident even when predictions are incorrect or inputs differ from training conditions.
A hybrid decision system should therefore evaluate:
Predictive uncertainty: How uncertain the model is about its output.
Data uncertainty: Whether relevant input information is incomplete or noisy.
Distributional uncertainty: Whether an input differs materially from familiar training conditions.
Human uncertainty: Whether the human possesses reliable knowledge for the relevant task.
System uncertainty: Whether integrating the available contributions will improve the outcome.
Calibration alone does not guarantee reliability under distribution shift.
A system may require uncertainty thresholds, input-validity checks, monitoring, and human escalation.
2.11. Trust Calibration in Human–AI Systems
Trust calibration concerns whether human reliance on an AI system corresponds appropriately to its actual competence.
Overtrust occurs when users rely on the system beyond its demonstrated capabilities.
Undertrust occurs when users reject reliable assistance without adequate reason.
Appropriate reliance requires accurate information about the system’s strengths, limitations, uncertainty, and expected performance.
Explanations may help users understand recommendations, but they do not automatically improve decision quality.
Bansal and colleagues examined whether AI explanations improve complementary human–AI team performance [15].
Their findings indicated that explanations could increase acceptance of AI recommendations without necessarily producing the desired improvement in joint performance.
This demonstrates the difference between making an AI recommendation persuasive and making it useful.
A trustworthy interface should support informed evaluation rather than merely increasing user agreement with machine outputs.
2.12. Human-in-the-Loop, Human-on-the-Loop, and Human Oversight
Hybrid systems can distribute authority across different functions.
Human-in-the-loop: A human participates directly in designated decisions or approval stages.
Human-on-the-loop: A system operates under defined authority while a human supervises and can intervene.
Human-directed operation: Humans establish goals and initiate tasks while computational systems perform delegated functions.
Machine-led assistance: Computational systems identify possible actions or decisions, but human authority remains defined by the application.
These arrangements can be applied differently to perception, analysis, planning, decision selection, and execution.
Parasuraman and colleagues demonstrated the importance of analyzing automation at the functional level rather than describing entire systems using only one autonomy label [12].
A system may autonomously retrieve data while requiring human authorization before transmitting confidential information.
It may independently propose software modifications but require approval before deployment.
Meaningful oversight requires appropriate information, intervention mechanisms, and sufficient time to respond.
2.13. Interactive Learning From Human Feedback
Human feedback can improve machine behaviour through annotation, correction, demonstration, ranking, or preference evaluation.
A supervised feedback cycle may include:
- The machine generates an output.
- A human evaluates the output.
- Feedback is recorded.
- Relevant examples are added to a training or evaluation dataset.
- The model is updated or its operation is adjusted.
- Revised performance is assessed on appropriate held-out data.
Human feedback can also influence retrieval rules, user preferences, system prompts, or decision policies without changing model parameters.
These mechanisms should be distinguished.
Correcting a generated answer in a user interface does not necessarily mean that the underlying model has learned from that interaction.
Likewise, storing a user preference is not the same as retraining a neural network.
A well-designed hybrid system should communicate which forms of adaptation actually occur.
2.14. Neural-Symbolic Hybrid Intelligence
Neural-symbolic AI combines the statistical representation-learning capabilities of neural networks with symbolic methods for representing and manipulating knowledge.
A simplified system may include:
Neural component: Converts raw data into structured predictions or representations.
Symbolic component: Applies explicit relationships, logical constraints, or inference rules.
Integration component: Transfers information between the neural and symbolic processes.
Verification component: Evaluates whether outputs satisfy the required conditions.
For example, an intelligent document-processing system might use a neural model to extract entities from text and a symbolic rule engine to check consistency against domain requirements.
A conceptual objective is:
L_total = L_data + λL_constraints
Where:
- L_data measures predictive error.
- L_constraints measures violations of selected symbolic constraints.
- λ controls their relative weight.
This is one possible integration approach, not a universal architecture.
Marra and colleagues surveyed relationships among neural-symbolic and statistical-relational approaches to learning and reasoning [11].
