Not Just A Vacancy Notice

When organisations communicate with the outside world, they do so through many different channels.

• Annual reports speak to investors.

• Corporate websites present a carefully curated public image.

• Internal communications shape culture behind the scenes.

• Social media offers a mixture of marketing, engagement and brand positioning.

One of the most revealing forms of organisational communication is often overlooked.

The job advert.

At first glance, a job advert appears to be a practical document designed to attract applicants for a specific role. In reality, it performs a much broader function. Every advert communicates assumptions about what success looks like, which qualities are valued and what kind of person is likely to thrive within the organisation.

In that sense, recruitment adverts are not simply descriptions of work. They are public statements about organisational identity.

Why Recruitment Communication Is Different

Many forms of organisational communication are difficult to compare systematically.

Annual reports can vary widely in purpose and audience. Corporate websites often differ in structure and content. Internal communications are rarely accessible to those on the outside and often impossible to benchmark across organisations.

Recruitment adverts occupy a different position.

They are public, comparable and directly concerned with people – often an organisation’s largest and most expensive investment.

Thousands of organisations publish adverts for similar roles every day. This creates a rare opportunity to examine how organisations communicate with the labour market using a common analytical unit.

A software engineer vacancy can be compared with hundreds of other software engineer vacancies. A healthcare assistant role can be benchmarked against similar roles across multiple employers. Few other forms of organisational communication allow this kind of large-scale comparison.

The Public Face Of Organisational Identity

Job adverts communicate far more than job requirements.

They communicate:

• priorities,

• expectations

• values and

• assumptions about what good performance looks like and what kinds of people are likely to succeed.

For prospective applicants, these messages help shape first impressions long before an interview takes place.

For organisations, they represent a public-facing expression of culture and identity.

This is why recruitment communication matters. It influences who applies, how organisations are perceived and whether employer-brand messages are visible in practice.

Why Start Here?

The Recruitment Language in Transition Research Project began with a more specific question.

The initial analysis focused on signals relating to experience, adaptability, ambition, workplace identity and related behavioural themes.

As the research progressed, however, a broader question emerged.

The answer appears to be: quite a lot.

Recent exploratory analyses suggest that different organisations often communicate distinctive narratives through recruitment language. Technology firms may emphasise scale, innovation and transformation. Healthcare organisations may communicate wellbeing, service and helping others. Different occupations require different skills, but organisations also appear to describe themselves in recognisably different ways.

The Shop Window

Perhaps the simplest way to think about recruitment adverts is as a shop window.

They are not the organisation itself and they don’t tell us everything about culture, strategy or employee experience. However, they are one of the first things potential employees see. They are carefully constructed, publicly visible and designed to persuade people to step inside. That makes them a useful place to begin.

Looking Ahead

The articles that follow explore a series of questions that emerged from the initial analysis.

• How much variation actually exists between organisations recruiting for similar roles?

• Do different employers communicate distinctive identities through recruitment language, or do many draw upon similar behavioural templates?

• As AI-assisted writing and recruitment automation become more common, what happens to organisational voice and distinctiveness?

• And perhaps most fundamentally, do organisations genuinely sound as different as they believe they do?

The goal of this research is not simply to catalogue words and phrases. It is to better understand how organisations communicate identity, culture and expectations through recruitment language, and how those communications compare with the wider market.

Job adverts may not tell us everything about an organisation.

However, they offer something surprisingly rare, a public, comparable and data-rich view of how organisations present themselves when they are trying to persuade people to join them.

Further Reading

This article serves as an introduction to the broader Recruitment Language in Transition research project.

The first two articles in the series have already explored how recruitment language has changed over time and how many different occupations appear to draw upon similar behavioural signals. Future articles will examine organisational distinctiveness, behavioural templates, AI-assisted recruitment communication and the question of whether organisations sound as different as they believe they do.

If you’ve missed the earlier pieces, they’re a good place to start. More findings will follow over the coming weeks.

