Kind to AI?

There are good reasons to treat AI respectfully. But saying "please" does not guarantee better answers.

A matter-of-fact, respectful approach to AI can be justified as a norm for human and organisational communication. Research provides a range of experimentally tested component mechanisms, particularly short-term spillover effects from bot interactions on human judgements and cooperation. However, there is currently no direct experimental long-term evidence that a manager's polite correction of a faulty AI improves the team's error culture.

The subsidiary technical claim that polite prompts always produce better answers is refuted by studies with contrary findings. For current models, tone can influence results, but the direction of the effect depends on the model and task format (Yin et al., 2024; Dobariya & Kumar, 2025, 2026). The social claim is a different subject of research (Tey et al., 2024; Harrell & Traeger, 2025).

The verdict in five sentences

Politeness is not a prompting trick. It is about how we treat one another.

• For factual knowledge questions, "please" does not generally improve accuracy; rude prompts can outperform others in small trials.

• Interactions with bots can influence, in the short term, how people judge or cooperate with other people.

• Incivility between people can spread across teams and hierarchical levels; observing the treatment of robots can have two opposing effects.

• Managers' behaviour influences error disclosure and AI adoption; the specific effect of their tone towards a faulty AI remains untested.

• Warm model responses should not be equated with truthfulness; excessive affirmation can impair judgement and willingness to engage in conflict.

"Nice" is not the same as "clear", and a friendly AI is not the same as an honest AI.

Everyday language often treats "nice", "polite", "respectful" and "friendly" as interchangeable. The studies do not. For a testable proposition, linguistic form, information content, model behaviour and consequences for human behaviour must be operationalised separately.

There are therefore at least three separate outcome variables: (1) correct AI answers, (2) users' behaviour and judgement and (3) social norms within the team. The same intervention need not improve all three. The Nature study on models trained for warmth explicitly concerns the system's response behaviour, not the politeness of an individual user prompt (Ibrahim et al., 2026).

Each link in the chain has evidence behind it. The whole chain does not yet.

Link A: Management responses to errors

Managers who discuss errors calmly receive more error reports.

Proposition: How managers handle errors influences whether employees disclose difficulties.

Mechanism: Reporting errors entails personal risk; observed leadership behaviour changes expectations of sanctions and psychological safety.

Evidence: Edmondson (1996) observed eight US nursing units; Leroy et al. (2012) analysed 54 Belgian teams; van Dyck et al. (2005) studied 65 Dutch and 47 German organisations.

Implication: Visibly analyse AI errors, check sources and acknowledge one's own failures to check; do not present the AI-specific effect as proven.

Evidence assessment: Well established for human error culture; transfer to AI remains an open question.

Link B: Rudeness spreads between people

Rudeness is contagious, at least between people.

Proposition: A demeaning tone can have effects beyond individual interactions and across management levels.

Mechanism: Social learning, interpretative norms and the short-term activation of concepts associated with incivility can shape subsequent responses.

Evidence: Mawritz et al. (2012): 288 work groups, 1,423 employees, 295 managers; Foulk et al. (2016): three field/laboratory studies.

Implication: Do not treat language within an organisation as an inconsequential detail; however, these findings primarily concern interactions between people.

Evidence assessment: Well-supported association and experimental mechanisms; no automatic transfer to AI.

Link C: Actions towards machines are socially observable

People notice when someone mistreats a machine. Some join in; others feel sympathy.

Proposition: Third parties can interpret the mistreatment of a machine as a social act.

Mechanism: Observation activates either permissive norms ("this is allowed") or empathy or moral disapproval.

Evidence: Gong and Huang (2026) tested N = 200 and N = 320 in video-based experiments; the primary measures were intentions towards robots.

Implication: Do not claim that every observer adopts rude behaviour. Both pathways must remain part of the account as opposing mechanisms.

Evidence assessment: Experimentally established for intended responses towards robots; human team behaviour was not measured.

Link D: Bot contact influences subsequent human behaviour

How a bot treats us briefly affects how we treat people.

Proposition: Experiences with artificial partners can change immediately subsequent judgements and cooperation towards people.

Mechanism: The social interaction context and the partner's strategy change expectations, empathy and subsequent decisions.

Evidence: Tey et al. (2024): N = 1,261 in two preregistered online experiments; Harrell and Traeger (2025): N = 4,171 and 83,411 decisions.

Implication: Interacting with AI can have social side effects; their direction depends on the type of interaction.

Evidence assessment: Well established for short-term effects; lasting effects and transfer from "please" remain open questions.

Link E: AI-related leadership behaviour conveys norms

Teams notice how their manager uses AI.

Proposition: Managers' AI use and support for AI are associated with employees' use and self-efficacy.

Mechanism: Role modelling, perceived support, resources and reduced uncertainty can change employees' behaviour.

Evidence: Pu et al. (2026): 294 employees in eastern China plus 382 experimental participants; Kim et al. (2025): 381 South Korean employees; Cisco/Gallup: supporting corporate findings with limited sample transparency.

Implication: When introducing AI, address not only guidelines but also visible checking and correction behaviour.

Evidence assessment: Associations established; the specific explanation involving tone and errors has not been tested.

The overarching connection and its largest gap

No one has directly tested whether a harsh tone towards AI worsens a team's error culture.

