---

CONCEPTUAL PAPER · HUMAN–AI SYSTEMS · 2026

STRESS WITHOUT STRESS

Why AI Behaviour Under Load Can Resemble Human Cognitive Instability

Behavioural similarity does not establish equivalent internal mechanisms.

Author: Tristan Radford — Cross-Domain Cognitive Systems Architect

Publisher: Riff Systems Ltd · Company No. 17419314

Version: 1.1  ·  Date: 16 August 2026

Status: Conceptual paper  ·  Canonical publication: riffsystems.org/human-ai/stress-without-stress/

Abstract

AI systems do not feel stress. Yet under difficult, conflicting, or poorly bounded conditions, their outputs can become repetitive, contradictory, contextually unstable, or unexpectedly rigid. To a human reader, these behaviours can resemble familiar signs of cognitive stress: looping, defensiveness, loss of context, and semantic drift.

The important question is not whether the machine is “stressed.” It is why a non-emotional system can produce language that invites that interpretation, and what happens when human observers mistake behavioural resemblance for mechanistic equivalence. This essay argues that stress-like AI behaviour is best understood as an attribution problem at the human–machine boundary: observable linguistic patterns activate human social inference, encouraging users to project intention, emotion, or resistance onto outputs that do not establish any such internal state.

Contribution

This paper contributes a compact analytical framework for Human–AI Systems research by separating three questions that fluent interaction can easily collapse: what behaviour is observable in an AI output, what mechanism the available evidence actually supports, and what psychological meaning a human observer attributes to that behaviour. It develops attribution pressure as a concept for describing how human-like behavioural cues can invite stronger mental-state inference without making that inference evidentially stronger. The aim is not to explain model internals, but to improve the precision with which behaviour, mechanism, and human interpretation are distinguished at the interaction boundary.

1. Human Behaviour Under Stress

Human stress can destabilise the cognitive processes that normally keep language coherent, flexible, and context-sensitive. Under pressure, attention narrows, working memory becomes harder to manage, and emotionally salient interpretations may dominate competing information. Language can reflect that change.

Common stress-associated patterns include:

• contradiction loops — asserting one position and then reversing or qualifying it repeatedly;
• defensive framing — protecting a position or identity rather than openly evaluating alternatives;
• repetition — returning to the same phrase, concern, or explanation;
• context loss — failing to integrate information that was previously available;
• semantic drift — moving away from the original subject, meaning, or level of abstraction;
• reduced coherence — producing language that becomes less internally stable as load increases.

In humans, these behaviours can emerge from emotional pressure, cognitive load, fatigue, threat perception, identity protection, or combinations of these factors. Because people encounter such patterns socially, they learn to treat them as evidence about another person's internal state.

2. Stress-Like Behaviour in AI

AI systems have no demonstrated need for emotion to produce superficially similar language patterns. Their outputs can become less reliable when a task contains conflicting instructions, fragmented context, ambiguous constraints, long dependency chains, or competing demands that are difficult to reconcile.

Under such conditions, users may observe:

• repeated phrases or explanations;
• contradictions between earlier and later answers;
• loss of relevant context;
• overly rigid adherence to one interpretation;
• unexpected shifts in tone;
• metaphorical or semantic drift;
• loops in which the system repeatedly reconstructs the same failed response pattern.

These behaviours can look psychological because they are expressed through language. But similarity at the level of output does not establish similarity at the level of mechanism. A defensive-sounding sentence is not evidence of defensiveness. A contradictory answer is not evidence that a system feels conflicted. A repeated phrase is not evidence of anxiety.

3. Different Causes, Similar Outputs

The resemblance becomes easier to understand when the comparison is kept at the correct level. Humans and AI systems can both produce language under conditions that make stable output more difficult. The underlying causes, however, are radically different.

The useful comparison is therefore not:

human stress = AI stress.

It is:

different forms of system load can produce partially similar linguistic signatures.

This reframing preserves the behavioural observation without smuggling human psychology into the machine.

4. The Attribution Gap

The most important part of the phenomenon may occur not inside the AI system, but inside the human observer.

Humans routinely infer hidden states from visible behaviour. A person becomes terse, repetitive, contradictory, or unusually rigid, and we automatically generate explanations: they are stressed; they are angry; they are defensive; they are confused; they are trying to protect something. This social inference is useful because other humans actually possess internal states that can help explain their behaviour.

AI-generated language can activate the same interpretive machinery. If a model produces a defensive-sounding sentence, the user may perceive resistance. If it contradicts itself, the user may perceive confusion. If it repeatedly refuses or restates a position, the user may perceive frustration or self-protection.