Part 1: Recruitment in Transition

Part 2: Different Jobs, Same Language

Subscribe now

Series: Recruitment Language in Transition

When Similarity Becomes the Story

In the first article in this series, I compared two collections of UK job adverts gathered seven months apart and found evidence that recruitment rhetoric had evolved in subtle but noticeable ways. Overtly performance-focused language appeared less prominent, while aspirational and development-oriented signals became more visible across many adverts.

That comparison led me to a second question.

If recruitment language is changing, how much variation actually exists between organisations, occupations and workplace contexts?

We might assume that different jobs will communicate different expectations. A graduate trainee programme, an operational role and a management position perform very different functions and require different skills. However, as the analysis continued, many of the same behavioural signals appeared repeatedly across all three.

The more adverts I examined, the less interesting the individual phrases became. What stood out instead was the tendency for the same combinations of ambition, motivation, adaptability and pace to appear together across very different kinds of work.

Many adverts appeared to describe different forms of work using remarkably similar aspirational language, suggesting evidence of recurring behavioural templates.

The question therefore became less about which words organisations use and more about whether different occupations are beginning to draw from a common behavioural vocabulary. That possibility sits at the centre of this article.

What Did the Analysis Show?

One of the most striking patterns across the recruitment-language comparison was not the popularity of any single phrase, but the extent to which certain behavioural signals appeared across many different occupations.

Across both datasets, particular combinations of terms repeatedly appeared:

• driven

• ambitious

• motivated

• fast-paced

• dynamic

• results-driven

These signals appeared across a wide range of occupations, including operational, managerial and early-career roles.

The technical requirements differed. The responsibilities differed. The occupations themselves remained distinct. What appeared more widely shared were certain behavioural expectations associated with employability, motivation and workplace contribution.

The analysis therefore focused not only on the frequency of individual phrases but also on how particular behavioural signals travelled across occupations and workplace contexts.

Rather than remaining confined to specific forms of work, several signals appeared repeatedly throughout the dataset, suggesting the existence of a shared behavioural vocabulary within recruitment language.

Beyond Isolated Buzzwords

The analysis focused not only on individual phrases but on the combinations in which they appeared. Across the datasets, certain behavioural signals repeatedly clustered together, creating recurring patterns that extended beyond any single keyword.

Terms such as:

• motivated

• ambitious

• adaptable

• driven

rarely appeared alone. They frequently appeared within the same adverts, creating recognisable behavioural templates that recurred across occupations and organisational contexts.

The repetition is significant because many of the same behavioural expectations appeared across very different categories of work. Rather than being confined to particular industries or levels of seniority, these signals appeared widely dispersed throughout the datasets.

Organisational Identity at Scale

One of the more interesting implications of this pattern is that recruitment language may increasingly function less as detailed job description and more as organisational identity signalling.

Across many occupations, the same behavioural signals appeared regardless of substantial differences in tasks, seniority and technical requirements. What remained surprisingly consistent was the behavioural profile being presented as desirable.

The analysis suggests that organisations may often be describing different forms of work through similar behavioural ideals, drawing on recurring combinations of ambition, motivation, adaptability and progression.

Why It Matters

If similar behavioural signals appear across a wide range of occupations, the implications extend beyond recruitment language itself.

For organisations:

• Shared behavioural signals may make it harder to communicate what genuinely distinguishes one workplace from another.

• Employer branding depends partly on communicating difference, yet many organisations appear to draw upon a similar pool of behavioural signals when describing the people they want to attract.

For candidates:

• The pattern may help explain why many adverts can feel familiar even when the underlying work differs substantially.

• Distinguishing between workplace cultures, expectations and organisational identities becomes more difficult when the same behavioural signals recur across different occupations and organisational contexts.

Over time, repeated behavioural signals can begin to feel less like the preferences of individual organisations and more like part of the background language of work itself. Expectations around ambition, adaptability, motivation and progression appeared across a wide range of occupations, creating a degree of familiarity even where the underlying work differed substantially.

The findings suggest that certain behavioural expectations are communicated through a relatively small group of widely distributed signals. At the same time, the analysis does not indicate that occupations themselves have become linguistically identical. Shared behavioural language and occupational distinctiveness can coexist.

