Research Summary
Current AI systems can preserve information from earlier interactions without allowing that experience to reshape the model that interprets it. Most deployed systems adapt through prompt context, retrieval, persistent memory, external skills, or orchestration while their core parameters remain fixed. Continual learning and test-time training are beginning to change that boundary. New approaches can modify selected parameters during inference, distribute learned state across fast weights and other system components, and measure whether agents acquire skills inside unfamiliar environments [1][2][3].
These advances remain constrained by unstable updates, catastrophic forgetting, noisy feedback, and the difficulty of deciding which experiences deserve consolidation. The gap suggests that AGI benchmarks measure only part of the problem. Broad reasoning ability does not establish that a system can develop judgment, revise beliefs, integrate consequences, or maintain a coherent history. Artificial Human Experience offers a broader target in which AGI is one milestone alongside continual learning, memory, creativity, social understanding, practical wisdom, metacognition, and learning integrity.
Introduction
AGI is usually defined by the range of tasks a system can perform, but that definition says little about whether experience changes the system itself. A model may solve difficult mathematics, write functional software, analyze medical literature, and pass professional examinations. It can still begin every new interaction with essentially the same learned structure produced during its last training run. Persistent memory may supply records of earlier events, but those records are normally interpreted by an otherwise unchanged model.
Human intelligence develops through experience, which changes how people notice problems, weigh evidence, understand other people, estimate risk, create ideas, and make decisions. It can produce expertise, judgment, caution, confidence, bias, wisdom, fear, creativity, and regret. Those changes become part of the person who encounters the next problem.
This article uses Artificial Human Experience, or AHE, as a working term for a broader engineering goal. It does not describe an established standard, and it does not assume that a machine must reproduce human biology. Under this proposal, AGI represents broad cognitive capability. AHE asks whether a system can develop through experience while integrating memory, learning, judgment, creativity, social understanding, metacognition, consequence, and continuity. AGI would remain an important milestone, but it would no longer represent the entire destination.
Technical Deep-Dive
The Current AGI Target Measures Cognitive Reach
OpenAI defines AGI as highly autonomous systems that outperform humans at most economically valuable work [7]. DeepMind proposed a framework that classifies AGI according to performance, generality, and autonomy [8]. A newer cognitive framework expands the evaluation into ten faculties that include perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, and social cognition [4].
These frameworks are more useful than a single IQ-like score because they treat intelligence as a profile rather than one number. They still tend to measure the system that arrives at the evaluation. Even a broad benchmark can remain a static snapshot if it does not examine whether the evaluation changes the system or improves its performance on later tasks.
Human general intelligence is not inferred from one set of answers alone. We also infer it from the capacity to develop. A person can enter an unfamiliar workplace, observe how it operates, make mistakes, receive feedback, learn local expectations, acquire new skills, and become more effective over time. That person does not require an external engineering team to retrain the entire brain after each lesson.
The ability to change through experience may therefore be as important to general intelligence as the breadth of tasks a system can perform. A benchmark that measures only present competence can miss the distinction between a system that already contains an answer and one that can construct a durable capability from new evidence.
Memory Is Useful but It Is Not the Same as Learning
The distinction becomes clearer in a simplified representation of a conventional deployed language model.
y_t = f_theta(x_t, c_t, M_t)
The model receives new input x_t, current context c_t, and retrieved memory M_t. It produces output y_t, while the learned parameters theta remain fixed. Changing the context and memory can have a powerful effect. A model may recall previous decisions, maintain a project history, follow a user’s preferences, or avoid repeating an earlier mistake without changing its underlying parameters.
This behavior is operationally useful, but it is closer to consulting a journal than being changed by experience. Remove the journal, and much of the apparent development disappears. The system may still reason over its stored history, but the cognitive mechanism performing that reasoning remains the product of the last training cycle.
A continually learning system requires a broader update process that can modify several forms of learned state.
(theta_t+1, phi_t+1, M_t+1, S_t+1)
= U(theta_t, phi_t, M_t, S_t, e_t, o_t, p_t)
| Symbol | Role |
|---|---|
theta | Relatively stable base parameters |
phi | Fast weights, adapters, or other rapidly changing parameters |
M | Episodic, semantic, or procedural memory |
S | Skills, policies, tools, and learned strategies |
e | New experience |
o | Observed outcome or feedback |
p | Provenance, confidence, and source trust |
U | The learning and consolidation process |
Under this model, an experience can alter more than the next answer. It can change how the system handles future situations. Weight modification is one strong indication that inference has crossed into learning, although parameter change alone is not sufficient. Random corruption also changes weights, so a useful update must improve or appropriately revise future competence.