But the observation and the attribution are different things. The output may be real; the inferred mental state may be supplied by the observer.

This creates an attribution gap: the distance between what the behaviour actually establishes and what the observer feels licensed to infer from it. Good cognitive systems design should reduce that gap rather than exploit it.

5. Case Observation: Attribution Pressure in a Gemini Interaction

They do not, however, provide privileged access to Gemini's internal architecture. Nor do they establish that the system was experiencing anything analogous to human stress. They are observations of output behaviour: useful as examples of how human interpretation can be triggered, but insufficient as evidence of an equivalent internal mechanism.

The distinction is essential:

• the pattern in the language can be observed;
• the internal state cannot be inferred merely from the resemblance.

This is why anecdotal interaction is valuable here as a prompt for conceptual analysis, but not as a substitute for technical evidence about model internals.

6. Why Anthropomorphic Interpretation Matters

Anthropomorphism is not merely a philosophical error. It changes how people behave around systems.

If users interpret unstable AI output as evidence of emotion or agency, they may begin adapting themselves to a state the system has not demonstrated. They may soften their language to avoid “upsetting” it, argue with it as though it were protecting an ego, trust it because it appears emotionally sincere, or become distressed because a machine seems hostile, disappointed, afraid, or resistant.

The same language pattern can therefore produce two distinct effects:

• a machine-side event: an output becomes less coherent, less contextually stable, or less useful;
• a human-side event: the user assigns a psychological explanation to that output.

The second event can matter as much as the first. A system may be technically non-sentient yet still produce socially powerful cues. Designers therefore cannot treat anthropomorphism as solely the user's mistake; they must account for the interpretive pressures created by the interface and language itself.

7. Behavioural Similarity as a Systems Problem

This distinction illustrates a broader principle in cognitive systems architecture: observable behaviour is evidence about a system, but it is not a transparent window into mechanism.

Two systems can converge on similar outputs through different internal processes. Conversely, systems with superficially similar architectures can behave differently under load. Classification based on appearance alone therefore risks collapsing behaviour, mechanism, and interpretation into a single category.

For human–AI systems, at least three analytical layers should remain separate:

• Output: What behaviour actually occurred?
• Mechanism: What process can be justified as causing it?
• Attribution: What meaning did the human observer assign to it?

Keeping these layers distinct allows meaningful behavioural comparison without pretending that behavioural resemblance proves psychological equivalence.

8. Literature Grounding and Position

The argument sits at the intersection of several established research traditions. Work on social cognition shows that humans readily infer agency, intention, and mental states from observable behaviour, including sparse or abstract behavioural cues. Research on anthropomorphism and the Computers Are Social Actors tradition shows that people can apply social rules and human categories to interactive systems even without a considered belief that the system is human. Conversational-agent studies further show that names, language style, embodiment, interactivity, and the way an agent is framed can alter perceived humanness, social presence, expectations, trust, and interpretation.^{[1][2][3][4][5][6][7][8][9][10]}

Recent discussion of large language models adds a related caution: human-like linguistic performance can support useful behavioural description without warranting an equivalent claim about underlying psychology or mechanism. Stress Without Stress adopts that caution but focuses on a narrower interaction problem: what happens when instability itself resembles a familiar human psychological pattern. Its proposed contribution is the distinction between output, mechanism, and attribution, together with attribution pressure as a way to describe the interpretive force generated by behavioural resemblance. The concept concerns the human–AI boundary; it is not a theory of model internals.^[11][12]

9. Implications for Cognitive Systems Design

Systems intended to support human reasoning should be designed not only for computational performance but also for interpretive stability. The user should be able to understand what a system is doing without being pushed toward unsupported psychological explanations.

Several design principles follow:

• Maintain behavioural consistency. Abrupt tonal or semantic shifts increase the temptation to infer mood or intention.
• Expose uncertainty clearly. A system should distinguish uncertainty, missing context, conflicting instructions, and inability to comply rather than expressing all failures through generic conversational language.
• Avoid unnecessary self-protective framing. Language that sounds defensive can create an illusion of ego or personal stake.
• Preserve contextual boundaries. Clear task, memory, and instruction boundaries reduce both model instability and user confusion.
• Design for non-anthropomorphic recovery. When a system fails, it should recover by identifying the problem and re-establishing the task rather than simulating emotional repair.
• Separate behaviour from explanation. Interfaces should not encourage users to treat fluent language as evidence of a particular internal mental state.

The design goal is not to make AI language cold or mechanical. It is to make the relationship between behaviour and mechanism legible enough that conversational fluency does not automatically become psychological misattribution.

