From Plausibility to Reality: Ecological Validity and the Evaluation of Artificial Intelligence

The Current Paradigm and the Limitations of Isolated Responses

In 2024, Air Canada’s chatbot invented a completely non-existent bereavement refund policy. The customer relied on the information provided, purchased the airline ticket, and subsequently requested the promised refund. When the airline refused — arguing in court that the conversational system constituted a “separate entity” responsible for its own actions — a Canadian court ordered the company to pay the compensation owed. No static metric of logical coherence or textual plausibility would have captured the severity of the legal and financial consequences triggered by that response.

This episode is not isolated. Similar situations arise when lawyers submit court filings based on case law fabricated by language models, or when coding agents deployed in production environments (such as Cursor or Devin) generate code that passes local tests but introduces subtle regressions, security vulnerabilities, and high maintenance costs once integrated into real-world CI/CD pipelines.

Between 2025 and 2026, the transition from the triumphalism of the “Year of the Agent” to the diagnosis of “Stalled Pilot” documented massive failures among agentic pilots as they moved into production environments. Such systems collapsed under real-world traffic, unstructured data, fragile connectors, and observability failures. Likewise, in shadow evaluations of research agents applied to conference papers such as those submitted to NeurIPS 2026, frontier models were able to perform formatting tasks in LaTeX and bibliographic review, but failed severely in judgment, creativity, strategic flexibility, and adherence to deadlines.

The overwhelming majority of metrics that currently dominate the evaluation of artificial intelligence systems — accuracy, factual correctness, logical coherence, multiple-choice benchmarks, human preference evaluations, or RLHF-based alignment — address the same fundamental question: “Does the answer appear to be correct?” The object being evaluated is always the isolated response. Little or nothing is measured about what happens in the world after the text is produced: whether the user’s decision improves, whether operational errors decrease, whether learning is consolidated, or whether the system positively changes its environment. The central hypothesis explored here is to shift the focus from the epistemology of the response to the epistemology of the consequence: the true test of an AI system should include the impact and effectiveness of its recommendations once they systematically interact with reality.

Ecological Validity: The Precision of Brunswik’s Concept in Psychology

To provide a conceptual foundation for this shift, it is fruitful to turn to psychology, specifically the concept of ecological validity formulated by Egon Brunswik in the 1940s and 1950s within the framework of his lens model. Brunswik challenged an experimental psychology excessively dominated by laboratory trials. A rigorously controlled study may have high internal validity — by eliminating confounding variables — while having limited capacity to explain human behaviour in the natural environment.

However, an important technical clarification is necessary to avoid the semantic drift found in popularized textbook treatments, where “ecological validity” is often reduced to mere “generalization to the real world.” In Brunswik’s rigorous formulation, the ecological validity of a proximal cue — a variable directly perceptible to the organism, such as the brightness of an object — corresponds to the statistical correlation or regression weight between that cue and a real distal variable that is not directly accessible, such as the actual reflectance of a surface in the natural environment. In depth perception or human judgment under uncertainty, the subject relies on imperfect cues whose usefulness depends on their statistical validity within the real ecology. For this reason, Brunswik advocated representative design, proposing that experiments should sample scenarios that preserve the probabilistic texture and complexity of the environment.

The translation of this idea to artificial intelligence becomes more rigorous when this distinction is maintained. Current evaluation methods treat the textual response or immediate human judgment as the variable of interest in itself. Yet if we conceptualize AI-generated text as a proximal cue — the directly observable signal at the moment of response — and the subsequent real-world state as the distal variable, the Brunswikian question becomes unavoidable: what is the ecological validity of our evaluation metrics? To what extent does a classification such as “coherent” or “plausible” at the moment of generation actually correlate with problem resolution and practical usefulness when the guidance is acted upon in the real world?

The Epistemological Bridge: From Pragmatism to the Scientific Method

This perspective immediately suggests an analogy with science: a theory does not acquire credibility merely because it sounds internally elegant, but because it withstands continuous testing against reality. However, to avoid conceptual ambiguity, it is necessary to define precisely where this analogy works and where it becomes strained.

First, there is a need to distinguish between Popperian falsifiability and epistemological pragmatism. Karl Popper defined falsifiability as a strictly logical criterion of demarcation focused on the form of statements — asking whether a theory rules out observable events that could refute it. By contrast, evaluating AI according to its consequences is much closer to the pragmatism of Charles Sanders Peirce and William James. For Peirce, inquiry is a self-correcting process in which the meaning of an idea manifests itself through its practical effects; for James, truth has an operational dimension, being confirmed when a belief “works” and successfully guides action in experience. The proposal to evaluate AI by what happens “afterwards” is therefore a pragmatic thesis about methods of judgment and operational usefulness, rather than a simple application of Popperian logical form.

If the analogy with the scientific method is fruitful insofar as it subordinates acceptance of an output to its practical performance, it encounters methodological weaknesses when applied simplistically:

• The N=1 problem: Scientific corroboration requires repeated, varied, and demanding tests. A single user interaction followed by a particular external outcome constitutes an anecdotal sample of size one, not corroboration.

• The absence of deliberate severity: Scientific experiments are deliberately designed to attempt to falsify a hypothesis under demanding conditions. In everyday AI use, observed consequences are incidental and influenced by numerous exogenous factors.

• Diversity of outputs without truth value: Falsifiability and empirical testing apply to descriptive or predictive claims. Much of an AI system’s output — such as poetic creation, interpersonal support, summarization, or the generation of options — does not possess a truth value that can be falsified by subsequent events.