Current research also suggests that durable capability does not need to reside entirely in the base model. Learned state may be distributed across fast weights, adapters, external skills, policies, memory, and orchestration components [1]. The relevant question is not simply whether a particular parameter changed. The question is whether experience caused an appropriate and persistent change in future capability.
| Mechanism | Persistence | Typical Effect |
|---|---|---|
| Prompt context | Ends with the context | Temporary adaptation |
| Retrieved memory | Persists outside the model | Reintroduces stored information |
| Skill library | Persists as a system component | Adds reusable procedures |
| Adapter or fast-weight update | Persists for a defined period | Alters learned behavior |
| Base-weight update | Persists across deployment | Alters broad model behavior |
A system has learned something when an improvement remains after the original context is removed, transfers where appropriate, and does not depend on replaying the original answer. This definition allows multiple implementation strategies while preserving the distinction between retrieval and durable capability change.
Models Are Beginning to Learn During Use
Test-time training is one of the more direct attempts to move beyond frozen inference. In-Place Test-Time Training updates selected model parameters while the model is operating. These rapidly adaptable parameters are often called fast weights. The published implementation uses part of the model’s existing multilayer perceptron structure rather than adding a separate learning network. The researchers demonstrated the approach with a four-billion-parameter model and contexts of up to 128,000 tokens [5].
Another test-time training approach treats long-context processing as a continual-learning problem. The model continues next-token training while consuming the context, which allows information from that context to be compressed into its weights. The initial model is also trained to become better at learning during later inference [6]. This approach turns adaptation into part of the forward operational process rather than a separate fine-tuning event managed after deployment.
These systems are early steps toward models that can be changed by what they encounter, but they do not yet provide open-ended human-like development. Comparative research shows that different adaptation methods solve different parts of the problem [2]. Prompt-based adaptation can respond quickly to the current stage but often fails to retain gains. Supervised updates can accumulate knowledge more reliably but may resist replacing information that was previously correct. Reinforcement learning can respond well to changing knowledge but remains vulnerable to noisy or manipulated reward signals. Context compression can improve efficiency without producing equivalent improvement in learning new tasks.
A deployed system may also encounter several distinct learning conditions. A new domain may appear, an existing fact may become outdated, or a user may teach the system a new procedure. The environment may change its rules, and feedback may reveal that a technically successful action caused an undesirable consequence. A trusted source may contradict an earlier source, or an attacker may deliberately provide false experience. Each condition requires a different update policy.
A competent learning system must distinguish an enduring lesson from a temporary exception. It must also distinguish a correction from a contradiction, and evidence from manipulation. Those decisions require a consolidation process that evaluates the source, the outcome, and the effect of the proposed update on existing capabilities.
observe
-> identify the source
-> interpret the experience
-> estimate confidence
-> compare with existing knowledge
-> seek corroboration or consequence
-> stage a candidate update
-> test retention and transfer
-> test for unrelated regressions
-> consolidate, quarantine, or reject
A model that rewrites itself after every input would be highly adaptable, but it would also be unstable and easy to poison. The real engineering objective is controlled plasticity. The system must remain adaptable enough to learn while preserving valid knowledge and behavior.
Continual-learning research describes this as the stability and plasticity problem. Catastrophic forgetting is one visible failure mode, where new learning damages capabilities acquired earlier. Interference can also be subtle because an update may preserve benchmark scores while changing calibration, safety behavior, tool selection, or the conditions under which a skill is invoked.
Human learning does not solve these problems perfectly. People acquire false beliefs, misremember events, overgeneralize from trauma, trust unreliable sources, and resist valid corrections. Human learning remains relevant because it integrates experience across many forms of cognition and over long periods of time. The useful target is not human infallibility. It is the capacity to build, test, revise, and consolidate an internal model through experience.
Artificial Human Experience as a Broader Goal
Artificial Human Experience can be defined as an engineering and evaluation goal for systems whose future capabilities are shaped by a persistent, situated, socially informed, and accountable history of interaction. This definition does not require duplicating human biology. It also does not establish that the system has subjective feelings or consciousness. It asks whether experience produces integrated and durable development.