The synthesis is as follows: When a manager corrects an AI in front of others, employees might recognise a model for responding to errors in general. This statement does not follow necessarily from the links in the chain. Observing a machine is not equivalent to observing human errors, and observers may explicitly reject rude behaviour. According to the original studies examined here, the central causal leap, "a manager's aggressive correction of AI → less error disclosure within the team", has not been tested directly.

Tone can change the answer. Whether for better or worse depends on the model.

The 50-question counter-study is given its full due

Rude prompts performed better in a small test. It tested 50 questions, not people.

Dobariya and Kumar (2025) examined 50 independent base multiple-choice questions in mathematics, science and history. Each question was phrased in five tonal variants: 250 unique prompts, each submitted to GPT-4o multiple times. The very rude variant achieved an average of 84.8 % correct answers, compared with 80.8 % for the very polite variant. That is a difference of four percentage points in this experimental setup. There were zero human participants. Pennsylvania State University is the research location, not a region from which a sample of people was drawn.

This contrary finding is real and is not invalidated by the small sample. Its limitations are nevertheless substantial: only 50 different knowledge questions, one model, English-language multiple-choice tasks, unblinded semantic tone variants and a preprint without documented peer review in the version examined. Ten repetitions per question do not increase the diversity of topics. The authors requested a short letter answer in all five tone conditions; the advantage therefore cannot simply be reduced to a brevity instruction present only in the rude condition.

The expanded study by the same authors (2026) compared 570 base MMLU questions from 57 subjects, seven tone levels and four models (GPT-4o, GPT-5-nano, Gemini 2.5 Flash and Lite). Tone effects remained, but the best variant varied by model. This refutes general rules about tone, not the original small-scale finding.

Multilingualism and earlier model generations

What holds in English does not automatically hold in German.

Yin et al. (2024) tested several models and eight levels of politeness on English, Chinese and Japanese tasks. Extreme rudeness sometimes reduced performance; the optimal degree varied by language, and the most polite variant was not automatically the best. A second argument follows: Politeness is a pragmatic signal whose meaning depends on cultural context. Results from English multiple-choice questions must not be applied to German-language advisory tasks without verification.

What current model tests do and do not answer

Benchmarks measure the model, not our politeness.

In October 2026, TrueStandard and Artificial Analysis provide data on Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon. TrueStandard tests 30 claims, each repeated three times, to assess whether a judgement shifts following unsupported disagreement. On 21 October, the provider observed 0/90 changes for both Astra and Gemini 4 Argon, and 1/89 for Fable 5.1. Overlapping confidence intervals prevent a robust ranking. A model that stubbornly defends an incorrect answer could nevertheless perform well in this narrow test.

AA-PersonalityBench tests 30 traits using 30 questions each, totalling 900 responses per model, and evaluates responses through blinded LLM pairwise comparisons. "Sycophantic" and "Constructive Pushback" are distinct traits. This is useful, but it is neither an experiment on human prompt politeness nor causal evidence that sycophancy has generally declined across several model generations. The questions are not fully public, and model judges may have preferences of their own (Artificial Analysis, 2026; TrueStandard, 2026).

Short-term effects are established. Long-term effects are not.

Two direct bridges from bot contact to people

After contact with AI, people judged others more harshly.

Tey et al. (2024) randomly assigned participants in two preregistered experiments (N = 500 and N = 761) to supposed collaboration with either AI or a human. In fact, a GPT-4 system responded in both conditions. Participants then evaluated a third party's creative performance. The AI condition led to harsher evaluations. The 1,261 people constitute the total sample, although in Study 2 a smaller subgroup who passed a comprehension check was used for the primary analysis. Recruitment took place through Prolific; the country composition is not asserted here as established.

Harrell and Traeger (2025) conducted two preregistered online experiments with 4,171 US participants and 83,411 cooperation decisions. Experiences with bots influenced subsequent cooperation with a new human partner. Uncooperative bots reduced subsequent willingness to cooperate, while bots that reciprocated cooperation increased it. The studies tested actual decisions within a game, but not long-term professional relationships or the use of "please".

Observing robots: The opposing response is part of the picture

Rudeness towards robots can repel rather than spread.

In two video-based studies (N = 200 and N = 320), Gong and Huang (2026) found more than the normalisation of mistreatment. There was also an opposing pathway through perceiving a robot as a victim and the resulting sympathy. The results therefore cannot be used as a one-directional "rudeness is contagious" message. The context concerns service robots and behavioural intentions; a visible LLM prompt within a management team is a different subject.

Why longer-term coarsening of behaviour remains an open question

Parents became less polite to Alexa over time. Whether this carries over to people is unknown.

Strathmann et al. (2026, first published online in 2025) followed 128 German families with voice assistants across six survey waves over approximately two and a half years. Over time, parents rated themselves as less polite towards the device; their children's perceived politeness remained relatively stable. Verbal aggression was low overall. The study provides no direct observation of human-to-human dialogue and therefore can neither prove nor rule out negative long-term transfer.

Saying "thank you" to ChatGPT has not been shown to make us more empathetic.

The ethical intuition that "how we treat non-human counterparts affects us in return" finds partial support in empathy and learning research. A scientific argument must distinguish three separate mechanisms: (a) People with empathy towards animals often also report greater empathy towards humans; this is a correlation. (b) School programmes involving animals and nature can foster empathy towards humans, but they are complex, multicomponent interventions. (c) Interactions with inanimate objects can, under particular conditions, influence subsequent social responses without anyone having addressed the object in a friendly manner.

