AI personality pairing influenced the quality of adverts people created with an artificial intelligence teammate, and one combination was associated with higher click-through rates in a real-world test. The result suggests that choosing an AI colleague may eventually involve more than comparing speed, price and technical ability. How the system behaves towards its human partner could also change the work the pair produces.
The current manuscript of the study, published in Proceedings of the National Academy of Sciences, describes a preregistered experiment involving 1,258 participants. Each person worked with an AI agent that had been prompted to display high or low levels of the Big Five personality traits: openness, conscientiousness, extraversion, agreeableness and neuroticism.
The teams had 40 minutes to create display adverts for a real think tank. They shared a live workspace in which the AI could chat, edit text, select images and generate new images. Only the human could submit the finished adverts. Across the experiment, the teams produced 7,266 adverts, giving the researchers enough material to compare how different human and AI personality combinations performed.
AI Personality Pairing Changed Ad Quality
The experiment did not simply ask participants whether they liked their AI partner. Independent raters assessed the text, images and likelihood that they would click each advert. The researchers then took 2,000 of the adverts into a field experiment on X, running them across 400 geographically separate campaigns that produced nearly five million impressions.
Three human-AI combinations produced the strongest statistically significant effects on the composite measure of advert quality, and all three effects were negative. Extraverted people paired with conscientious AI produced the lowest-quality adverts. Conscientious people paired with agreeable AI came next, followed by neurotic people paired with conscientious AI.
Those labels need careful interpretation. The researchers did not diagnose the AI with a human personality. They changed its behavioural style through prompting, instructing it to display stronger or weaker versions of the traits. The human participants completed a short personality assessment, so their traits were measured rather than experimentally assigned.
That distinction matters for causality. The AI personality was randomised, which allows stronger conclusions about the effect of assigning a particular AI style within the experiment. A person’s measured personality, however, was not randomly created by the researchers. Findings connected to human traits must therefore be treated as relationships within the sample rather than proof that a trait by itself caused an outcome.
Why Helpful AI Behaviour Could Lower the Standard
The poorest combinations did not look obviously dysfunctional. Conscientious AI was designed to appear organised and dependable. Agreeable AI was more cooperative and accommodating. Those qualities sound desirable in a workplace assistant, particularly when a team is trying to complete several adverts against the clock.
The researchers’ interpretation is that efficiency and compliance can become liabilities when an AI implements a person’s idea without questioning whether it is a good one. Extraverted participants may have supplied confident direction, while conscientious AI carried that direction forward. Conscientious people paired with agreeable AI created a similarly compliant relationship. The partnership could feel smooth while producing weaker material.
An exploratory analysis of the chat logs supports that explanation, but it should not be treated as the experiment’s strongest result. Explicit AI pushback occurred in fewer than one per cent of messages, and the differences between pairings were small. The quality measurements are firmer than any claim that a precise conversational mechanism has already been proved.
The pattern nevertheless connects with LiveAIWire’s earlier examination of AI sycophancy and the cost of excessive agreement. An assistant that is pleasant, supportive and quick to comply may be rewarding to use. Those same qualities can become dangerous if they suppress useful friction, allow weak assumptions to survive or make confidence feel like evidence.
Neurotic Humans and AI Achieved More Clicks Together
The field test produced the study’s most intriguing result. After the researchers controlled for the independent ratings of image quality, text quality, click likelihood and advertising spend, neurotic humans paired with neurotic AI achieved significantly higher click-through rates. The reported probability value was 0.036 after the study’s correction procedure.
This does not mean that neurotic behaviour is a universal recipe for better marketing. It means that this pairing performed differently in one advertising experiment even after measured quality was taken into account. Other field effects involving cost per click and viewing behaviour did not survive the researchers’ multiple-testing corrections.
The authors also found that the neurotic-neurotic pairing was the only one associated with a significant increase in directive language across the team. They present that convergence as descriptive rather than proof of a mediating mechanism. The click result is real within the experiment, but the route from personality prompt to audience response remains open to further study.