Hybrid computational systems may offer stronger structural consistency for some tasks, but their advantages depend on the quality of the symbolic representation and the integration method.
A correct rule engine cannot repair every error originating from an incorrect neural interpretation.
2.15. Neuro-Fuzzy and Probabilistic Hybrid Systems
Some hybrid AI systems combine neural networks with fuzzy logic or probabilistic reasoning.
Fuzzy logic represents graded membership rather than only binary true-or-false categories.
A fuzzy membership function is:
μA(x) ∈ [0, 1]
Where μA(x) represents the degree to which x belongs to fuzzy set A.
A neuro-fuzzy system may learn parameters associated with fuzzy rules or membership functions.
Probabilistic hybrid systems represent uncertainty using probability distributions while also incorporating logical or structured relationships.
For example:
P(H | E) = P(E | H)P(H) / P(E)
Where H represents a hypothesis and E represents evidence.
These approaches can be valuable when tasks involve uncertainty, structured knowledge, and adaptive learning.
However, integrating several paradigms does not inherently guarantee improved performance.
Additional layers may increase complexity, introduce incompatible assumptions, and complicate validation.
2.16. Brain–Computer Interfaces as Biological–Computational Integration
A brain–computer interface (BCI) creates a communication pathway between measured neural activity and an external computational system.
A basic BCI architecture includes:
- Neural signal acquisition.
- Signal preprocessing.
- Feature extraction.
- Intent or state decoding.
- Command generation.
- External device control.
- User feedback and adaptation.
Let z represent measured neural features.
A decoder may estimate an intended action:
â = fθ(z)
The decoded command may then be transmitted to an authorized external device.
Brain–computer interfaces demonstrate a form of biological–computational integration, but they should not automatically be equated with the broader research definition of hybrid intelligence.
Some systems are primarily communication or motor-control interfaces rather than joint cognitive decision-making systems.
Wolpaw and colleagues surveyed foundational BCI approaches for communication and control [17].
Hochberg and colleagues demonstrated neurally controlled robotic-arm movement in research participants with tetraplegia [18].
These studies provide empirical evidence that neural signals can be integrated with computational decoding and external action.
They do not establish unrestricted mind-reading, general cognitive enhancement, or the merger of human and machine consciousness.
2.17. Multimodal Hybrid Intelligence
Hybrid systems may integrate text, images, audio, sensor measurements, structured records, and human observations.
Let:
X = {X_text, X_image, X_audio, X_sensor}
A multimodal model may combine these inputs through a fusion function:
z = F(f₁(X_text), f₂(X_image), f₃(X_audio), f₄(X_sensor))
Where each fᵢ extracts a representation from a particular input modality.
The resulting representation z may support prediction, reasoning, or decision-making.
However, multimodal AI is not necessarily human–AI hybrid intelligence.
Combining several machine-readable input types is a form of multimodal integration, while hybrid intelligence in the human–AI sense requires meaningful integration of human and artificial capabilities.
A system can be both multimodal and hybrid, but the terms describe different architectural properties.
2.18. Verification and Runtime Authority
An intelligent model may recommend an action that falls outside its legitimate authority.
Therefore, decision generation and action authorization should be separated.
Let A(x) represent technically possible actions in context x.
Let A_authorized(x) represent the subset permitted by the applicable rules.
Then:
A_authorized(x) ⊆ A(x)
A valid execution mechanism should permit only authorized operations.
For a proposed action a:
Execute(a) only if Authorized(a, x) = true
This rule represents an architectural requirement rather than a claim that authorization checks eliminate all operational risk.
An authorization service may itself contain implementation errors.
It must therefore be independently tested and protected against bypass.
High-impact operations may also require human approval, audit logging, transaction safeguards, and recovery mechanisms.
These controls are especially important in systems involving financial transactions, sensitive data, healthcare, cybersecurity, or physical machinery.
2.19. Evaluating Hybrid Performance
Evaluation should compare the combined system against appropriate alternatives.
Relevant conditions include:
Human-alone performance: The human completes the task without AI assistance.
Machine-alone performance: The AI system completes the task without human intervention, where such testing is appropriate.
