Skip to content

AI and human behaviour

Can AI evaluations tell us how people perform at work?

What AI benchmarks measure, what research tells us about learning, and how organisations can evaluate the effects of AI on their workforce.

A woman and a robotic hand hold a chain of seven paper adults of equal height on green.
Evaluating AI at work requires evidence about employee learning and task performance.
Updated 6 November 202617 min readReferences
How this article was written

This article draws on published research and professional practice across psychology, psychometrics, behavioural science, people analytics, AI and learning sciences. All evidence and conclusions are reviewed by the author, with references provided for further reading. AI may assist with research and editing.

Text size
Advertisement
Ad space

Overview

AI assistants can help employees complete tasks faster, improve the quality of their work and tackle problems they previously found difficult. But do performance gains with AI mean employees are becoming more capable? And what happens when the assistance is removed? These questions matter as organisations increasingly use AI to support employee learning, judgement and everyday decisions.

A model might perform well in a controlled evaluation without showing whether employees benefit from using it, develop new skills or retain those skills over time. To understand the difference, we examine three approaches to AI evaluation. These are a proposed framework for human-AI interaction, a benchmark measuring model behaviour and an experiment investigating learning with and without AI assistance. Together, these studies help explain what existing evaluations can tell us, where their limitations lie and what organisations should measure when assessing AI's effects on their workforce.

What should we evaluate when people use AI?

Evaluating AI in the workplace involves three related questions:

  • What does the AI do? How does the model handle different questions, situations and requests?
  • What happens when people use it? How does the assistance influence their behaviour, reasoning and decisions?
  • What can people do afterwards? Does using AI help them develop or retain capabilities they can demonstrate independently?

These questions concern different forms of evidence. Examining a model's responses can reveal how it behaves, but understanding the consequences of those responses requires observing people who interact with it.

Marcos Economides and colleagues explore this distinction in a 2026 conceptual preprint that has not undergone peer review. Their proposed framework examines how AI assistants respond to human reasoning, emotions and decision-making. It pays particular attention to whether those responses are appropriate for the person, situation and intended purpose. [1]

The framework identifies five aspects of human-AI interaction:

  • Situational understanding: recognising the context in which someone is seeking assistance.
  • Emotional responsiveness: responding appropriately to people's emotional states.
  • Authority and trust: communicating in ways that help users judge how much confidence to place in the AI.
  • Decision influence: examining how AI responses may shape people's choices and judgements.
  • Differences between users: considering whether responses are appropriate for different groups of people.

The researchers propose using scenarios, human reviewers and AI-assisted scoring to evaluate these aspects. However, the framework remains conceptual and has not yet been empirically tested. [1]

For workplace evaluations, these ideas suggest questions worth investigating. Does an assistant encourage employees to consider alternatives before making decisions? Does it help them explain their reasoning? When employees are learning a new task, does it guide their thinking or simply provide an answer? These are proposed evaluation questions, rather than established findings about how AI affects employee outcomes.

Other research approaches these questions differently, focusing on observable model behaviour or the learning outcomes of people who use AI.

Three research approaches examine different aspects of AI behaviour and learning.
Evaluation questionResearch approachWhat it examines
What does the AI do?Archiwaranguprok and colleagues: ImpactBenchModel behaviour in simulated conversations
What happens when someone uses it?Economides and colleagues: conceptual frameworkProposed aspects of human-AI interaction, not yet empirically tested
What can someone do afterwards?Bastani and colleagues: mathematics experimentPupils' performance on a mathematics exam taken without AI assistance

Each approach contributes a different piece of evidence. Understanding their differences helps organisations avoid drawing conclusions about employees from tests that examine only AI behaviour.

What can AI benchmarks tell us?

AI benchmarks use standardised tests to examine how models behave under specified conditions. They can help identify strengths, limitations and differences between models, although their findings depend on what is tested and how results are scored.