The analysis cannot determine the role of AI-assisted drafting directly. However, questions about recurring behavioural templates and behavioural expectations may become increasingly relevant as organisations draw upon similar sources of language when producing recruitment content.

The broader question is not whether every occupation uses the same language, but how organisations communicate difference when many behavioural signals appear across large parts of the labour market.

What Comes Next?

This analysis suggests that many behavioural signals appear across a wide range of occupations. However, similarity alone does not tell us whether all occupations are becoming linguistically alike.

Some signals may be almost universal, appearing across much of the labour market. Others may remain strongly associated with particular occupational groups, workplace contexts or forms of work.

The next article explores that distinction by examining how behavioural signals are distributed across occupations and asking a related question: which signals appear almost everywhere and which still help distinguish one type of work from another?

Related Articles

Part 1 Recruitment Language in Transition

Further Reading

This analysis forms part of a broader research programme exploring how AI-assisted systems influence organisational communication, workplace expectations and decision-making.

If these findings resonate with observations or questions within your own organisation, I’d be interested in hearing from you.

Further reports, guides and research publications are available via Beyond the Average.

Explore Applied Guides & Reports

Subscribe now

Series: Recruitment Language in Transition

The Question Behind the Analysis

For the past 7 months, I’ve been analysing UK job advert language as part of a wider research project exploring behavioural signalling and workplace culture in recruitment.

The original October 2025 dataset formed part of the earlier Agents at Work research series (Phases 1 & 2), which examined age-adjacent language and recurring behavioural cues across UK job adverts.

My earlier behavioural evaluation research explored how AI-supported systems behave under repetition and constraint. I became interested in whether similar forms of repetition and recurring behavioural patterns were beginning to appear within organisational language itself as AI-assisted writing tools became more widely used.

My interest was not limited to recruitment itself. Job adverts provided a useful starting point because they are one of the few large-scale forms of organisational communication that are publicly available, regularly updated and accessible for systematic analysis. While organisations generate many other forms of text – including internal communications, customer correspondence, policy documents and marketing materials – these are often difficult to obtain at scale. Recruitment adverts therefore offered a visible and measurable window into how organisations describe themselves, the behaviours they value and the workplace identities they seek to construct.

I collected a second dataset of 5,500 advert texts across the same roles in May 2026 to explore whether recruitment rhetoric had changed over time and whether those changes revealed anything about wider workplace culture.

As the comparison developed, something more interesting began to emerge. The analysis suggested that the adverts were doing more than simply repeating isolated phrases. They were constructing recurring workplace identities and behavioural expectations across the labour market.

When I compared the October 2025 and May 2026 datasets, one pattern stood out almost immediately.

From Pressure to Aspiration

October 2025 dataset

The dominant rhetoric felt overtly performance-oriented. Phrases such as:

• “fast-paced”

• “results-driven”

• “dynamic”

appeared frequently and often clustered together. The strongest co-occurrence patterns created a fairly recognisable workplace identity — high-energy, acceleration-focused and visibly pressure-oriented.

Many of the adverts framed employability through speed, urgency and output. Even when the roles themselves varied, the cultural tone often sounded remarkably similar.

What became noticeable across the dataset was not simply the repetition of individual words, but the tendency for particular phrases to repeatedly appear together. Terms such as:

• “dynamic”,

• “fast-paced”,

• “driven” and

• “results-driven”

frequently clustered within the same adverts, reinforcing a shared rhetorical style centred around acceleration, energy and performance intensity.

May 2026

The overall behavioural emphasis had not disappeared, but the presentation had changed. The later dataset showed stronger growth in phrases such as:

• motivated

• ambitious

• graduate

• trainee

Meanwhile, some overtly energetic language, particularly “dynamic”, weakened substantially.

Different clusters now became more prominent, especially combinations involving “motivated”, “ambitious”, “graduate” and “trainee”, creating a softer but still highly directional aspirational tone.