Human–AI performance: The combined system completes the same task under defined interaction procedures.
Alternative collaboration designs: Different delegation, explanation, or approval strategies are tested.
Important measures include:
Augmentation Gain = J(HM) − J(H)
Synergy Gain = J(HM) − max[J(H), J(M)]
Intervention Rate = Intervened Tasks / Evaluated Tasks
Deferral Rate = Deferred Decisions / Eligible Decisions
Error Rate = Incorrect Outcomes / Evaluated Outcomes
Constraint Violation Rate = Executions With Violations / Evaluated Executions
These metrics should be supplemented by resource costs, decision time, robustness, fairness, uncertainty calibration, and user workload.
Where larger metric values represent worse outcomes, such as error or cost, improvement calculations must use the appropriate direction.
A lower error rate does not automatically justify unrestricted machine autonomy.
2.20. Empirical Evidence on Human–AI Performance
The effectiveness of hybrid intelligence is an empirical question.
Vaccaro, Almaatouq, and Malone conducted a systematic review and meta-analysis of 106 experimental studies reporting 370 effect sizes [16].
Their findings indicated that, on average across the included experiments, human–AI combinations performed worse than the better of the human-alone and AI-alone baselines.
The overall reported effect on synergy was negative.
The analysis also identified differences between task categories, including less favourable average results for decision-making tasks than for content-creation tasks.
These findings do not establish that hybrid intelligence is inherently ineffective.
They show that successful collaboration depends on task characteristics, component capabilities, interaction design, and the ability to integrate complementary information.
The scientific objective should be to identify when collaboration creates measurable value, when it introduces avoidable costs, and how the system can be improved.
3. ARCHITECTURAL TEXT MAP
3.1. General Machine Intelligence Architecture
LAYER 1 — TASK AND OPERATING ENVIRONMENT
User objectives, input conditions, external constraints, and available resources.
↓
LAYER 2 — INPUT ACQUISITION
Text, images, audio, sensor information, documents, structured data, or other observations.
↓
LAYER 3 — REPRESENTATION AND PREPROCESSING
Data validation, feature extraction, normalization, tokenization, or state construction.
↓
LAYER 4 — COMPUTATIONAL INTELLIGENCE
Machine-learning models, symbolic reasoning, probabilistic inference, search, planning, or learned decision policies.
↓
LAYER 5 — OUTPUT GENERATION
Predictions, classifications, recommendations, plans, generated content, or proposed actions.
↓
LAYER 6 — VALIDATION AND AUTHORIZATION
Quality checks, uncertainty assessment, rule enforcement, safety controls, and applicable execution permissions.
↓
LAYER 7 — EXECUTION OR DELIVERY
Authorized actions, software outputs, robotic control signals, or user-facing information.
↓
LAYER 8 — FEEDBACK AND EVALUATION
Observed outcomes, error detection, performance assessment, and revision requirements.
↓
LAYER 9 — ADAPTATION OR TERMINATION
Continue, revise, learn through an authorized update mechanism, request assistance, or stop.
3.2. Human–AI Hybrid Intelligence Architecture
LAYER 1 — HUMAN OR ORGANIZATIONAL PURPOSE
Legitimate task request, intended outcome, accountability, and operating constraints.
↓
LAYER 2 — HYBRID TASK ANALYSIS
Task decomposition, uncertainty assessment, and identification of required expertise.
↓
LAYER 3 — CAPABILITY AND AUTHORITY ALLOCATION
Assignment of responsibilities between human participants and artificial systems.
↓
LAYER 4 — PARALLEL OR SEQUENTIAL PROCESSING
Human contextual reasoning, professional judgment, and domain knowledge.
Machine prediction, retrieval, computation, and structured analysis.
↓
LAYER 5 — INFORMATION EXCHANGE
Recommendations, evidence, explanations, uncertainty estimates, and human feedback.
↓
LAYER 6 — JOINT EVALUATION
Comparison of alternatives, contradiction detection, complementary reasoning, and assessment of task requirements.
↓
LAYER 7 — DECISION AUTHORIZATION
Human approval where required, machine permission validation, and safety controls.