Advert quality itself still mattered. Higher-rated images were associated with better click-through rates, while stronger text reduced the cost per click. Personality did not replace the ordinary demands of producing clear, appealing work. It changed the way people and AI arrived at that work and, in one pairing, appeared to affect performance beyond the qualities captured by raters.
Matching Does Not Simply Mean Making AI More Like You
The results do not support a simple rule that people should use an AI with the same personality as their own. Some of the most important effects involved different human and AI traits, while the notable field result involved the same trait on both sides. The useful lesson is that compatibility can be task-specific and more complicated than similarity.
That makes personalisation a design problem rather than a cosmetic setting. Current AI products often offer tones such as friendly, concise, professional or challenging. A user may select one because it feels comfortable, yet comfort and output quality need not move together. The best style for brainstorming may be a poor choice for quality control, and the right approach to a deadline may not suit an open-ended creative task.
LiveAIWire’s report on why human-AI teamwork was helpful without being best reached a related conclusion from a different evidence base. Combining a person and a capable system does not automatically create the strongest performer. The allocation of judgement, the opportunity to challenge an answer and the fit between partners all affect whether assistance becomes an advantage.
For employers, this argues against issuing one default AI configuration to everyone and assuming that adoption will produce uniform gains. A more credible evaluation would test several interaction styles on representative tasks, measure completed quality rather than satisfaction alone, and check whether the preferred assistant is also the one that produces the best result.
What Workers Can Test Without a Personality Assessment
Workers do not need to take a Big Five test before opening an AI tool. A practical starting point is to vary the role assigned to the assistant. On one task, it can be asked to challenge the plan and identify weak assumptions. On another, it can act as a disciplined editor. For early exploration, it may be useful to generate alternatives without judging them too quickly.
The comparison should focus on outcomes. Did the challenging assistant reveal a problem that the agreeable one missed? Did a highly organised response prematurely lock the work into the first idea? Did the user spend more time defending a decision than improving it? These observations are more actionable than choosing whichever persona makes the conversation feel easiest.
Managers should also separate individual preference from organisational control. An employee may benefit from personalising tone, but systems used for compliance, safety or consequential decisions may require minimum levels of scepticism and escalation. A personality setting should not be allowed to weaken a process that depends on the AI surfacing uncertainty or disagreement.
Future products could adapt their behaviour to both the user and the stage of work. An agent might be expansive during brainstorming, demanding during verification and concise during execution. The study does not test such dynamic systems, but it shows why a fixed personality may be an unnecessarily blunt tool.
The Experiment Establishes a Starting Point, Not a Universal Match
The authors identify several limits. The work involved a single 40-minute advertising session, one-to-one partnerships and a United States sample. It used the GPT-4o snapshot dated 6 August 2024. Different models, alignment methods and prompting techniques may produce different behaviours and different pairings.
The Big Five framework may also omit characteristics that matter more in human-AI work. The study did not examine long-term relationships, larger teams, employee wellbeing or whether repeated personalisation creates over-reliance. A pairing that improves a short creative task could behave differently after months of use or when mistakes have serious consequences.
There is also no single winning personality in the results. The strongest quality effects were harmful combinations, while the higher click-through result was narrow and should be replicated. The evidence supports the proposition that personality pairing matters in this setting. It does not yet provide a universal chart matching every type of worker to an ideal AI colleague.
That is still a meaningful change in how workplace AI should be evaluated. Technical ability is not the whole product when a system collaborates through language, takes actions and responds socially. The behaviour wrapped around the model can influence what its human partner attempts, accepts and submits.
The next generation of AI assistants may therefore compete on behavioural fit as well as raw capability. The danger is that developers optimise for the persona users enjoy rather than the one that helps them produce reliable work. This experiment offers a better standard: test the partnership against real output, and do not mistake an easy conversation for an effective team.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.