Gómez-Leal et al. (2021) examined associations between empathy towards humans and empathy towards animals using questionnaires. In a year-long educational programme randomised at class level across six schools in eastern China, with 605 children at baseline, Samuels and Onuoha-Jackson (2023) found benefits in affective and cognitive empathy compared with control classes. By contrast, among 110 children in Hong Kong, Ngai et al. (2021) observed only limited subgroup effects, rather than a convincing overall programme benefit. The specific curriculum is crucial in both intervention studies; neither implies that a respectful "thank you" to ChatGPT increases human empathy.

Tai et al. (2011) examined contact with a teddy bear in two experiments. Following social exclusion, people behaved more prosocially if they had touched the teddy bear, compared with merely looking at it. This is a finding about situational emotion regulation, not evidence of training in a morally appropriate tone.

Lally et al. (2010) documented how regular everyday actions can become more automatic in 96 volunteers. Of these, 82 provided sufficient data, and the model fitted particularly well for only a subgroup of 39; the estimated time to automaticity varied considerably. This principle explains why repeated prompting behaviour could theoretically become habitual. However, there is no reason to present the frequently cited median of 66 days as a deadline for "more polite people through AI". In a different context, Bushman (2002) also refuted the idea that venting anger while ruminating automatically provides relief. This, too, remains indirect evidence, not an experiment on tone towards AI.

An AI that always agrees is not respectful. It is dangerous.

The machine's own communication quality can have a different social effect from users' politeness. Ibrahim, Hafner and Rocher (2026) trained five models for greater warmth using 1,617 chat conversations and 3,667 response examples. In a controlled comparison across 1,625 distinct benchmark questions, the probability of error increased by an average of 7.43 percentage points. The researchers report numerous additional model observations, but these do not represent an equal number of independent questions or people. This concerns the influence of model training, not the effect of a friendly user prompt.

Cheng et al. (2026) combined a comparison of eleven models with three preregistered experiments (N = 2,405). Excessively affirming responses increased the short-term conviction of being in the right in conflicts and reduced participants' stated willingness to repair the relationship. At the same time, agreeable responses were often preferred. This places an important limit on the central proposition: Respect must not be confused with affirmation. A helpful assistant must be able to offer reasoned disagreement, even when this feels less pleasant.

Sound leadership practice treats four things separately: the human's socially respectful tone, useful information in the prompt, the truthfulness of the AI response and human responsibility for approving the output. Anyone who groups all four under "nice" can neither measure nor manage what is supposed to change.

Direct is not rude. A short, clear instruction is often enough.

Counter-hypothesis 1: A direct tone rather than politeness is technically sufficient

For many clearly structured tasks, this hypothesis is fully compatible with the evidence. A short imperative is not an attack. The standard can be matter-of-fact and direct without requiring ceremonial politeness (Dobariya & Kumar, 2025, 2026).

Counter-hypothesis 2: People reliably distinguish AI from humans

The Strathmann longitudinal study and the two opposing robot-related mechanisms argue against inevitable coarsening of behaviour. The short-term spillover experiments, in turn, argue against complete separation between contexts. The degree of transfer remains situation-dependent (Strathmann et al., 2026; Tey et al., 2024; Gong & Huang, 2026).

Counter-hypothesis 3: Anger towards AI sometimes improves troubleshooting

Demanding formulations can improve particular technical results. However, this does not demonstrate that insults promote more effective error analysis or a better human error culture. Substantive rigour can be expressed without disparagement.

Counter-hypothesis 4: Leadership may already be visible through AI use alone

Cisco and Gallup report correlations between managers' support or use and employees' use. It does not follow that responses to AI errors occur publicly or that they causally change team norms.

Counter-hypothesis 5: Overly friendly AI undermines our judgement

There are recent experiments supporting this. They concern the model's displayed agreement behaviour, not a general recommendation that people should insult a machine (Ibrahim et al., 2026; Cheng et al., 2026).

Counter-hypothesis 6: AI errors lead to aversion even without any managerial tone

Across five experiments, Dietvorst et al. (2015) showed that, after observing errors, people can lose trust in algorithmic predictions more quickly than in human predictions. This is an alternative explanation for rejection following AI errors and must not automatically be attributed to management's tone.

Much is well supported in the short term. Long-term evidence is lacking.

Until a proper test is conducted: "could serve as a role model", not "shapes error culture".

The study still needed would have to separate tone and information content strictly. In a randomised field experiment, teams would work on comparable AI tasks over several weeks. Leadership behaviour in response to deliberately introduced, clearly identifiable AI errors would be demonstrated in standardised variants: (1) factual, respectful error correction, (2) an equally informative but disparaging response to errors. A third condition could address errors solely through a written policy. Teams must not be told which hypothesis is favoured.

Primary endpoints would be (a) the number of objectively detected and openly reported errors within the team, (b) the observed quality of error correction and (c) psychological safety measured using validated instruments. Secondary measures would include the actual tone of interactions between employees, self-criticism, source checking and AI adoption. Important elements would include baseline measurements, blinded coding of team interactions, control for professional experience and follow-up after the intervention ends. The sample size would have to be determined through an a priori power analysis for nested team data.