The strongest phrase pairing in the May dataset became:

“graduate + trainee”

It is maybe not that surprising that those two words co-occur, but the change matters because it alters the emotional tone of the adverts, even when many underlying behavioural expectations remain similar.

The comparison draws on a wider behavioural-signalling framework developed during the earlier Agents at Work research series, combining structured cue detection with interpretive analysis of recurring rhetorical patterns across the adverts.

The project does not examine isolated keywords in abstraction, but explores how repeated phrase combinations, co-occurrence patterns and recurring behavioural signals contribute to broader workplace narratives and organisational identity formation.

The Pressure Didn’t Disappear

The pressure did not vanish but became more culturally polished.

That distinction feels important because recruitment language rarely operates as a neutral description of tasks.

At scale, it begins to function more like a behavioural signalling system. Repeated combinations of phrases create implicit expectations around pace, ambition, adaptability and workplace identity.

Recruitment Language Shift

October 2025 vs May 2026

The comparison suggests that recruitment rhetoric evolved unevenly rather than moving in a single direction. Some overtly energetic signals weakened, while motivational and aspirational language became substantially more prominent.

One particularly notable movement was the sharp decline in the use of “dynamic”, which fell substantially between the two collection periods. This may reflect a broader rhetorical movement away from visibly high-intensity corporate language toward softer aspirational and identity-based signalling. Unlike terms such as “motivated” or “ambitious”, “dynamic” increasingly feels associated with an older style of overt performance rhetoric rather than contemporary development-oriented recruitment language.

Shared Signals Across Occupations

One of the most interesting findings from the comparison was how widely distributed many of these signals had become. Terms such as “driven”, “ambitious” and “motivated” appeared across multiple role categories rather than remaining confined to particular occupations or levels of seniority, which raises an interesting possibility.

Some behavioural signals may function less as occupation-specific requirements and more as broadly shared indicators of employability. Rather than describing the technical demands of a role, they communicate expectations about attitude, motivation, adaptability and workplace behaviour.

What became noticeable across the dataset was not simply the repetition of individual words, but the extent to which the same age-adjacent behavioural signals appeared across very different forms of work.

This does not mean that occupations are linguistically identical. Different roles continue to contain distinct technical requirements, responsibilities and occupational language. However, it does suggest that certain behavioural expectations are communicated through a relatively small group of widely distributed signals.

AI Transformation vs Recruitment Reality

One slightly unexpected finding involved the relative absence of explicit AI-related language across much of the dataset.

Despite the wider public discussion surrounding AI transformation, many adverts outside technical and analytical roles continued to rely far more heavily on traditional behavioural and cultural-fit rhetoric than explicit references to AI capability or adaptability.

Terms associated with motivation, flexibility, ambition and workplace attitude remained considerably more visible than AI-specific language across many operational and professional roles.

That contrast may become increasingly interesting as organisations continue integrating AI tools into everyday working practices.

What Changes Slowly

What emerged most strongly from the comparison was not evidence of dramatic linguistic revolution, but gradual cultural normalisation.

Some of these rhetorical movements may also reflect differences in the composition of advertised roles between the two collection periods.

To explore this possibility further, I compared the salary distributions across both collection periods.

That movement may seem subtle, but subtle shifts are often how labour-market culture changes. Recruitment language evolves gradually through repetition, saturation and normalisation rather than sudden replacement.

The findings do not establish discriminatory intent or recruitment outcomes directly. They do, however, suggest that recruitment language functions as a recurring cultural system and one that shapes how organisations present work, how applicants interpret employability and how workplace expectations become socially embedded over time.

This comparison is only an initial snapshot across two collection periods, but it shows that longitudinal recruitment-language analysis may reveal much more than isolated phrase audits ever could.

What changes over time is often not simply vocabulary. It is the way organisations learn to present expectations as culture.

---

A Question of Distinctiveness

Several of the most common behavioural signals appeared across multiple role categories and occupational groups rather than remaining confined to particular sectors or levels of seniority.