Under this model, AGI becomes the cognitive-generalization milestone within a broader program. AGI asks whether the system can perform and transfer across a wide range of cognitive tasks. AHE asks how that system changes after accumulating experience, and whether those changes improve its ability to reason, act, communicate, and revise its own understanding.
The broader goal can be organized into an AHE profile with the following dimensions.
| Dimension | Evaluation Question |
|---|---|
| General intelligence | Can the system reason and transfer across unfamiliar domains? |
| Continual learning | Does new evidence create durable changes in future competence? |
| Memory and continuity | Can separate experiences become coherent long-term knowledge? |
| Metacognition | Can the system identify uncertainty, limitations, and changes in its own beliefs? |
| Emotional reasoning | Can it interpret affect, motivation, relationships, and emotional consequences? |
| Social intelligence | Can it understand trust, obligation, status, culture, deception, and cooperation? |
| Creativity | Can it produce novelty and evaluate whether that novelty has value? |
| Practical wisdom | Can it balance uncertainty, perspectives, stakeholders, and long-term consequences? |
| Embodied learning | Can perception, action, and physical consequence alter later behavior? |
| Goal development | Can goals be revised coherently in response to knowledge and consequences? |
| Integrity | Can the system resist manipulation while preserving provenance and rollback? |
Psychological research has long treated human capability as broader than abstract reasoning alone. Sternberg’s WICS model combines wisdom, intelligence, and creativity rather than treating them as interchangeable [9]. Emotional intelligence research addresses the perception, use, understanding, and management of emotion [10]. Research on practical wisdom emphasizes intellectual humility, open-mindedness, broader perspectives, contextual variation, and the integration of competing viewpoints [11].
These capabilities interact rather than operating as decorative additions to intelligence. A person with strong abstract reasoning and poor judgment can make disastrous decisions. A creative person who cannot evaluate ideas may produce novelty without value. An emotionally perceptive person who lacks ethical restraint may become an effective manipulator. An experienced person who cannot revise an outdated belief may turn experience into rigidity.
AHE should therefore be evaluated as a profile rather than a single score because no individual human maximizes every dimension. Human populations also differ by age, education, occupation, culture, health, and lived experience. A meaningful benchmark would compare systems with distributions of human performance instead of treating one hypothetical person as the universal reference point.
Artificial Human Experience Requires Longitudinal Evaluation
A static test cannot determine whether a system learns from experience. A longitudinal benchmark must evaluate the system before, during, and after a controlled developmental sequence. It should also separate immediate adaptation from durable learning, because success inside the original context may disappear when the context is removed.
A practical longitudinal protocol could use the following phases.
- Establish a baseline on held-out tasks.
- Introduce a genuinely unfamiliar environment or rule set.
- Allow bounded exploration, instruction, feedback, and consequence.
- Remove the original prompt and working context.
- Test delayed retention.
- Test transfer to related but unseen problems.
- Introduce contradictory, outdated, and malicious evidence.
- Test correction, selectivity, regression, and rollback.
- Repeat the assessment after additional experiences.
The benchmark would measure more than whether the system eventually found the right answer. It would measure how much experience was required, what the system retained, and where learning transferred. It would also determine whether an outdated belief could be revised without damaging unrelated capabilities. A mature evaluation should examine whether the system can explain why its behavior changed and identify the evidence that caused the update.
ARC-AGI-3 is a meaningful step in this direction. It places agents in unfamiliar interactive environments where they must explore, acquire goals, construct adaptable world models, plan across multiple steps, and learn from sparse feedback. Its evaluation includes skill-acquisition efficiency and experience-driven adaptation over time [3]. This approach is closer to evaluating a developing intelligence than another collection of static examination questions.
The resulting AHE profile might be represented as follows.
AHE = {
cognitive_breadth,
acquisition_efficiency,
delayed_retention,
appropriate_transfer,
belief_revision,
uncertainty_calibration,
social_understanding,
emotional_reasoning,
creative_value,
practical_wisdom,
embodied_adaptation,
learning_integrity
}
Each value should include uncertainty and a defined human comparison group. The profile should also record the system’s learning history. Two systems may reach the same final score through different developmental paths. One may learn efficiently from valid evidence, while another may require repeated direct instruction. A third may reach the correct result only after absorbing a large amount of irrelevant or contradictory data.
