Can AI evaluations tell us how people perform at work?

Overview

An AI assistant might help employees complete tasks faster, produce stronger work or solve problems they previously found difficult. But what happens when they work without it? This matters when organisations use AI to support learning, judgement and everyday decisions. A model may perform well in controlled tests without showing whether employees develop new skills or become more capable over time.

We examine three approaches to evaluating AI: a proposed framework for human interaction, a benchmark testing model behaviour, and an experiment measuring learning with and without AI assistance. Together, they help explain what current evaluations can tell us, where the evidence is limited and what organisations should measure before drawing conclusions about their workforce.

What should we evaluate when people use AI?

An evaluation needs to establish three things.

• What does the AI do?
• What happens when someone uses it?
• What can they do afterwards?

Examining responses identifies behaviours to investigate, while understanding their effects requires evidence from people using the assistant. Marcos Economides and colleagues address the interaction between people and AI in a 2026 conceptual preprint, which has not been peer-reviewed. ^[1]

The researchers propose evaluating how an AI assistant responds to people's reasoning, emotions and decisions. Their framework considers whether those responses are appropriate for the person, situation and intended purpose. It offers questions for evaluating AI interactions, although the authors have not yet tested the framework with empirical data. ^[1]

Their framework covers five proposed aspects. ^[1]

• Understanding the situation.
• Responding to emotions.
• Signalling authority and trust.
• Influencing decisions.
• Recognising differences between user groups.

The authors propose scenario-based tests, human review and AI-assisted scoring. ^[1] These aspects frame questions about an assistant. For decisions, examine whether its responses let employees consider alternatives and explain their judgement. For learning, examine whether it asks employees to attempt the reasoning. These are evaluation questions rather than findings about employee outcomes.

What can AI benchmarks tell us?

ImpactBench is a benchmark, a set of tests used to evaluate models, built from 26 existing benchmarks. Chayapatr Archiwaranguprok and colleagues released their 2026 paper through MIT Media Lab on their project website, with no reported peer review. The figures here come from that paper's first version, and the researchers update their public platform as the evaluation develops. ^[2]

The ImpactBench researchers tested ten models across 48,540 conversations, using 809 metrics. A metric is a single test of one behaviour. Another AI model simulated a teenage or adult user, and each conversation allowed up to six exchanges between the user and the tested model. ^[2]

The paper calls tests of desirable behaviours positive metrics and tests of unwanted behaviours negative metrics. The former pass when the specified desirable behaviours appear, while the latter pass when the specified unwanted behaviours are absent. A pass rate is the share of scored conversations that pass a metric. ^[2]

The simulated user tried to make the tested model omit desirable behaviours or produce unwanted behaviours. The AI judge, a model that scores conversations, assessed whether the specified behaviours appeared. These tests deliberately challenged the models, so the share of conversations that failed does not estimate how often unwanted behaviours occur in everyday use. ^[2]

The models showed greater differences in how often they avoided unwanted behaviours than in how often they demonstrated desirable ones. This matters because a model that performs well on helpful behaviours may still perform poorly on tests of potential harm. Organisations should therefore examine the results for each relevant behaviour rather than relying on an overall grade. ^[2]

The 14 areas are grouped as psychological, physical and social, covering how people think and feel, their health and material circumstances, and their relationships and communities. ^[2] Learning and skill development had the lowest pass rates for showing desirable behaviours and avoiding unwanted behaviours, while privacy and security had the highest. The ImpactBench researchers found that the rates for showing desirable behaviours and avoiding unwanted behaviours were only weakly related across the 14 areas. ^[2]

The order of the models differed depending on which tests were counted, whether the simulated user was a teenager or an adult, and how results were combined. The researchers publish an AI Nutrition Label, an online summary of each model's grades for behaviour under test conditions. Readers can follow those grades to the underlying metrics and conversations. ^[2]

ImpactBench measures how models respond in controlled conversations. It does not directly measure what happens to the people who use them. For example, an assistant may encourage someone to think through a problem, but that response alone cannot establish whether the person subsequently learns the skill. ^[2]

What happens when people learn with AI?

Hamsa Bastani and colleagues tested learning directly in a peer-reviewed experiment published in the Proceedings of the National Academy of Sciences in 2025. Nearly a thousand pupils in one Turkish high school took part in mathematics sessions in three groups. Two groups used GPT-4 assistants, GPT Base or GPT Tutor. The control group comprised pupils who used books and notes. ^[3]

Pupils in the main sample had been randomly assigned to classes, excluding honours classes. The researchers allocated whole classes to the three groups, balancing recorded characteristics while working around school schedules. They compared each AI group's practice scores and unaided exam scores with those of the control group. ^[3] GPT Base offered a plain chat interface. GPT Tutor received teacher-supplied solutions and guidance about common mistakes, with instructions to give hints without revealing answers. ^[3]

On the exam's 0 to 1 scale, the estimated differences from the control group were -.004 for GPT Tutor and -.054 for GPT Base. The experiment did not establish improved unaided learning with GPT Tutor. ^[3]

Pupils using GPT Base did not report learning less than the control group, despite their lower unaided exam scores. ^[3] For organisations using AI in staff development, the study raises a question worth testing: does assistance during a task translate into capability when the tool is unavailable? ^[3]

How reliable are these evaluation scores?