↓
LAYER 8 — COORDINATED EXECUTION
Approved actions, generated outputs, external operations, or formal decisions.
↓
LAYER 9 — OUTCOME VERIFICATION
Task success, error analysis, constraint compliance, and operational consequences.
↓
LAYER 10 — MUTUAL ADAPTATION
Human learning, workflow improvement, model evaluation, and explicitly authorized system updates.
3.3. Computational Hybrid AI Architecture
INPUT DATA
Raw observations, structured information, documents, or sensor signals.
↓
NEURAL REPRESENTATION LAYER
Feature extraction, embeddings, perceptual classification, or learned prediction.
↓
SYMBOLIC OR PROBABILISTIC LAYER
Knowledge graphs, rules, constraints, logical inference, or probabilistic reasoning.
↓
INTEGRATION LAYER
Conversion between learned representations and structured reasoning components.
↓
CONSISTENCY AND CONSTRAINT CHECKING
Validation of relationships, applicable rules, and formal requirements.
↓
OUTPUT OR ACTION SELECTION
A verified prediction, structured explanation, decision recommendation, or proposed action.
↓
EXTERNAL EVALUATION
Performance testing, uncertainty analysis, and error monitoring.
3.4. Biological–Computational Interface Architecture
BIOLOGICAL NEURAL ACTIVITY
Task-related signals generated within the nervous system.
↓
SIGNAL ACQUISITION
Suitable sensors or interfaces collect measurable signals.
↓
PREPROCESSING
Noise reduction, filtering, artifact detection, and signal preparation.
↓
FEATURE EXTRACTION
Identification of signal characteristics relevant to the intended interface.
↓
COMPUTATIONAL DECODING
A model estimates the intended command or control state.
↓
AUTHORIZED DEVICE CONTROL
The system executes permitted commands through an external application or device.
↓
SENSORY FEEDBACK
The user observes or otherwise receives information about the result.
↓
ADAPTIVE CONTROL
The interface may be recalibrated through an explicitly defined procedure.
3.5. The Unified Hybrid Intelligence Feedback Cycle
The general interaction cycle can be summarized as:
Human Goal → Machine Assistance → Evidence Exchange → Joint Evaluation → Authorized Decision → Action → Feedback → Adaptation
The fundamental architectural challenge is to make the interaction productive without creating unnecessary cognitive burden, hidden dependencies, or unjustified trust.
A successful hybrid system should preserve the ability to identify which component contributed particular information, what evidence supports the result, and who possesses authority over consequential actions.
4. USE-CASES & LIMITATIONS
4.1. Healthcare and Clinical Decision Support
Machine intelligence can assist with medical image analysis, diagnostic predictions, documentation, and clinical workflow prioritization.
Hybrid intelligence can combine computational findings with professional medical judgment.
The AI system may identify patterns or estimate risks.
A qualified clinician may evaluate the recommendation against medical history, clinical observations, patient circumstances, and established standards of care.
Potential benefits: Faster information processing, improved prioritization in validated settings, and additional decision-support information.
Limitations: Dataset bias, uncertain calibration, incomplete records, automation bias, distribution shift, and clinical errors.
Human involvement alone does not guarantee safety.
The interaction process must be designed to support meaningful review and professional accountability.
4.2. Scientific Research and Knowledge Discovery
Machine intelligence can support literature retrieval, pattern identification, simulation, and data analysis.
Hybrid research systems can combine these capabilities with human expertise in hypothesis formulation, experimental design, scientific interpretation, and evidence evaluation.
For example, a computational agent may identify correlations in a dataset.
A researcher must then investigate whether those relationships reflect causal mechanisms, confounding variables, measurement errors, or statistical artifacts.
Limitations: Fabricated references, incorrect assumptions, reproducibility failures, unsupported causal claims, and excessive confidence in generated explanations.
Scientific reliability requires independent verification rather than agreement between several computational systems.
4.3. Software Engineering and Autonomous Development
AI systems can assist with code generation, debugging, testing, refactoring, documentation, and development planning.
A hybrid engineering workflow may delegate routine implementation to AI while retaining human control over architecture, security, requirements, and production deployment.