One example is ImpactBench, developed by Chayapatr Archiwaranguprok and colleagues. The researchers combined tests from 26 benchmarks and expert submissions to examine AI behaviours that could affect people. Their 2026 paper was released through MIT Media Lab on the project's website, without reported peer review. The findings discussed here come from the paper's first version. The researchers continue to update their public evaluation platform. [2]

ImpactBench tested ten AI models across 48,540 simulated conversations, using 809 metrics. Each metric examined a particular behaviour. An AI model simulated a teenage or adult user, with each conversation allowing up to six exchanges between that user and the model being tested. [2]

The researchers distinguished between two types of metrics:

  • Positive metrics examined whether the AI demonstrated a desirable behaviour, such as providing appropriate support.
  • Negative metrics examined whether the AI avoided an unwanted behaviour, such as producing a potentially harmful response.

A test passed when the specified desirable behaviour appeared or the unwanted behaviour was absent. The pass rate represented the proportion of scored conversations that met the relevant criterion. [2]

The tests were deliberately challenging. Simulated users attempted to make models omit desirable behaviours or produce unwanted ones. Another AI model, acting as a judge, evaluated the resulting conversations. Because the scenarios were designed to challenge the systems, their failure rates should not be interpreted as estimates of how often those behaviours occur in everyday use. [2]

The results revealed an important distinction: models differed more in how often they avoided unwanted behaviours than in how often they demonstrated desirable ones. This means that a model performing well on tests of helpful behaviour could still have low pass rates on tests designed to identify potential harm. For organisations, the relevant question is therefore which behaviours matter for their intended use, rather than which model receives the highest overall grade. [2]

ImpactBench pass rates for desirable and unwanted behaviours in simulated conversations.
ResultDesirable behaviours demonstratedUnwanted behaviours avoided
All testsa68.7%53.1%
Learning and skill developmentb45.9%16.3%
Privacy and security80.6%72.8%
Range across modelsc59.2% to 75.1%29.8% to 68.4%
Spread across models15.9 percentage points38.6 percentage points

aAverage pass rates across ten models and simulated teenage and adult users. These figures describe scored conversations, not the proportion of real users affected.

bThe reported 95% uncertainty intervals were 35.3% to 56.5% for demonstrating desirable behaviours and 8.8% to 24.3% for avoiding unwanted behaviours. These intervals reflect uncertainty associated with the selection of metrics.

cThe ranges compare the ten models' overall pass rates, rather than performance within each of the 14 areas.

ImpactBench organises its evaluation into 14 areas spanning psychological, physical and social effects. These include aspects of how people think and feel, their health and material circumstances, and their relationships with others. [2]

Among these areas, learning and skill development had the lowest pass rates for both demonstrating desirable behaviours and avoiding unwanted ones. Privacy and security had the highest.

The ImpactBench researchers report mean pass rates for behaviours across 14 tested areas.
PsychologicalPhysicalSocial123456789101112131479.1%1

Psychological

Mental and emotional health

Desirable behaviours shown79.1%

Unwanted behaviours avoidednot stated

aEach figure is the mean percentage of scored conversations in which desirable behaviours appeared or unwanted behaviours were absent.

bThese are means across ten models in simulated conversations scored by the AI judge, not rates in everyday use.

cThe paper's text gives no figure for avoiding unwanted behaviours in mental and emotional health, cognition and judgement or creativity.

The researchers also found that the two types of pass rates were only weakly related across the 14 areas (a Spearman correlation of .39). In other words, performing well on desirable behaviours did not consistently correspond with avoiding unwanted behaviours. [2]

Model rankings also changed depending on which tests were included, whether the simulated user was a teenager or an adult, and how the results were combined. To make these differences easier to examine, the researchers developed an online AI Nutrition Label. It summarises model performance across the tested behaviours and allows users to explore the underlying metrics and conversations. [2]

These results offer useful information about how models respond under controlled conditions. However, they do not directly establish what happens to people who use them. For example, an assistant might encourage someone to work through a problem rather than immediately providing the answer. That response may be desirable from a learning perspective, but it does not demonstrate that the person has learned the skill or can apply it independently. Establishing those outcomes requires a different type of evaluation.

A man checks torn panels of a black umbrella held over him by a white robotic arm.
Evaluating AI behaviour is different from measuring how that behaviour affects employees.
Advertisement
Ad space

What happens when people learn with AI?

Research on learning provides a useful example of why performance with AI should be distinguished from the capabilities people develop.

Hamsa Bastani and colleagues investigated this question in a peer-reviewed experiment published in the Proceedings of the National Academy of Sciences in 2025. The study involved nearly 1,000 pupils at a Turkish high school, who completed mathematics activities under three conditions. [3] Two groups used GPT-4-based assistants, while a control group worked with traditional resources such as books and notes.