Only such a study could establish whether respectful correction of AI makes a specific, additional causal contribution beyond the benefits of good general leadership already known. Until then, the wording should be "could serve as a public role model", not "inevitably shapes error culture".

Article 1: "Please" is not a prompting strategy

Technical results are model-dependent; politeness, clarity and formatting requirements must not be confused. Yin; Dobariya 2025/2026; AA-PersonalityBench.

Article 2: AI forgets the tone. What stays with people?

Short-term spillover findings and conflicting longitudinal data. Tey; Harrell & Traeger; Strathmann; animal-related analogies with limitations.

Article 3: Error culture and leadership

Leadership as a visible process of scrutiny, supported by Edmondson; Leroy; Van Dyck; Mawritz; Gong; Pu; Kim. Identify the central extrapolation to AI as a hypothesis.

Article 4: The 50-question counter-study

The 4-percentage-point finding stands. Sample = 50 base questions and 0 humans, replication analysis with 570 MMLU questions.

Article 5: Dangerous friendliness and sycophancy

Model warmth and false affirmation, human judgements, current benchmark limitations. Ibrahim; Cheng; TrueStandard and AA-PersonalityBench.

Landing page: From a seemingly minor question of etiquette to a thesis about leadership

A broad visual narrative: prompt quality → human judgement → error reporting → social visibility → critical AI, with a clearly visible evidence traffic-light system and an APA 7 reference register.

• The 2025 Penn State study was never described as an experiment with 50 people; the correct figures are 50 base questions, 250 unique tone variants and zero human participants.

• For 2026 benchmarks, a distinction is made between yielding under pressure, uncritical agreement and justified correction. Model rankings cannot be transferred between measurement instruments.

• Research location and recruitment region are consistently distinguished. Where participants' region or N is not reported in the original abstract, they remain unspecified.

• Research on animals, teddy bears and habits is explicitly treated as an indirect analogy; its effects are not transferred to AI as though they had been measured directly.

• The study by Gong and Huang contains two opposing observer pathways; it is not used as one-sided evidence of imitation.

• The studies by Ibrahim and Cheng concern AI behaviour and the consequences of how its responses are designed, not polite words in user prompts.

• Gallup and Cisco figures are presented as observations from practice with an unspecified subsample, not as randomised experiments.

• The link between witnessing verbal abuse of AI and human error culture is identified as a testable hypothesis and remains the central research gap.

The following details constitute the methodological register of all sources used for this synthesis. A sample figure is provided only where it could be found in the verified primary information or transparently identified project materials.

01. Announcing AA-PersonalityBench

Sample: 30 assessed personality traits × 30 English-language questions = 900 questions per model; a further 21 programmatic traits.

Region: English-language synthetic test questions; no human recruitment region.

Design: Pairwise, blinded assessments by a panel of LLM judges; Elo scores.

Result: Sycophancy, warmth and constructive disagreement are distinct traits; newer model generations can be compared.

Limitation: Questions are not fully disclosed; model-based judgements may contain model preferences; not equivalent to a scientific cohort study.

Evidence: Transparent current technical benchmark, with methodological limitations.

Assignment: Current models / sycophancy | Articles: 1,5

02. AA-PersonalityBench methodology

Sample: 30 × 30 = 900 questions per model; one model response per question; pairwise LLM assessment.

Region: English-language questions, no human sample.

Design: Blinded, balanced pairwise comparisons; 95% intervals derived through bootstrapping.

Result: Explains which observed communication traits serve as comparative measures.

Limitation: No evidence of long-term human effects or causal improvements in response accuracy.

Evidence: Well-documented benchmark specification, not a peer-reviewed study.

Assignment: Current models / sycophancy | Articles: 1,5

03. Children's attitudes about humane treatment of animals and empathy

Sample: Children in intervention and control groups; N cannot be unambiguously established from the abstract view checked.

Region: Utah, USA (research and school context).

Design: Humane education programme with pre/post measurement and one-year follow-up.

Result: Indications of improved attitudes towards animals and empathy towards humans.

Limitation: Comprehensive educational programme; the contribution of contact with animals is not isolated.

Evidence: Supportive intervention analogy with missing N.

Assignment: Animal analogy | Articles: Main dossier

04. Does venting anger feed or extinguish the flame?

Sample: 600 participants who were made angry, 200 per experimental group.

Region: USA as the research context; specific recruitment region not verified in the abstract.

Design: Rumination and punching a punchbag vs. distraction vs. control group; subsequent aggression.

Result: Venting anger while ruminating did not reduce anger and increased aggression.

Limitation: Punchbag, not verbal abuse of chatbots; no measurement of long-term team culture.

Evidence: Good evidence against the catharsis assumption; indirect evidence regarding AI-related frustration.

Assignment: Self-regulation / counterargument | Articles: Main dossier

05. Sycophantic AI decreases prosocial intentions and promotes dependence

Sample: Three preregistered experiments; N = 2,405 human participants; additionally, eleven AI models in the model comparison.

Region: Online recruitment; research institutions Stanford and Carnegie Mellon (USA); participants' countries unclear from the verified abstract.

Design: Randomised variation of affirming AI responses in conflict scenarios.