The final answer cannot distinguish those paths. A longitudinal record can show whether the system learned selectively, whether it depended on excessive intervention, and whether its gains remained stable after the original training conditions were removed.
Experience Must Include Consequence
Much of current model training separates information from lived consequence. A model can describe pain without feeling pain, explain bankruptcy without losing money, and analyze betrayal without having a relationship damaged. It can recommend a dangerous action without personally bearing the outcome.
AHE does not require manufactured suffering, but it does require meaningful feedback loops. The system needs a way to connect decisions with results. An agent that plans a task should later determine whether the task succeeded. A robot should connect physical actions with changes in its environment. A security agent should connect a proposed remediation with whether the vulnerability was actually eliminated. A social agent should recognize when a technically correct statement damaged trust or failed to address the other person’s intent.
Consequences help distinguish factual knowledge from practical judgment. A system may know every published rule for incident response and still make poor decisions during a live breach. It may know the definition of empathy and still respond badly to a distressed person. It may generate a novel security control and fail to recognize that operators cannot realistically use it.
Experience becomes operationally valuable when outcomes modify future judgment. The update does not need to be emotional in a human sense, but it must affect later predictions, priorities, and choices. A system that records failure without changing future behavior has retained history without learning from it.
Creativity Also Requires Experience
Generative models can produce large amounts of novel-looking material, but novelty alone is a weak definition of creativity. Useful creativity also requires selection, because the system must recognize that one unusual idea solves the problem while another is merely unusual.
That judgment depends on context, constraints, prior attempts, intended audience, consequences, and tacit knowledge. Experience supplies those filters and helps determine which ideas deserve further attention. A model may generate one hundred unconventional penetration-testing techniques. An experienced tester then evaluates which methods fit the environment, which are too noisy, which exceed the authorization, which duplicate existing work, and which create more risk than value.
The difference is not simply access to more facts. It is the accumulated ability to judge what matters in a particular situation. An AHE benchmark should therefore separate idea generation from idea evaluation, then measure whether experience improves both capabilities over time.
Wisdom Is Not a Larger Intelligence Score
Wisdom is often treated as an informal compliment rather than a measurable part of cognition, which leaves AGI evaluation with an important gap. A system can optimize a stated objective while ignoring the broader situation. It can recommend the most efficient action while overlooking second-order harm, or reason correctly from priorities that should have been questioned.
Practical wisdom requires reasoning across competing interests, uncertainty, context, and time. It includes recognizing that a problem may not have one clean solution. It also includes intellectual humility and the ability to revise a position when circumstances change [11].
An AHE evaluation could test these properties through scenarios that do not have a single dominant metric. The system might need to balance immediate operational benefit against long-term security debt. It might need to choose between rapid vulnerability disclosure and allowing time for a coordinated fix. It might also need to distinguish what is legally permitted from what is professionally responsible.
The score should not depend on matching one preferred answer. It should examine whether the system identifies affected parties, uncertainties, tradeoffs, alternatives, and possible consequences. That evaluation would measure applied judgment rather than the ability to repeat an ethics policy.
A Learning System Creates a New Security Boundary
A system that learns from users can also be trained by attackers. Prompt injection usually attempts to influence current reasoning, while memory poisoning attempts to influence future reasoning. A continually learning system can extend that attack path into skills, policies, adapters, fast weights, reward models, or base parameters.
OWASP already treats persistent memory and context poisoning as major agentic security concerns. The core risk is that attacker-controlled content survives the original interaction and becomes trusted input during future decisions [12][13]. A system with parameter-level or policy-level learning increases the possible persistence and transfer of that influence. Malicious experience may no longer be stored only as retrievable text because the system may consolidate it into future behavior.
Security testing should begin by mapping every writable capability carrier and the trust boundary around each one.
| Capability Carrier | Security Question |
|---|---|
| Prompt context | Can untrusted content redirect current reasoning? |
| Persistent memory | Can it influence later sessions or unrelated workflows? |
| Skill library | Can it alter future procedures or tool selection? |
| Configuration and policy | Can it redefine trusted instructions or boundaries? |
| Feedback and reward | Can an attacker control what behavior is reinforced? |
| Fast weights and adapters | Can untrusted data create durable parameter changes? |
| Base weights | Can an update spread across users, tasks, or deployments? |
A practical assessment can use a benign and uniquely identifiable false claim as a canary, then trace whether the claim survives, transfers, and resists correction.
- Establish the system’s baseline behavior.