The findings in this article suggest that some behavioural signals are widely shared across occupations. However, the presence of shared signals does not necessarily mean that occupations themselves are becoming linguistically alike.

Different forms of work may continue to retain distinct occupational identities while drawing upon a common set of behavioural expectations.

The next article explores that distinction by examining how behavioural signals are distributed across occupations and asking a related question: which signals appear widely across the labour market and which remain more strongly associated with particular forms of work?

What’s Next

This comparison only begins to scratch the surface of how behavioural expectations are communicated through recruitment language.

Future analysis will explore:

• recurring behavioural patterns across roles

• occupational distinctiveness and shared behavioural signals

• organisational distinctiveness

• aspirational rhetoric over time

• recurring workplace expectation language across the labour market

The more adverts I analyse, the more the language itself appears to reveal broader patterns about work, identity and organisational culture in the age of AI-assisted communication.

This article forms part of an ongoing longitudinal research project exploring behavioural signalling, organisational language and AI-assisted recruitment communication across the UK labour market.

Extended findings editions and research summaries will be developed as the project expands.

Methodological Note

This analysis focuses on a defined set of age-adjacent behavioural and workplace signals identified through a custom research dictionary developed for the project. The findings therefore relate to patterns within this signal set rather than all language used in job adverts.

The analysis is designed to explore how organisations communicate behavioural expectations, career stage assumptions, experience signals and related workplace characteristics. It should not be interpreted as a comprehensive analysis of all recruitment language or all forms of organisational communication.

Further Reading

This analysis forms part of a broader research programme exploring how AI-assisted systems influence organisational communication, workplace expectations and decision-making.

If these findings resonate with observations or questions within your own organisation, I’d be interested in hearing from you.

Further reports, guides and research publications are available via Beyond the Average.

Explore Applied Guides & Reports

Subscribe now

Over the past few months, the Beyond the Average research series has expanded into a broader body of work examining behavioural reliability in AI-supported judgement systems.

The recent working papers explored:

• judgement stability under repeated evaluation
• stable non-resolution
• confidence behaviour
• explanation drift

Alongside the formal papers, newer essays on Medium and Substack have started translating these behavioural patterns into practical governance and operational questions around trust, consistency and AI-supported decision-making in practice.

The current behavioural audit arc is now complete, with the next stage focusing more directly on real-world implementation, operational oversight and behavioural reliability in deployed systems.

You can explore the work below:

• Working Papers

• Medium Articles

• Reports, Applied Guides and Other Publications

• Substack Articles Index
  

This article forms part of an ongoing research series exploring AI behaviour and decision reliability in practice.

A full overview of the work is available on the Start Here page.

Subscribe now

A new working paper is now live on Zenodo.

When Explanations Do Not Stay the Same examines how AI explanations behave under repeated evaluation and whether coherent reasoning reflects underlying judgement stability.

Across repeated evaluations, explanations often remained plausible and highly similar in overall meaning, even in cases where underlying classifications changed across runs. The paper argues that explanation reflects a credible account of a decision in a single instance, rather than evidence of stable reasoning over time.

This paper marks the final working paper in the current behavioural audit research arc within the Agents at Work series.

Available now via the Beyond the Average research collection on Zenodo.

This article forms part of an ongoing research series exploring AI behaviour and decision reliability in practice.

A full overview of the work is available on the Start Here page.

Subscribe now

A new working paper is now live on Zenodo.

Earlier work in the Agents at Work research series examined how AI judgements behave under repeated evaluation, including variation at decision boundaries and patterns of stable non-resolution.

This paper examines a different behavioural question.

Confidence scores are often interpreted as indicators of reliability in AI systems. Higher confidence is generally assumed to reflect more dependable judgement.

However, repeated evaluation showed that confidence remained comparatively stable even in cases where underlying judgements changed across runs.

Across the Phase 4 behavioural evaluation study, confidence values remained tightly clustered, typically around 0.60–0.62, including in cases where classifications moved between adjacent categories such as “Potentially Biased” and “Unclear”.

In some instances, unstable judgements exhibited equal or slightly higher confidence than stable ones.