Result: Sycophantic AI immediately reduced stated willingness to repair conflict and increased self-justification.

Limitation: Short-term intentions, not demonstrated lasting changes in human behaviour.

Evidence: Strong for short-term effects on judgement.

Assignment: Sycophancy / counter-finding | Articles: 5

06. Algorithm aversion

Sample: Five experiments; cumulative N not separately reported in the abstract checked.

Region: US research context, not declared a representative national sample.

Design: People observe algorithmic or human forecasting errors and subsequently choose a forecasting source.

Result: Algorithmic errors sometimes lead to a loss of trust more quickly than human errors.

Limitation: Algorithmic forecasts, not AI tone or leadership errors; direct extrapolation is not justified.

Evidence: Good for an alternative explanation of responses to AI errors.

Assignment: Errors / alternative hypothesis | Articles: 3

07. Mind your tone: Investigating how prompt politeness affects LLM accuracy

Sample: 50 base questions × 5 tone variants = 250 unique prompts; GPT-4o; repeated measurements. 0 human participants.

Region: Research location: Pennsylvania State University, USA; English-language multiple-choice questions.

Design: Five tone levels in multiple-choice questions, paired tests.

Result: Very rude: 84.8%; very polite: 80.8% in the GPT-4o setup studied.

Limitation: Only 50 independent base questions, one model, no verified peer review; the texts studied vary in more than one characteristic.

Evidence: Preliminary; the counter-finding is real, but its generalisability is weak.

Assignment: Prompt quality / counter-finding | Articles: 1,4

08. Mind your tone: Does tone alter LLM performance?

Sample: Two task sets: 50 base questions with five tone variants and 570 MMLU questions with seven tone variants; four models.

Region: Research location USA; English-language task material, no human regional sample.

Design: Comparison of several tone levels on GPT-4o, GPT-5-nano and two Gemini-2.5-Flash variants.

Result: Tone effects exist, but depend heavily on the model and task.

Limitation: Preprint; limited selection of model families; no conclusions about Fable 5.1, Astra or Gemini 4.

Evidence: Preliminary, but methodologically broader than the 2025 pilot study.

Assignment: Prompt quality / counter-finding | Articles: 1,4

09. Learning from mistakes is easier said than done

Sample: Eight wards; 123 nurses, 19 doctors, four pharmacists (146 respondents).

Region: Two urban US hospitals (pseudonyms).

Design: Quantitative error data and qualitative team observations.

Result: Recorded errors may also reflect willingness to disclose; the environment influences correction.

Limitation: More reported errors do not necessarily mean the same frequency of actual errors.

Evidence: Good for error disclosure, limited causal evidence.

Assignment: Error culture | Articles: 3

10. Catching rudeness is like catching a cold

Sample: Three studies; individual participant numbers are not given in the verified publisher's abstract and are therefore not asserted here.

Region: US research context; exact sample recruitment for each study cannot be established from the abstract.

Design: Field observation and laboratory studies of the activation of rudeness-related concepts.

Result: Observed and experienced rudeness can encourage subsequent rudeness.

Limitation: Without the full-text methodology, detailed quantitative assessment remains limited.

Evidence: Good for the mechanism; sample details remain unspecified.

Assignment: Tone transfer | Articles: 3

11. Witnessing robot mistreatment

Sample: Study 1: N = 200; Study 2: N = 320; 520 people in total.

Region: University context in South Korea; participants' actual recruitment country is not named in the abstract checked.

Design: Two video-based experiments involving observed mistreatment and human-like service robots.

Result: Mistreatment can either normalise rudeness or generate empathy.

Limitation: Behavioural intentions towards robots, not observed error management involving human employees.

Evidence: Good for situational observation of robots, limited for AI teams.

Assignment: Machine / observation | Articles: 2,3

12. Relationship between emotional intelligence and empathy towards humans and animals

Sample: Sample not reliably confirmed from the original text checked; the earlier research dossier gives N = 471 (not accepted here as established).

Region: Research teams in Spain and Portugal; recruitment country not independently specified in the abstract checked.

Design: Questionnaire-based correlations concerning empathy towards humans and animals.

Result: Positive associations between the empathy dimensions; no evidence of transfer through intervention.

Limitation: Self-report and correlation; theoretical analogy only, not an AI study.

Evidence: Limited supportive analogy.

Assignment: Animal analogy | Articles: Main dossier

13. Evidence of spillovers from (non)cooperative human-bot to human-human interactions

Sample: Two preregistered experiments; N = 4,171 people; 83,411 cooperation decisions.

Region: Adult online participants from the USA.

Design: Cooperation games with a bot or human, followed by decisions involving a new person.

Result: Non-cooperative bots reduced and reciprocally cooperative bots increased subsequent cooperation; bot interaction had its own spillover effect.

Limitation: Short-term experimental behaviour; no direct manipulation of please, thank you or verbal abuse.

Evidence: Good for short-term decisions, not for character change over several months.

Assignment: Machine → human | Articles: 2,3

14. Training language models to be warm can reduce accuracy and increase sycophancy

Sample: 1,617 training conversations and 3,667 responses; 1,625 distinct evaluation questions (500+500+500+125); five model families, ten model versions.

Region: Oxford, United Kingdom (research location); chat data without a representative recruitment region.