Potential benefits: Faster exploration of candidate implementations and reduced effort for well-specified development tasks.
Limitations: Insecure code, unintended behaviour, incorrect dependencies, incomplete testing, and misleading claims of successful execution.
A generated code modification should not be considered correct merely because it is syntactically plausible.
Independent build verification, automated tests, and appropriate review remain essential.
4.4. Education and Adaptive Learning
Machine intelligence can evaluate learner responses, recommend practice activities, and adapt instructional sequences.
Hybrid educational intelligence integrates these functions with human instruction, curriculum design, and learner support.
Teachers may identify motivational, cultural, or personal circumstances that are insufficiently represented in a learner model.
AI systems may provide additional practice opportunities and computationally generated feedback.
Limitations: Incorrect assessments, biased recommendations, excessive personalization, unreliable feedback, and substitution of fluent explanations for genuine understanding.
Educational performance should be measured through meaningful learning outcomes rather than only engagement or completion rates.
4.5. Robotics and Physical Human–Machine Collaboration
Collaborative robots can assist humans with manufacturing, inspection, logistics, and physical manipulation.
Human participants may provide high-level objectives or situational judgments.
Robotic systems may perform repetitive or precision-dependent actions.
Potential benefits: Reduced physical workload, improved consistency, and assistance with appropriately selected operations.
Limitations: Sensor uncertainty, collision risks, inaccurate intent inference, control latency, and unsafe authority allocation.
Physical safety requires independently validated control systems and operational constraints.
Natural-language instructions alone are not sufficient safeguards for high-risk robotic execution.
4.6. Brain–Computer Interfaces and Assistive Technology
Brain–computer interfaces can provide communication or device-control pathways for individuals with certain neurological or motor impairments.
They may translate measured neural activity into control signals for communication software or assistive devices.
Experimental research has demonstrated neurally controlled robotic movement [18].
Limitations: Signal instability, calibration requirements, training burden, interface reliability, physical risks in invasive systems, privacy concerns, and limited generalization across users.
These systems should not be presented as unrestricted cognitive enhancement technologies.
Their capabilities depend on the sensing method, decoding model, user training, and validated operational conditions.
4.7. Cybersecurity and Threat Analysis
Machine intelligence can detect anomalous network activity, classify suspicious events, and correlate security information.
Hybrid intelligence can integrate computational monitoring with human threat analysis and incident-response decisions.
Potential benefits: Faster identification of relevant events and improved prioritization of large information streams.
Limitations: Adversarial manipulation, false alerts, missed attacks, unauthorized containment, and disruption of legitimate services.
Security agents should have clearly bounded execution permissions.
High-impact operations may require independent validation or human authorization.
4.8. Business Strategy and Organizational Decision-Making
Organizations may combine predictive analytics, simulation, and AI-assisted information synthesis with human strategic judgment.
A machine may identify historical patterns in sales or operations.
Human decision-makers may interpret regulatory, cultural, competitive, and organizational factors absent from the model’s input data.
Limitations: Historical bias, changing market conditions, incomplete objectives, overreliance on forecasts, and failure to communicate uncertainty.
A predictive model can estimate outcomes under its assumptions without determining which organizational values or strategic priorities should prevail.
4.9. Accessibility and Communication
Machine intelligence can support speech recognition, transcription, captioning, translation, document accessibility, and assistive interfaces.
Hybrid systems may combine automated processing with human review or correction.
Potential benefits: Reduced communication barriers and greater access to information.
Limitations: Recognition errors, language and dialect disparities, loss of context, privacy concerns, and inaccurate interpretation of user intentions.
Accessibility evaluation should include the needs and feedback of affected users.
4.10. Explainability Is Not the Same as Reliability
An AI system can produce an explanation that sounds coherent without that explanation faithfully representing the mechanism responsible for the output.
An explanation may also increase user confidence despite an incorrect recommendation.
Research on complementary human–AI performance has demonstrated that explanations do not automatically produce better joint decisions [15].
Therefore, explanation quality should be evaluated separately from task accuracy, calibration, robustness, and practical usefulness.