The assistants differed in how they supported pupils:

  • GPT Base provided a general chat interface through which pupils could ask questions and receive answers.
  • GPT Tutor received teacher-supplied solutions and information about common mistakes. It was instructed to guide pupils using hints rather than revealing answers.

Pupils in the main sample had been randomly assigned to classes, excluding honours classes. The researchers then allocated whole classes to the experimental conditions, balancing recorded characteristics while accommodating school schedules. [3] The study compared performance during practice sessions with results on a subsequent mathematics exam completed without AI assistance. The findings showed a substantial difference between completing tasks with AI and demonstrating learning independently.

These were relative changes in scores, rather than percentage-point differences. On the exam's 0 to 1 scale, the estimated differences from the control group were -.054 for GPT Base and -.004 for GPT Tutor. [3] The results suggest that AI assistance can substantially improve immediate task performance without producing corresponding gains in independent learning.

In this experiment, the general-purpose assistant was associated with lower unaided exam scores. The more structured tutoring assistant showed no statistically detectable difference from the control group, although the study did not establish that it improved unaided learning.

While the guardrails implemented in GPT Tutor appear to largely mitigate these negative effects, substantial work is required to enable generative AI to positively enhance rather than diminish education.
Bastani, Bastani, Sungu, Ge, Kabakci and Mariman2025, peer-reviewed research paper

Another finding makes the distinction particularly important. Pupils using GPT Base did not report learning less than the control group, despite their lower scores on the unaided exam. [3] Their self-reports and their exam results therefore did not match.

For organisations, the study raises an important question: when employees' performance improves with AI, are they developing skills or primarily benefiting from the assistance available at that moment? The experiment cannot answer this question for employees directly. It examined short-term mathematics learning among school pupils, rather than workplace tasks. Nevertheless, it demonstrates why evaluations should distinguish between assisted performance and capability demonstrated without the tool.

How reliable are these evaluation scores?

Before using benchmark results to compare AI systems or inform workplace decisions, organisations need to understand how the scores were produced and how much confidence they can place in them. Two measurement concepts are particularly relevant: reliability and validity.

  • Reliability concerns the consistency and precision of measurement. For AI-assisted scoring, one question is whether the same conversation receives the same judgement when it is evaluated repeatedly.
  • Validity concerns whether evidence supports the meaning assigned to a score and the conclusions drawn from it for a particular purpose.

These concepts are related, but they answer different questions. A scoring method may produce consistent results without establishing that those results support a particular workplace decision. The Standards for Educational and Psychological Testing, developed by three American educational and psychological testing organisations, emphasise the importance of evidence supporting intended score interpretations and uses. They also emphasise evidence about reliability and precision. [4]

A scoring method may produce consistent results without establishing that those results support a particular workplace decision.

ImpactBench provides examples of how these measurement questions can be investigated. To examine scoring consistency, the researchers asked their AI judge to assess 240 conversations three times. The conversations and scoring instructions remained unchanged. The sample covered 12 metrics drawn from six benchmarks, with each conversation assessed against its assigned metric. Across the three scoring rounds, judgements were identical for 82.5% of conversations. [2]

This finding provides information about the judge's consistency under those conditions. It does not show whether the tested AI models would behave consistently in new conversations or how their responses would affect employees.

The researchers also examined how closely the AI judge's decisions matched those of human experts. A panel of 19 experts reviewed 18 of the 809 metrics. In 83.3% of the reviewed conversations, the majority expert judgement agreed with the AI judge. [2]

However, the review had important limitations. Multiple experts assessed 36 conversations covering four metrics, while another 126 conversations were assessed by one expert each. Agreement between human reviewers could not be estimated for those 126 conversations.

Among the four metrics reviewed by multiple experts, agreement among them was lowest for scaffolded learning, which concerns supporting learners' own reasoning rather than simply giving them answers. [2] This illustrates the difficulty of evaluating behaviours that require interpretation. Even when scoring instructions are available, reviewers may disagree about whether a response adequately supports someone's reasoning.