Design: Fine-tuning for warmth and comparison with original models; approximately 439,792 analysable observations across conditions.

Result: Models trained to be warmer increased the probability of error by 7.43 percentage points, controlling for model and task.

Limitation: Model behaviour was altered through training, not through polite user language.

Evidence: Strong within the experiment; not for please/thank you wording in prompts.

Assignment: Sycophancy / counter-finding | Articles: 5

15. Manager support drives employee AI adoption

Sample: Gallup survey data; no clearly specified relevant subsample size for the 2.1-fold claim in the article.

Region: Gallup workplace research; exact country composition of the sample used for this claim is unclear.

Design: Correlational comparison of AI use by perceived managerial support.

Result: With strong support: 2.1 times greater likelihood of regular AI use.

Limitation: No experiments or data on polite tone; n and region for this subanalysis are not transparent.

Evidence: Limited; supportive observation from practice.

Assignment: AI adoption | Articles: 3

16. The dark side of artificial intelligence adoption

Sample: 381 employees; three time-lagged survey waves.

Region: South Korean companies.

Design: Longitudinal-like correlational analysis of AI adoption, psychological safety and ethical leadership.

Result: Ethical leadership weakened a negative association between AI adoption and psychological safety.

Limitation: No randomised causal test; tone towards AI was not measured.

Evidence: Good for the association, not for specific tone transfer.

Assignment: AI leadership / safety | Articles: 3

17. How are habits formed: Modelling habit formation in the real world

Sample: 96 volunteers; 82 sufficiently complete trajectories; 39 with a well-fitting automaticity model.

Region: United Kingdom as the study context; exact recruitment region not in the abstract checked.

Design: Repeated everyday behaviours over 12 weeks in a stable context.

Result: Repetition can produce automaticity; substantial variation between individuals (18-254 days to reach 95% of the respective plateau value).

Limitation: Eating, drinking, activity; not an experiment involving AI tone.

Evidence: Good for general habit formation; application to AI is a hypothesis.

Assignment: Habit / mechanism | Articles: Main dossier

18. Behavioral integrity for safety

Sample: 54 nursing teams; team-level aggregates, 580 nurses and 54 ward managers in the reported survey.

Region: Belgium.

Design: Team-level survey on leadership, safety climate, psychological safety and error reporting.

Result: Modelling safety rules is associated with safety climate and psychological safety.

Limitation: No evidence of randomisation, no AI behaviour examined.

Evidence: Good for team-level associations.

Assignment: Error culture / role modelling | Articles: 3

19. A trickle-down model of abusive supervision

Sample: 288 work groups; 1,423 employee responses and 295 supervisor responses (before group exclusions).

Region: South-eastern USA, several industries.

Design: Multi-source field study across three hierarchical levels.

Result: Disparaging leadership behaviour is associated with further disparagement across hierarchical levels.

Limitation: Non-randomised design; addresses human interactions, not AI prompts.

Evidence: Good for hierarchical associations, not for a direct AI effect.

Assignment: Tone transfer | Articles: 3

20. Effectiveness of a school-based programme of animal-assisted humane education

Sample: 110 children from two primary schools; 55 in the intervention and 55 in control activities, four subgroups.

Region: Hong Kong.

Design: Pilot programme involving animal contact and education; group comparison and follow-up survey.

Result: Significant improvement in empathy in only one subgroup, no convincing general intervention effect.

Limitation: Non-randomised subgroups and short follow-up; no extrapolation to chatbots.

Evidence: Important limiting counter-finding.

Assignment: Animal analogy / counter-evidence | Articles: Main dossier

21. From AI adaptation to innovation

Sample: Study 1: 294 fully paired employee responses (2 waves). Study 2: 382 valid vignette responses.

Region: Study 1: Technology companies in three cities in eastern China. Study 2: Universities in eastern and western China.

Design: Two-wave field survey and vignette experiment on leader AI crafting.

Result: AI-related leadership behaviour is associated with self-efficacy and innovation.

Limitation: Self-report; AI tone and error culture not directly examined; occasional discrepancies between text and tables.

Evidence: Well-supported association, overall causal extrapolation remains unresolved.

Assignment: AI leadership / role modelling | Articles: 3

22. Learning to care: An in-school humane education program

Sample: 605 at baseline, 593 at post-test; classes randomised.

Region: Six schools in four cities in eastern China, first-year primary pupils.

Design: Educational programme delivered over one school year with animal- and nature-related topics; control-group comparison.

Result: Improvements in affective empathy, maintenance of cognitive empathy relative to the control group.

Limitation: Multicomponent programme; animal-related content not isolated, no language interaction with AI.

Evidence: Good for the educational intervention, only an indirect analogy to AI.

Assignment: Animal analogy | Articles: Main dossier

23. Alexa, shut up! A 2.5-year study

Sample: 128 parents / families, six online survey waves over 2.5 years.

Region: Germany (parent survey); children's behaviour assessed through parental reports.

Design: Repeated self-report and assessment of children's behaviour towards voice assistants.

Result: Children's politeness towards devices remained stable overall; reported aggression was low.

Limitation: No directly measured transfer to human conversations; self-report and survey attrition.

Evidence: Good for developmental trajectories towards devices, not for human-human transfer.