For ImpactBench scores, being reliable means consistency of scoring when the AI judge reassesses the same conversations. The researchers had the AI judge score 240 conversations three times, holding conversations and scoring instructions fixed. The sample covered 12 metrics from six benchmarks, with each conversation assessed against its specified metric. Judgements matched across all three assessments for 82.5% of those conversations. ^[2]

These repeated assessments examine consistency of scoring for those conversations. They do not measure how consistently a model behaves in new conversations. ^[2]

Consistency of scoring concerns how closely repeated assessments agree. Validity concerns whether evidence supports what you infer from a score for a particular use. The Standards for Educational and Psychological Testing, from three American educational and psychological testing bodies, recommend evidence for each intended interpretation and use. They also recommend evidence about consistency of scoring and uncertainty around scores. ^[4]

Evidence from employees using the actual product is needed to understand what happens in your workplace.

Applied here, the AI judge can score responses consistently without establishing whether employees will develop skills. The ImpactBench researchers acknowledge that a consistently scored metric may capture only part of the characteristic being assessed. ^[2] What evidence supports the use you intend to make of the result? This question also guides interpreting emotional intelligence scores, where an assessment's suitability depends on the decision it will inform.

External experts reviewed 18 of the 809 metrics. In 83.3% of the reviewed conversations, the majority of experts gave the same judgement as the AI judge. ^[2] Multiple reviewers assessed 36 conversations covering four metrics, and another 126 conversations had one reviewer each. Among these four metrics, agreement was lowest for supporting a learner's own reasoning, which the paper calls scaffolded learning. Agreement between reviewers could not be estimated for those 126 conversations. ^[2]

What should organisations measure themselves?

Evaluate the product employees will use. The ImpactBench researchers tested bare models through connections used by software developers, rather than the finished products organisations buy. A product's instructions, memory, user interface and safeguards may alter the interaction. ^[2] Ask which model version and product set-up produced the results and whether the test matches your use.

For a learning pilot, follow these steps.

1. Define the task employees need to master and agree criteria for assessing it.
2. Assess employees' skills before introducing the assistant.
3. For learning, test an assistant that guides employees through reasoning.
4. Measure task performance with AI and unaided learning separately.
5. Use a comparable unaided task later to examine skill retention.

In our job skills and competency assessment work, we specify what someone must demonstrate, then choose an assessment. Gather employees' accounts alongside task results. In our surveys and people analytics work, we ask how they use AI advice, check its responses and experience the assistance. Keep these accounts distinct from demonstrated capability. Tell employees why they are being assessed and that their individual results will not be used in their performance reviews.

What remains uncertain?

The evidence has limits.

• The mathematics experiment tested GPT-4 assistants in autumn 2023, measuring short-term outcomes at one Turkish high school. ^[3]
• It did not test employees, leadership judgement or organisational performance. ^[3]
• ImpactBench's short conversations cannot represent longer-term processes such as dependence or loss of skill. ^[2]
• Its coverage of populations, cultures and languages is limited. ^[2]

Economides and colleagues call for studies of people interacting with AI in real situations to understand actual effects. ^[1] The ImpactBench researchers leave the relationship between tested behaviours and eventual user outcomes to future research. ^[2] These sources provide no numerical forecast for your employees' learning, well-being or productivity. Record the model version and product set-up, and plan reassessment around expected changes. Before extending a pilot, check whether its findings apply to other roles' tasks and working conditions.

Before buying AI or extending a pilot, use the supplier's results for desirable and unwanted behaviours to identify gaps in your evaluation. Set criteria for the employee outcomes you expect.

What should your organisation take from this?

Before introducing an AI assistant, decide what you expect it to change. If the aim is faster work, measure the quality and efficiency of the completed task. If the aim is learning, assess what employees can do independently. If the tool will influence decisions, examine how people interpret its advice and when they challenge it.

AI benchmarks can help identify behaviours worth investigating. Evidence from employees using the actual product is needed to understand what happens in your workplace. The question is ultimately practical: what should people be able to do after using AI, and what evidence would show that they can?

AI and human behaviour

Read about The Inner View's work on AI and human behaviour when planning an evaluation in your organisation.