A hybrid system should help users recognize when a recommendation may be wrong, not merely make the recommendation easier to accept.
4.11. Automation Bias and Human Overreliance
Automation bias occurs when users give excessive weight to machine-generated recommendations.
A professional may overlook contradictory evidence because the AI system expresses high confidence.
In other situations, users may reject reliable assistance because they distrust the technology.
Both patterns can reduce joint performance.
Mitigation measures may include calibrated uncertainty, independent initial judgments, meaningful evidence presentation, and interfaces that support correction.
However, requiring constant manual confirmation can create excessive workload and encourage superficial approval.
Oversight mechanisms must therefore be designed according to risk, task characteristics, and human capacity.
4.12. Human Skill Degradation and Deskilling
Repeated delegation of cognitive tasks may reduce opportunities for humans to practice important skills.
Potential concerns include reduced independent problem-solving, weaker error detection, and loss of familiarity with operational procedures.
However, the extent of skill degradation depends on the task, training arrangements, frequency of independent practice, and system design.
It should not be assumed that every form of AI assistance causes deskilling.
Appropriate designs may preserve meaningful human engagement through training, periodic independent assessments, and deliberate allocation of responsibilities.
4.13. Correlated Errors and False Complementarity
Human and machine errors may not be independent.
A human may rely on the same incomplete data used to train the AI system.
Several AI agents may share similar architectures, sources, or assumptions.
Consequently, agreement between components may provide less evidence than it appears to provide.
A system can also produce false complementarity, in which separate components seem to contribute different perspectives but ultimately reproduce the same underlying error.
Evaluation should examine error correlations, source independence, and actual performance gains.
The mere presence of multiple participants does not establish meaningful hybrid intelligence.
4.14. Distribution Shift and Unknown Environments
A machine-learning model may perform well on familiar data but deteriorate when the deployment environment changes.
A medical model may encounter patients whose characteristics differ from those represented in its development data.
A financial model may experience unfamiliar economic conditions.
A language-model agent may receive misleading or unusually formatted external information.
Human participation may help identify some unfamiliar conditions, but it does not guarantee successful correction.
A robust hybrid system should monitor input validity, model uncertainty, environmental changes, and the need for escalation.
4.15. Bias, Fairness, and Unequal Performance
Machine-learning systems can reproduce or amplify patterns present in their training data.
Hybrid systems may inherit those patterns while also introducing human judgment biases.
Possible harms include unequal error rates, inappropriate recommendations, exclusion of relevant groups, and inconsistent treatment across populations.
Fairness cannot be established merely by introducing human oversight.
Human reviewers may possess similar biases or lack the information necessary to correct model errors.
Appropriate evaluation should examine performance across relevant populations and deployment conditions.
The applicable fairness criteria must be selected according to the task and its social, legal, and operational context.
4.16. Computational and Coordination Costs
Hybrid intelligence can introduce additional costs.
These may include:
- Human review time.
- Model inference costs.
- Communication latency.
- Data preparation.
- Interface development.
- Monitoring and auditing.
- Training and retraining.
- Organizational coordination.
- Integration maintenance.
A combined system may achieve marginally higher accuracy while requiring substantially more resources.
A meaningful evaluation should therefore consider resource-adjusted performance.
A conceptual net-benefit measure is:
Net Benefit = Value of Improved Outcomes − Integration Costs − Operational Costs − Expected Failure Costs
Each term must be defined for the actual application.
This framework does not imply that every relevant value can be expressed adequately in monetary terms.
Safety, rights, and legitimate authority may require independent constraints.
4.17. The Risks of Excessive Integration Complexity
Combining multiple systems can create failure modes absent from the individual components.
A neural model may generate incorrect structured data.
A symbolic component may reason correctly from those incorrect premises.
A human reviewer may misunderstand an explanation.
An execution layer may perform an action that satisfies its technical rules but violates the user’s broader intentions.
These failures demonstrate why system-level verification is necessary.
A hybrid architecture should not be considered reliable simply because each component performs well in isolation.
Interactions, interfaces, data transformations, and authority boundaries must also be tested.