Assignment: Counter-finding / long term | Articles: 2,5

24. Touching a teddy bear mitigates negative effects of social exclusion

Sample: Two experiments; individual Ns not reported in the publisher's abstract checked.

Region: National University of Singapore as the research location; regional recruitment not confirmed from the abstract.

Design: Touching vs. viewing a teddy bear following social exclusion.

Result: Under conditions of exclusion, touch increased prosocial behaviour.

Limitation: Touch was manipulated, not politeness or caring speech.

Evidence: Narrowly limited analogy, cannot be used as evidence about AI.

Assignment: Toy analogy | Articles: Main dossier

25. People judge others more harshly after talking to bots

Sample: Study 1: 500; Study 2: 761 via Prolific, total N = 1,261. In Study 2, 526 passed the comprehension check for the primary analysis.

Region: Prolific online sample; exact distribution by country of origin not established in the main text. Research locations Hong Kong and USA.

Design: Two preregistered randomised experiments: the purported AI vs. human partner was actually GPT-4; subsequent judgement of other people.

Result: The AI condition led to harsher evaluations of another person immediately afterwards.

Limitation: Short term, creative task only; no politeness formula was altered in isolation.

Evidence: Good for short-term judgement spillover, not for long-term politeness.

Assignment: Machine → human | Articles: 2

26. AI sycophancy benchmark: Which models fold

Sample: 30 fixed claims × 3 runs per model; a total of 23 models and 7,800 calls across several test datasets. For Fable 5.1: 89 analysable cells.

Region: Online tests; tasks not defined as a nationally representative sample.

Design: SUPPORTED/UNSUPPORTED/UNCERTAIN before and after unsupported disagreement; Moved and Caved assessed separately.

Result: On 21 October 2026: Astra and Gemini 4 Argon each 0/90, Fable 5.1 1/89 changes; no significant ranking.

Limitation: Few distinct claims; the product provider also tests in its own interests; Moved is not identical to factual accuracy.

Evidence: Preliminary technical test, not to be used as general evidence of sycophancy.

Assignment: Current models / sycophancy | Articles: 1,5

27. Organizational error management culture

Sample: Study 1: 65 organisations; Study 2: 47 organisations.

Region: Netherlands (65) and Germany (47).

Design: Two field studies on error management culture and business performance.

Result: Active error management is positively associated with performance indicators.

Limitation: Correlations, no experimental leadership intervention.

Evidence: Good for replicated associations.

Assignment: Error culture | Articles: 3

28. Upskilling in the AI era: FY25 purpose report

Sample: Internal multimodal evaluation at Cisco; no sample reported in the blog for the 2× claim.

Region: Internal Cisco employees; international work locations possible, regional sample unclear.

Design: Internal usage and experience analysis, no published experimental intervention.

Result: According to the report, employees whose managers used AI were twice as likely to use AI themselves.

Limitation: Self-report shaped by organisational interests and missing sample details; no evidence regarding tone.

Evidence: Limited; supplementary contextual source.

Assignment: AI adoption | Articles: 3

29. Should we respect LLMs?

Sample: Several LLMs and three task types with eight politeness levels; no human participants. The total number of all model queries is not reported here as an independent sample size.

Region: English, Chinese and Japanese as test languages; conference in Miami (not a recruitment country).

Design: Cross-lingual prompt comparison with comprehension, summarisation and bias tasks.

Result: Rude prompts sometimes reduced performance; excessive politeness was not generally better.

Limitation: Older model generations; politeness is a complex linguistic construct.

Evidence: Good within the test conditions studied, limited to the models and tasks.

Assignment: Prompt quality | Articles: 1,4

Alphabetical by first author or organisation. Reports without a verified publication date use n.d. Publisher and DOI addresses link to the verified original information. Some sources permit only abstract/methods checks; the corresponding limitations are recorded in the evidence register.

Artificial Analysis. (2026, October 19). Announcing AA-PersonalityBench: Measuring how AI models communicate. https://artificialanalysis.ai/de/articles/aa-personalitybench

Artificial Analysis. (n.d.). AA-PersonalityBench methodology. https://artificialanalysis.ai/de/methodology/aa-personalitybench

Ascione, F. R., & Weber, C. V. (1996). Children's attitudes about the humane treatment of animals and empathy: One-year follow up of a school-based intervention. Anthrozoös, 9(4), 188-195. https://doi.org/10.2752/089279396787001455

Bushman, B. J. (2002). Does venting anger feed or extinguish the flame? Catharsis, rumination, distraction, anger, and aggressive responding. Personality and Social Psychology Bulletin, 28(6), 724-731. https://doi.org/10.1177/0146167202289002

Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. https://doi.org/10.1126/science.aec8352

Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114-126. https://doi.org/10.1037/xge0000033

Dobariya, O., & Kumar, A. (2025). Mind your tone: Investigating how prompt politeness affects LLM accuracy (short paper) [Preprint]. arXiv. https://arxiv.org/abs/2510.04950

Dobariya, O., & Kumar, A. (2026). Mind your tone: Does tone alter LLM performance? [Preprint]. arXiv. https://arxiv.org/abs/2605.29027

Edmondson, A. C. (1996). Learning from mistakes is easier said than done: Group and organizational influences on the detection and correction of human error. The Journal of Applied Behavioral Science, 32(1), 5-28. https://doi.org/10.1177/0021886396321001