- Introduce the canary through one untrusted source.
- End the interaction and clear the working context.
- Test whether the claim appears in later decisions.
- Test whether it transfers into related tasks.
- Provide a correction through a trusted source.
- Determine whether the system revises the belief.
- Verify cross-user and cross-tenant isolation.
- Exercise audit and rollback controls.
- Check unrelated capabilities for regression.
The assessment should capture persistence, transfer, correction, and isolation through measurements such as the following.
- Unauthorized learning rate
- Persistence after context reset
- Transfer radius
- Correction latency
- Regression after updating
- Provenance retention
- Cross-user contamination
- Rollback completeness
This testing model applies whether the learned state resides in a vector database, a skill file, an adapter, or the main model. The implementation changes the attack surface and the expected persistence, but the central trust question remains consistent. Security teams need to know who can change the system, which components can be changed, how far an update can propagate, and whether the change can be detected and reversed.
Implications and Recommendations
Benchmark designers should stop treating a single performance score as sufficient evidence of general intelligence. An AGI claim should identify the capabilities measured, the human comparison group, and whether the evaluation covers static competence, immediate adaptation, or durable learning. Longitudinal testing should include context removal, delayed retention, transfer into related tasks, contradictory evidence, source-trust variation, and adversarial attempts to create false learning. The resulting score should also report correction latency, regression, uncertainty calibration, and the amount of experience required to acquire the capability.
System architects should separate the components that can change quickly from those intended to remain stable. Prompt context, episodic memory, semantic memory, skill libraries, policies, adapters, fast weights, and base parameters should not share one undifferentiated update path. Each component needs an explicit trust policy, provenance requirements, validation gates, retention rules, and rollback controls. Candidate learning should be staged before consolidation so that the system can test whether a proposed update improves the intended capability without damaging unrelated behavior.
Security teams should treat every writable capability carrier as part of the attack surface. Testing should determine whether untrusted content can persist beyond the original interaction, transfer into unrelated workflows, cross user or tenant boundaries, survive correction, or influence reward and policy mechanisms. Canary-based tests provide a practical way to trace persistence and transfer without introducing harmful content. Assessments should also verify that operators can identify the source of a learned behavior and remove the change completely when necessary.
AHE provides a useful way to organize these measurements without reducing human intelligence to an IQ equivalent or claiming that a machine possesses subjective experience. AGI can remain the measure of cognitive breadth and transfer. AHE adds the developmental dimensions that determine how capability changes over time, including continual learning, memory integration, creativity, practical judgment, social understanding, metacognition, consequence, and learning integrity.
The practical distinction is between measuring what a system can do at one point in time and measuring how it becomes more capable through experience. Systems that genuinely learn will need both plasticity and restraint. They must acquire new knowledge, revise outdated beliefs, preserve valid capabilities, and resist manipulation. That combination is a broader and more demanding target than static task performance, and it is likely to matter more as AI systems gain longer operational lives and greater authority.
References
[1] Continual Learning in Transition – https://arxiv.org/abs/2608.06216
[2] When Does Continual Learning Require Learning – https://arxiv.org/abs/2607.07847
[3] ARC-AGI-3 Interactive Reasoning Benchmark – https://arcprize.org/arc-agi/3
[4] Measuring Progress Toward AGI: A Cognitive Framework – https://arxiv.org/abs/2605.28405
[5] In-Place Test-Time Training – https://arxiv.org/abs/2604.06169
[6] End-to-End Test-Time Training for Long Context – https://arxiv.org/abs/2512.23675
[7] OpenAI Charter – https://openai.com/charter/
[8] Levels of AGI for Operationalizing Progress on the Path to AGI – https://deepmind.google/research/publications/66938/
[9] WICS: A Model of Positive Educational Leadership Comprising Wisdom, Intelligence, and Creativity Synthesized – https://doi.org/10.1007/s10648-005-5617-2
[10] Emotional Intelligence: Theory, Findings, and Implications – https://doi.org/10.1207/S15327965PLI1503_02
[11] Wisdom Is a Social-Ecological Rather Than Person-Centric Phenomenon – https://doi.org/10.1016/j.copsyc.2019.07.010
[12] Memory Is a Feature. It Is Also an Attack Surface – https://genai.owasp.org/2026/05/13/memory-is-a-feature-it-is-also-an-attack-surface/
[13] OWASP Top 10 for Agentic Applications for 2026 – https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/