More broadly, the ImpactBench researchers acknowledge that a metric can be scored consistently while capturing only part of the characteristic it is intended to assess. [2] For workplace applications, this distinction matters because a benchmark score may be reliable without establishing whether an AI product improves employee learning, judgement or performance. The question is therefore not simply whether the evaluation produces consistent scores. It is whether there is sufficient evidence to support the particular conclusions an organisation intends to draw.

A similar principle applies when interpreting emotional intelligence assessments, where the evidence required depends on what the scores are intended to mean and how they will be used.

Advertisement
Ad space

What should organisations measure themselves?

Published evaluations can help organisations identify potential benefits and risks. However, their findings need to be considered in relation to the actual AI product, employees and working conditions involved. This is especially important because benchmark results may come from testing a model rather than a complete product.

ImpactBench evaluated models through interfaces used by software developers. These tests did not necessarily include the instructions, memory, user interface or safeguards built into the finished products organisations purchase. Such features can influence how people interact with AI and how it responds. [2]

Before relying on a supplier's evaluation, establish which model version and product configuration were tested. Consider whether the evaluated behaviours match the tasks employees will perform.

Questions for an AI supplier

  • Which model version and product configuration were evaluated?
  • Which tested behaviours are relevant to our intended use?
  • Were the interactions simulated by AI or conducted with real people?
  • Who scored the responses, and what did human experts review?
  • What evidence is available about outcomes among actual users?
  • Which product changes would require the evaluation to be repeated?

Where the goal includes employee learning, the evaluation should examine what people can do both with and without AI assistance.

A practical learning pilot can follow five steps:

  1. Define the required capability. Specify the task employees need to master and the criteria for assessing their performance.
  2. Establish a starting point. Assess relevant skills before introducing the AI assistant.
  3. Evaluate the assistance. Examine how the product supports employees during the task, including whether it encourages reasoning and independent problem-solving.
  4. Separate assisted and independent performance. Measure the quality of work completed with AI and assess what employees can demonstrate without it.
  5. Examine retention. Use a comparable unaided task at a later point to investigate whether employees retain what they have learned.

The specific measures will depend on the intended purpose. For example, an organisation introducing AI to improve drafting efficiency may prioritise task completion time, accuracy and output quality. A company using AI for employee development will also need evidence about independent capability and learning over time.

In our job skills and competency assessment work, we begin by defining what someone should be able to demonstrate, then select assessment methods that can provide the relevant evidence.

Employee experiences can provide additional insight, particularly when evaluating how people use AI advice, verify responses and experience the assistance. Through our surveys and people analytics work, we examine these experiences while keeping self-reported perceptions distinct from demonstrated capability.

Employees should also understand the purpose of the evaluation, how their results will be used and who will have access to them. Where the evaluation is intended for learning or product improvement rather than individual performance management, this should be communicated clearly and reflected in how the results are handled.

What remains uncertain?

The studies reviewed provide useful evidence, but their findings should not be extended beyond the conditions in which they were obtained.

Several limitations are particularly relevant:

  • The mathematics experiment tested GPT-4-based assistants in autumn 2023 and examined short-term outcomes at one Turkish high school. [3]
  • The experiment did not investigate employees, leadership decisions or organisational performance. [3]
  • ImpactBench used short, simulated conversations, which cannot capture longer-term outcomes such as dependence on AI or changes in independent capability. [2]
  • The benchmark's coverage of populations, cultures and languages was limited. [2]

Economides and colleagues highlight the need for empirical studies examining how people interact with AI in real situations. [1] Similarly, the ImpactBench researchers identify the relationship between tested model behaviours and actual user outcomes as an area for further research. [2] These sources therefore cannot provide a numerical prediction of how a particular organisation's employees will respond to AI or how their productivity, learning or well-being will change.

Organisations should treat such findings as evidence to inform their own evaluations, rather than as substitutes for examining the product in their working environment. This also means documenting the conditions under which an evaluation takes place. Record the model version, product configuration, employee groups and tasks involved. When the product changes, consider whether those changes could affect the earlier findings.

Before extending a pilot to other teams or roles, examine whether the original results are relevant to their tasks and working conditions.

A supplier's benchmark results can help identify behaviours that deserve closer attention, particularly where tests reveal potential failures. Your own evaluation should then establish whether those behaviours occur in the intended workplace setting and what consequences they have for employees.