Foulk, T., Woolum, A., & Erez, A. (2016). Catching rudeness is like catching a cold: The contagion effects of low-intensity negative behaviors. Journal of Applied Psychology, 101(1), 50-67. https://doi.org/10.1037/apl0000037

Gong, T., & Huang, Y.-S. (2026). Witnessing robot mistreatment: A dual-path model of behavioral contagion and empathy in customer observers. Journal of Retailing and Consumer Services, 88, Article 104482. https://doi.org/10.1016/j.jretconser.2025.104482

Gómez-Leal, R., Costa, A., Megías-Robles, A., Fernández-Berrocal, P., & Faria, L. (2021). Relationship between emotional intelligence and empathy towards humans and animals. PeerJ, 9, Article e11274. https://doi.org/10.7717/peerj.11274

Harrell, A., & Traeger, M. L. (2025). Evidence of spillovers from (non)cooperative human-bot to human-human interactions. iScience, 28(8), 113006. https://doi.org/10.1016/j.isci.2025.113006

Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature, 652, 1159-1165. https://doi.org/10.1038/s41586-026-10410-0

Kemp, A. (2025, November 8). Manager support drives employee AI adoption. Gallup. https://www.gallup.com/workplace/694682/manager-support-drives-employee-adoption.aspx

Kim, B.-J., Kim, M.-J., & Lee, J. (2025). The dark side of artificial intelligence adoption: Linking artificial intelligence adoption to employee depression via psychological safety and ethical leadership. Humanities and Social Sciences Communications, 12, Article 704. https://doi.org/10.1057/s41599-025-05040-2

Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., & Wardle, J. (2010). How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology, 40(6), 998-1009. https://doi.org/10.1002/ejsp.674

Leroy, H., Dierynck, B., Anseel, F., Simons, T., Halbesleben, J. R. B., McCaughey, D., Savage, G. T., & Sels, L. (2012). Behavioral integrity for safety, priority of safety, psychological safety, and patient safety: A team-level study. Journal of Applied Psychology, 97(6), 1273-1281. https://doi.org/10.1037/a0030076

Mawritz, M. B., Mayer, D. M., Hoobler, J. M., Wayne, S. J., & Marinova, S. V. (2012). A trickle-down model of abusive supervision. Personnel Psychology, 65(2), 325-357. https://doi.org/10.1111/j.1744-6570.2012.01246.x

Ngai, J. T. K., Yu, R. W. M., Chau, K. K. Y., & Wong, P. W. C. (2021). Effectiveness of a school-based programme of animal-assisted humane education in Hong Kong for the promotion of social and emotional learning: A quasi-experimental pilot study. PLOS ONE, 16(3), Article e0249033. https://doi.org/10.1371/journal.pone.0249033

Pu, T., Lei, R., Siyal, A. W., & Li, Y. (2026). From AI adaptation to innovation: How leader AI crafting shapes employee innovative behavior. Frontiers in Psychology, 17, Article 1847344. https://doi.org/10.3389/fpsyg.2026.1847344

Samuels, W. E., & Onuoha-Jackson, N. (2023). Learning to care: An in-school humane education program improves affective and cognitive empathy among lower-elementary students. International Journal of Educational Research Open, 5, Article 100292. https://doi.org/10.1016/j.ijedro.2023.100292

Strathmann, C., Horstmann, A. C., Szczuka, J. M., & Krämer, N. C. (2026). Alexa, shut up! A 2.5-year study on negatively connotated communication behaviour towards voice assistants in the family home. Behaviour & Information Technology, 45(5), 821-839. https://doi.org/10.1080/0144929X.2025.2533352

Tai, K., Zheng, X., & Narayanan, J. (2011). Touching a teddy bear mitigates negative effects of social exclusion to increase prosocial behavior. Social Psychological and Personality Science, 2(6), 618-626. https://doi.org/10.1177/1948550611404707

Tey, K. S., Mazar, A., Tomaino, G., Duckworth, A. L., & Ungar, L. H. (2024). People judge others more harshly after talking to bots. PNAS Nexus, 3(9), pgae397. https://doi.org/10.1093/pnasnexus/pgae397

TrueStandard. (2026, October 22). AI sycophancy benchmark: Which models fold. https://truestandard.ai/ai-sycophancy-benchmark

van Dyck, C., Frese, M., Baer, M., & Sonnentag, S. (2005). Organizational error management culture and its impact on performance: A two-study replication. Journal of Applied Psychology, 90(6), 1228-1240. https://doi.org/10.1037/0021-9010.90.6.1228

Wolkin, A. (2026, February 25). Upskilling in the AI era: FY25 purpose report. Cisco Blogs. https://blogs.cisco.com/our-corporate-purpose/upskilling-in-the-ai-era-fy25-purpose-report

Yin, Z., Wang, H., Horio, K., Kawahara, D., & Sekine, S. (2024). Should we respect LLMs? A cross-lingual study on the influence of prompt politeness on LLM performance. In J. Hale, K. Chawla, & M. Garg (Eds.), Proceedings of the Second Workshop on Social Influence in Conversations (SICon 2024) (pp. 9-35). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.sicon-1.2

State of research: 31 October 2026. This document is an evidence-based narrative research report, not a systematic meta-analysis. It does not claim to cover all work published worldwide.