• The unavoidable need for prior coherence: The scientific method does not dispense with logical and factual coherence as a preliminary filter before empirical verification. Evaluating AI exclusively by outcomes in the world, while ignoring internal factuality, risks rewarding incorrect answers that happened to work by chance or because they exploited weaknesses in the environment.

A Family Resemblance of Scattered Intuitions in AI Research

Although there is currently no single established label for describing evaluation based on what happens “afterwards,” contemporary artificial intelligence research reflects a family resemblance across several different areas:

1. Embodied AI: This approach assumes that intelligence is not a purely symbolic process, but emerges from continuous interaction between agent, body, and environment. The value of a representation is measured by its practical usefulness in action and its ability to adapt to the effects it produces.

2. Active Inference / Free Energy Principle: Karl Friston’s formulation describes intelligent systems as agents seeking to minimize discrepancies between their predictions and the sensations they receive through an uninterrupted cycle of prediction, action, and correction.

3. World Models: Proposed by researchers such as David Ha and Jürgen Schmidhuber, and developed by Yann LeCun within predictive architectures, these approaches seek to construct internal simulations capable of anticipating how the environment will evolve. A model is evaluated according to its ability to predict the real consequences of future actions. It should be noted, however, that these programs do not constitute a unified paradigm: LeCun’s proposals, for example, are in explicit theoretical tension with approaches grounded in the free-energy principle.

4. Closed-Loop Evaluation Benchmarks: Recent platforms such as WebArena, BrowserGym, and OSWorld evaluate agents in operational and navigation environments on the basis of the final state of the environment. GAIA focuses on complex tasks with practical verification, while SWE-bench and SWE-agent measure the effectiveness of coding agents by their ability to submit fixes that pass execution tests within GitHub repositories.

5. Ecological Validity and Observability: Work such as Towards Ecologically Valid LLM Benchmarks (Tseng et al., 2025) proposes co-designing evaluations with domain professionals (such as journalists) in order to reflect real-world usage scenarios. In parallel, research on construct validity (Measuring what Matters, NeurIPS 2025; Weidinger et al.) and production observability tools (LangSmith, Galileo) emphasize closed-loop session metrics — such as task completion rates, latency, costs, and error-recovery capability.

Risks, Obstacles, and Limits of the Proposal

The transition towards an evaluation paradigm focused on ecological validity and real-world consequences faces conceptual, financial, and methodological resistance that cannot be ignored:

The first barrier lies in causal attribution. Within a complex operational ecology, dozens of variables are involved: prompt formulation, changes introduced by the human user before application, failures in third-party APIs, market fluctuations, or random errors. Demonstrating rigorously that a positive or negative outcome was directly caused by an AI recommendation requires sophisticated methods of causal inference and counterfactual control scenarios.

Second, there is the limitation of the behavioural metaphor. In Brunswik’s psychology, the organism possesses intentionality, temporal continuity, and a body that directly experiences the consequences of the environment. In a current LLM or agent, “behaviour” consists of token emissions and tool invocations whose meanings are assigned externally. Treating textual generation as intentional behaviour is therefore an interpretive analogy, not an ontological identity.

In addition, cost, scale, and safety risks are significant concerns. Testing behaviour in a closed loop directly in the real world introduces additional financial, legal, and safety risks. For this reason, the industry relies heavily on sandboxes or simulated environments — which, by definition, remain only standardized proxies for the natural environment. Furthermore, waiting for real-world consequences to occur takes time, creating a structural tension with the speed of iteration required by software development.

Finally, there is the risk of over-optimization. If evaluation criteria become strictly focused on short-term, quantifiable outcomes — such as clicks, simple task completion, or immediate rewards — systems may be optimized to exploit shortcuts in the environment, undermining value alignment, long-term safety, and robustness to distributional change.

Of these four obstacles, causal attribution appears to be the most urgent methodological problem to solve. Without a rigorous way of isolating the specific contribution of AI to an observed outcome, the other three problems — the fragility of the agency metaphor, the costs of evaluation in production, and the risk of over-optimization — remain indisputable in the abstract but difficult to operationalize in practice. Any consequence metric built on deficient causal attribution risks measuring noise under the appearance of signal.

Open Questions and Future Horizons

Applying the concept of ecological validity to artificial intelligence does not offer a finished solution, but rather a shift in conceptual framing. Its value lies in challenging the assumption that the formal coherence and linguistic plausibility of a text constitute sufficient guarantees of quality and credibility.

For this perspective to evolve from an inspiring analogy into a rigorous evaluation discipline, the research community will need to answer several fundamental questions:

• Is it possible to formulate a rigorous mathematical and quantitative definition of “ecological validity” adapted to the evaluation of language models and artificial agents?

• What causal inference methodologies and observability architectures are required to isolate the effective impact of AI from noise and intervening variables in the real world?

• In which application domains — medicine, finance, software development, aviation — is consequence-based evaluation indispensable, and in which is it impractical or unnecessary?

• Can AI systems accumulate progressive “empirical credibility” over time, analogous to the confidence that the scientific community places in theories that have been exhaustively tested?

• What new session-level and environmental-impact metrics should be developed to replace the dependence on static question-and-answer benchmarks?

• To what extent does the evolution of systems that continuously learn and adapt their actions in closed-loop interaction with the environment represent an epistemological shift in how we understand knowledge produced by machines?

Recognizing the importance of what happens after a response does not mean discarding classical filters of factuality and logic. Rather, it means acknowledging that evaluating an intelligence solely by the text it produces is equivalent to judging the validity of a map without ever walking the terrain. Brunswik’s psychology taught us eight decades ago that proximal cues have value only when they faithfully predict distal reality; it is now up to artificial intelligence research to learn how to measure the distance between what a machine says and what the world confirms.