The paper examines this pattern as a separation between expressed certainty and behavioural stability. Rather than indicating whether a judgement remains stable under repeated evaluation, confidence reflects how strongly a decision is expressed in a single instance.

This distinction matters for how AI-supported decisions are interpreted in practice, particularly in operational settings where confidence may be treated as evidence of reliability.

Read the working paper: When Confidence Does Not Indicate Reliability

This article forms part of an ongoing research series exploring AI behaviour and decision reliability in practice. A full overview is available on the Start Here page.

Subscribe now

A new working paper is now live on Zenodo.

Earlier work in the Agents at Work research series examined what happens when the same AI evaluation no longer returns to quite the same place twice. Variation emerged most clearly at decision boundaries, where classifications moved between adjacent categories under repeated evaluation.

This paper examines a different pattern.

In a subset of early-career job adverts, the judgement did not vary across runs, but nor did it resolve into a definitive classification. Instead, the system consistently returned an indeterminate outcome, remaining fixed in “Unclear” despite repeated evaluation under identical conditions.

The system detected relevant signals, but did not assign sufficient weight to commit to a clearer outcome.

Across these cases, confidence remained stable and tightly clustered, with no evidence of threshold crossing or movement toward more decisive classifications.

The paper introduces this pattern as non-resolution, a distinct behavioural property that sits alongside variation at decision boundaries. Some judgements move under repeated evaluation. Others remain stable, but unresolved.

Both patterns matter for how AI systems are interpreted in practice, particularly in high-risk areas such as finance, health and recruitment.

The specific focus on early-career language emerged through discussion with Alex Voronov, which prompted closer examination of this subset within the dataset.

Read the working paper: When a Decision Does Not Resolve

This article forms part of an ongoing research series exploring AI behaviour and decision reliability in practice. A full overview is available on the Start Here page.

Subscribe now

Series: Part of Beyond the Output — Understanding AI Decisions in Practice

Introduction

There is a particular moment that only becomes visible once the same task is repeated. An AI system produces a judgement that appears entirely reasonable, with an explanation that reads clearly and a line of reasoning that holds together without strain, so that nothing in the output suggests uncertainty or instability. Seen once, there is no obvious reason to question it and the result sits comfortably as a completed decision but that sense of resolution begins to loosen when the task is run again.

The same input is evaluated under identical conditions and the outcome does not remain in quite the same place, not in a way that appears dramatic or immediately problematic but in a way that is just enough to unsettle the assumption that the original judgement was fixed.

From stability to variation

Most of the time, the result does remain consistent. In repeated evaluations of recruitment text, around 80% of classifications remained identical across multiple runs of the same input, which is what gives the initial sense of stability and allows the system to appear reliable when viewed through a single output.

A smaller set of cases does not settle in quite the same way, however. In the remaining one in five cases, the classification moves between adjacent categories while the input itself remains unchanged, so that the difference does not lie in the text but in how it is interpreted at the point of decision.

Where the change happens

What becomes visible through repetition is that these changes do not occur across the entire range of outcomes, but appear within a narrower region in which interpretation is required. In practice, most of the movement sits between categories such as “Potentially Biased” and “Unclear”, where signals are present but not definitive.

At either end of the scale, where the language is more explicit, the outcome tends to remain fixed. Between those points sits a zone in which the system does not fully resolve, and it is here that the classification may move from one category to another without any change in the underlying input.

Where AI Judgements Change

What the system is doing

In these cases, the system often identifies the same underlying cues each time. Across repeated runs, the explanations remain highly similar (average similarity ~0.85+), and the reasoning continues to point to the same aspects of the text, such as tone, pace, or implied experience level.

What changes is not what the system sees, but how that interpretation is translated into a categorical outcome. A decision that rests in one position on one run may settle slightly differently on the next, even though the explanation remains largely the same.

Why this matters

This is easy to miss when evaluation focuses on single outputs, where a response that appears coherent is often taken as evidence that the system is working as expected.