4.18. Privacy and Cognitive Data Protection
Human–AI systems may process sensitive personal, professional, or organizational information.
Brain–computer interfaces may additionally process neural signals associated with an individual’s physiological activity or intended control actions.
These data can raise important questions involving consent, confidentiality, access, retention, and appropriate use.
Privacy risks depend on the information collected, the processing method, the context, and the available technical safeguards.
A responsible architecture should follow data minimization, purpose limitation, access control, and appropriate security practices.
For systems involving neural data, the limitations of inference must also be communicated accurately.
A decoder’s ability to classify a limited control signal does not imply unrestricted access to private thoughts.
4.19. Accountability and Decision Authority
Hybrid intelligence distributes computational work across participants, but this does not remove the need for accountable decision-making.
A machine may recommend an action.
A human may approve it.
Another software component may execute it.
The system should preserve sufficient records to determine:
- What information was available.
- Which component generated the recommendation.
- What uncertainty was communicated.
- Whether applicable constraints were checked.
- Who or what authorized the action.
- What operation was executed.
- What result was observed.
The NIST AI Risk Management Framework emphasizes governance, risk identification, measurement, and management across the AI lifecycle [19].
Accountability should be designed into the complete system rather than assigned retrospectively after a failure.
4.20. Why Hybrid Intelligence Does Not Guarantee Artificial General Intelligence
Combining several narrow AI systems does not automatically create a generally intelligent system.
Likewise, integrating human and machine intelligence does not establish that the machine independently possesses human-level general cognitive abilities.
A hybrid system may accomplish a broad set of tasks because humans provide interpretation, supervision, and expertise unavailable to the artificial components.
Its overall performance must therefore be distinguished from the independent capabilities of the machine.
Similarly, an integrated neural-symbolic architecture may perform better on a specific problem while remaining limited outside its intended domain.
General intelligence concerns broader capabilities and transfer across tasks, not simply the number of components incorporated into a system.
4.21. Empirical Limits of Human–AI Synergy
The results of Vaccaro and colleagues provide an important caution for the field [16].
Human–AI combinations do not consistently outperform their strongest individual components.
Possible causes include poor delegation, inaccurate trust calibration, communication problems, inadequate feedback, and failure to recognize which participant is more reliable.
Research should therefore compare:
Human-alone baseline: What humans achieve independently.
Machine-alone baseline: What artificial systems achieve independently.
Hybrid outcome: What the combined system achieves.
Interaction cost: Resources required to coordinate the components.
Failure distribution: Conditions under which the combined system becomes less reliable.
Generalization: Whether benefits persist across different tasks, environments, and participant populations.
The objective is not to demonstrate hybrid intelligence by definition.
It is to identify the conditions under which hybrid integration produces demonstrably better outcomes.
4.22. Principles for Responsible Machine and Hybrid Intelligence
A rigorous engineering framework should satisfy several requirements.
Define machine capabilities operationally. Evaluate systems according to observable performance rather than anthropomorphic assumptions.
Identify complementary strengths. Integrate participants because their capabilities address different needs, not merely because combining them is technologically possible.
Preserve appropriate human authority. Delegation should follow the task’s legitimate governance and safety requirements.
Measure genuine synergy. Compare combined performance against both human-alone and machine-alone baselines.
Calibrate reliance. Human users should understand when the machine is reliable and when independent judgment is necessary.
Maintain evidence provenance. Important outputs should remain traceable to their supporting information.
Verify integration boundaries. Errors can arise when information moves between models, tools, rules, and human participants.
Monitor distribution shift. Deployment conditions may differ from development environments.
Protect sensitive information. Data collection and sharing should remain proportionate to the task.
Support interruption and correction. High-impact systems need effective escalation, rollback, and termination mechanisms.
Evaluate costs and consequences. Improved task performance must be weighed against coordination demands, operational risks, and applicable constraints.
Avoid unsupported claims about consciousness. Computational competence and biological–machine connectivity do not establish subjective experience.
The central scientific conclusion is that machine intelligence concerns the capabilities produced by artificial computational systems, while hybrid intelligence concerns the effective integration of complementary capabilities across humans, machines, or computational paradigms.