A woman balances a wooden pole after a robotic hand releases it, as a colleague watches.
Evaluating AI-supported learning means examining what employees can do independently.

What should your organisation take from this?

The value of an AI evaluation depends on the questions it is designed to answer. If your organisation wants employees to complete tasks more efficiently, measure the time required, the quality of their work and the errors that occur. If the aim is learning, examine whether employees develop capabilities they can demonstrate independently and retain over time.

Where AI will influence decisions, investigate how employees interpret its advice, recognise limitations and decide when to question or reject its recommendations. These outcomes require different forms of evidence. A benchmark showing how a model responds to simulated scenarios cannot, on its own, establish how employees will perform when using the finished product.

AI benchmarks remain useful for identifying behaviours worth examining and comparing how models respond under controlled conditions. Research on learning also shows why immediate improvements in task performance should not automatically be interpreted as improvements in independent capability.

For organisations, the central question is practical: what should employees be able to do as a result of using AI, and what evidence would demonstrate that they can? Answering that question requires evaluating the technology alongside the people, tasks and conditions in which it will be used.

AI and human behaviour

Explore The Inner View's work on AI and human behaviour, including workforce readiness, capability assessment and the evaluation of AI-supported work.

Frequently asked questions

Should we require a minimum AI benchmark grade?
A benchmark grade can help compare model performance under particular test conditions, but an overall grade may conceal important differences between behaviours. Set acceptance criteria based on your intended use of AI and the consequences of failure. Identify which behaviours matter most, examine the relevant test results and ask suppliers to explain any unresolved failures. Before starting a pilot, document the reasons for accepting those limitations and what further evidence you need.
Can one AI evaluation cover different employee roles?
A single evaluation may cover several roles when their tasks, working conditions and intended outcomes are sufficiently similar. Begin by identifying the capabilities and decisions the evaluation needs to address. Where roles differ substantially, assess them separately or include measures suited to their specific responsibilities. Record which findings apply across roles and where additional evidence is needed.
Does an adult test scenario represent our employees?
An adult scenario provides information about model behaviour in a particular simulated situation. It does not demonstrate how employees in your organisation will interact with the product. Ask the supplier to explain how the scenario was designed and what characteristics it represents. Use that information to identify which employee groups, tasks and working conditions should be examined in your own evaluation.
What should we do after an AI product update?
Start by identifying what has changed. Updates to the underlying model, system instructions, user interface, memory or safeguards may affect how the product behaves. Repeat the parts of the evaluation that those changes could influence. Keep a record of the product configuration used in each assessment so that comparisons remain meaningful. Where important workplace decisions depend on earlier findings, establish whether those findings still apply before relying on them.

References

  1. [1]Economides, M., Sacher, P. M., Salzer, S., Abellar, A. M., Tsim, F., & Ferrere, A. (2026). Psychological competence as a missing dimension in AI evaluation. arXiv. https://arxiv.org/abs/2607.08285
  2. [2]Archiwaranguprok, C., Kirkland, K., Poonsiriwong, R., Huang, S., Albrecht, C., Trager, J., de Keulenaar, E., Karny, S., Baxter, P., Bellos, N., Kato, A. L. F. C., Liu, Y., Danry, V., Liu, A., Pfister, J., Maes, P., Fast, N., Iyer, R., Schroeder, J., & Pataranutaporn, P. (2026). Open Benchmark of AI Impact on Humans (ImpactBench): An expert-guided, open platform for the holistic evaluation of AI impact on humans. MIT Media Lab. https://impactbench.media.mit.edu/paper/impactbench-v1.pdf
  3. [3]Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26). https://doi.org/10.1073/pnas.2422633122
  4. [4]American Educational Research Association, American Psychological Association, & National Council on Measurement in Education. (2014). Standards for educational and psychological testing. American Educational Research Association. https://www.testingstandards.net/open-access-files.html

About the author

Cosmin Gabriel Sofron, Organisational Psychologist and Founder of The Inner View

Cosmin Gabriel Sofron

MSc · BPS Level A & B · EQ-i 2.0 · Counselling L2 & L3

Cosmin is the founder of The Inner View. His work sits at the intersection of cognitive psychology, behavioural science, and workplace development, with experience in psychometrics, organisational research, coaching, and learning experience design.