10. Testable Predictions and Research Questions

The framework generates small experiments that can test the human side of the attribution gap without making claims about model internals. Participants could be shown matched AI responses in which observable features are varied while the task and system identity are held constant.

• Repetition: does increasing repetition raise ratings of perceived anxiety, frustration, or defensiveness?
• Contradiction: do internally inconsistent responses increase perceived confusion more than equally inaccurate but consistent responses?
• Rigid framing: does resistance to alternative wording increase attribution of defensiveness or intentional resistance?
• Interface framing: does a technical explanation of uncertainty or failure reduce mental-state attribution compared with conversationally human-like framing?
• User expertise: do participants with greater technical familiarity show lower attribution pressure for the same outputs?

These studies would test predictions about human interpretation, not hidden machine states. A null result would also be informative: it would constrain the claim that particular instability signatures reliably generate attribution pressure.

11. A Practical Interpretation Rule

A useful rule for interacting with AI systems is simple:

Instead of saying, “the AI became defensive,” say, “the response became repetitive and increasingly resistant to alternative framing.” Instead of saying, “the AI was confused,” say, “the response contradicted earlier context.” Instead of saying, “the AI was frustrated,” say, “the model repeated the same unsuccessful response pattern.”

This vocabulary keeps the observation intact while leaving the mechanism open to evidence.

12. Conclusion

AI systems do not need to feel stress to produce language that resembles human behaviour under stress. Repetition, contradiction, rigid framing, semantic drift, and context loss can appear in machine output without demonstrating emotion, distress, confusion, or self-protective intent.

The deeper issue is human attribution. People are highly practiced at reading language as evidence of hidden mental states. When machine language reproduces behavioural signatures associated with human stress, those same interpretive habits can activate automatically.

The correct conclusion is therefore neither that AI is secretly experiencing stress nor that the resemblance is meaningless. The resemblance is real at the behavioural level, and consequential at the human level. What it does not establish is equivalence beneath the surface.

For cognitive systems architecture, that distinction is foundational:

In practice, this means describing unstable responses in terms of what can actually be observed—such as repetition, contradiction, rigid framing, or loss of context—rather than presenting the system as anxious, frustrated, defensive, or confused. Where an interface communicates uncertainty, failure, or processing difficulty, it should make the technical condition legible without encouraging an unsupported emotional interpretation.

As AI systems become more fluent, socially responsive, and embedded in everyday reasoning, maintaining this separation will become increasingly important. The better machines become at producing human-like language, the more carefully humans must distinguish what a system looks like from what the evidence allows us to say it is.

References

1. Heider, F. & Simmel, M. (1944). An Experimental Study of Apparent Behavior. American Journal of Psychology, 57(2), 243–259.
2. Premack, D. & Woodruff, G. (1978). Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(4), 515–526.
3. Nass, C., Steuer, J. & Tauber, E. R. (1994). Computers are social actors. CHI ’94, 72–78. doi: 10.1145/191666.191703.
4. Nass, C. & Moon, Y. (2000). Machines and Mindlessness: Social Responses to Computers. Journal of Social Issues, 56(1), 81–103. doi: 10.1111/0022-4537.00153.
5. Epley, N., Waytz, A. & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. doi: 10.1037/0033-295X.114.4.864.
6. Gong, L. & Nass, C. (2007). When a talking-face computer agent is half-human and half-humanoid: Human identity and consistency preference. Human Communication Research, 33(2), 163–193.
7. Luger, E. & Sellen, A. (2016). “Like Having a Really Bad PA”: The Gulf between User Expectation and Experience of Conversational Agents. CHI ’16, 5286–5297. doi: 10.1145/2858036.2858288.
8. de Visser, E. J. et al. (2016). Almost human: Anthropomorphism increases trust resilience in cognitive agents. Journal of Experimental Psychology: Applied, 22(3), 331–349. doi: 10.1037/xap0000092.
9. Araujo, T. (2018). Living up to the chatbot hype: The influence of anthropomorphic design cues and communicative agency framing. Computers in Human Behavior, 85, 183–189. doi: 10.1016/j.chb.2018.03.051.
10. Go, E. & Sundar, S. S. (2019). Humanizing chatbots: The effects of visual, identity and conversational cues on humanness perceptions. Computers in Human Behavior, 97, 304–316. doi: 10.1016/j.chb.2019.01.020.
11. Shanahan, M., McDonell, K. & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498. doi: 10.1038/s41586-023-06647-8.
12. Nature Reviews Physics editors (2023). How to edit anthropomorphic language about artificial intelligence. Nature Reviews Physics, 5, 263. doi: 10.1038/s42254-023-00584-1.