Their relationship is neither one of simple replacement nor automatic superiority.
Machine intelligence can automate specific cognitive functions, extend information-processing capacity, and support increasingly sophisticated forms of inference and action.
Hybrid intelligence can create opportunities to combine these capabilities with human expertise, structured reasoning, specialized algorithms, or biological control pathways.
However, successful integration depends on the quality of the components, the validity of their information, the design of their interactions, the clarity of their authority, and rigorous evidence that the combined system performs better for its intended purpose.
The long-term importance of machine and hybrid intelligence therefore lies not merely in constructing more powerful algorithms, but in developing intelligent systems that are capable, complementary, verifiable, adaptable, and appropriately governed.
5. RELATED ENCYCLOPEDIA TOPICS
Collective and Social Intelligence
Embodied and Situated Intelligence
6. REFERENCES
- McCulloch, W. S., & Pitts, W. (1943). A Logical Calculus of the Ideas Immanent in Nervous Activity. The Bulletin of Mathematical Biophysics, 5, 115–133. DOI: 10.1007/BF02478259.
- Turing, A. M. (1950). Computing Machinery and Intelligence. Mind, 59(236), 433–460. DOI: 10.1093/mind/LIX.236.433.
- Licklider, J. C. R. (1960). Man-Computer Symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1(1), 4–11. DOI: 10.1109/THFE2.1960.4503259.
- Engelbart, D. C. (1962). Augmenting Human Intellect: A Conceptual Framework. Stanford Research Institute, Summary Report AFOSR-3223.
- Russell, S. J., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Fourth Edition. Pearson.
- Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning Representations by Back-Propagating Errors. Nature, 323, 533–536. DOI: 10.1038/323533a0.
- LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep Learning. Nature, 521, 436–444. DOI: 10.1038/nature14539.
- Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems, 30.
- Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction. Second Edition. MIT Press.
- Dellermann, D., Ebel, P., Söllner, M., & Leimeister, J. M. (2019). Hybrid Intelligence. Business & Information Systems Engineering, 61(5), 637–643. DOI: 10.1007/s12599-019-00595-2.
- Marra, G., Dumančić, S., Manhaeve, R., & De Raedt, L. (2024). From Statistical Relational to Neurosymbolic Artificial Intelligence: A Survey. Artificial Intelligence, 328, Article 104062. DOI: 10.1016/j.artint.2023.104062.
- Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A Model for Types and Levels of Human Interaction with Automation. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30(3), 286–297. DOI: 10.1109/3468.844354.
- Amershi, S., et al. (2019). Guidelines for Human-AI Interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, Article 3, 1–13. DOI: 10.1145/3290605.3300233.
- Mozannar, H., & Sontag, D. (2020). Consistent Estimators for Learning to Defer to an Expert. Proceedings of the 37th International Conference on Machine Learning, Proceedings of Machine Learning Research, 119, 7076–7087.
- Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. S. (2021). Does the Whole Exceed Its Parts? The Effect of AI Explanations on Complementary Team Performance. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems.
- Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-Analysis. Nature Human Behaviour, 8, 2293–2303. DOI: 10.1038/s41562-024-02024-1.
- Wolpaw, J. R., Birbaumer, N., McFarland, D. J., Pfurtscheller, G., & Vaughan, T. M. (2002). Brain–Computer Interfaces for Communication and Control. Clinical Neurophysiology, 113(6), 767–791. DOI: 10.1016/S1388-2457(02)00057-3.
- Hochberg, L. R., et al. (2012). Reach and Grasp by People with Tetraplegia Using a Neurally Controlled Robotic Arm. Nature, 485, 372–375. DOI: 10.1038/nature11076.
- Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, NIST AI 100-1. DOI: 10.6028/NIST.AI.100-1.
- Organisation for Economic Co-operation and Development. (2024). Explanatory Memorandum on the Updated OECD Definition of an AI System. OECD Artificial Intelligence Papers, No. 8. OECD Publishing, Paris. DOI: 10.1787/623da898-en.