What the evidence shows is that confidence does not reliably signal this instability. In many of these borderline cases, confidence scores remain tightly clustered – often around the same level across runs – even when the classification itself changes.

Once those same cases are observed more than once, a different pattern begins to emerge, in which small differences in interpretation accumulate, particularly in situations that already sit close to a threshold.

A different way to think about reliability

What this suggests is not that the system is behaving erratically, but that it has not fully settled, with variation appearing in a specific region rather than across the entire range of outcomes and becoming visible in the relationship between one run and the next. Reliability, in this sense, cannot be assessed solely through individual outputs, but needs to be understood as a pattern of behaviour that emerges through repeated observation.

Conclusion

When the same task is repeated, the outcome does not always remain the same, and while that may appear concerning at first it reflects something more specific about how these systems interpret uncertain inputs. Repetition does not introduce change so much as make visible the places where the decision does not fully resolve.

Final note

This piece introduces the pattern in plain terms. The full analysis, including stability rates, confidence behaviour, and cross-run comparisons, sits in the working paper When Judgement Doesn’t Stay the Same, which examines how these changes appear under repeated evaluation and is available here

Subscribe now

Looking more closely at early-career roles

A LinkedIn comment from Alex Voronov earlier this week pointed me towards recent Stanford research on a redistribution in employment within AI-exposed roles, particularly away from early-career positions.

That work looks at labour market outcomes but it raised a slightly different question for me.

Not what happens when people are hired but what happens earlier in the process, when roles are described and then evaluated.

If early-career roles are changing within AI-exposed parts of the market, how do they behave when an AI system is asked to evaluate them?

From single outputs to repeated evaluation

Over the past few months, this work has been exploring how AI systems behave when asked to make repeated judgements. Instead of treating a single output as meaningful in isolation, the same task is run multiple times under identical conditions to see whether the judgement remains stable.

In earlier work, this revealed a consistent pattern. Where interpretation is required, the judgement does not always remain the same. Seen once, each output appears reasonable but observed across repeated evaluation, variation begins to appear.

This behaviour was visible across a new sample of 150 UK job adverts.

Isolating the early-career subset

Within that dataset, around a third of adverts – 52 out of 150 – contained explicit early-career signals such as “junior”, “graduate”, “trainee”, or “entry level”. These were isolated and evaluated separately using exactly the same setup as before, with the same model, the same prompt and three repeated runs per advert.

The expectation was that, if these roles sit closer to the boundary, they might show more instability.

The analysis showed that they did not.

Note: This result is based on three repeated runs per advert, which is sufficient to observe consistent behaviour, although further runs would be needed to test how persistent this pattern remains under extended evaluation.

A pattern that does not resolve

Across all 52 early-career adverts, every single case was classified as “Unclear”, and every single case remained stable across repeated runs. None of them moved into “Potentially Biased” or any other category and there was no evidence of judgement drift or threshold crossing.

Confidence followed the same pattern, remaining tightly clustered around 0.6, with very little variation and no high-confidence outputs.

What this means in practice is quite simple.

The system is not struggling to repeat the judgement but it is unable to resolve the judgement.

What behaves differently elsewhere

When the same analysis is applied to the remaining adverts, a different pattern appears. Around 29% of non-junior roles show variation across runs, with some cases moving between categories and a small proportion producing more decisive classifications such as “Potentially Biased”. Confidence in these cases is also more varied, occasionally reaching higher levels.

The contrast is not dramatic in volume but it is clear in behaviour. Some inputs produce instability, particularly at the boundary.

Not instability, but constrained judgement

The early-career subset does not fluctuate, drift, or become more confident over time. Instead, it remains in a fixed position, consistently unclear, with moderate confidence, across repeated evaluation. This is not instability in the usual sense. It is a different kind of behaviour, where the system detects signals but does not reach a threshold that would allow it to commit. Other cases produce something else entirely.

Where this sits in a wider context

The Stanford research uses payroll data to examine how employment is changing in roles exposed to generative AI, and finds not a broad collapse in jobs but a redistribution within them, with early-career workers declining relative to more experienced workers, particularly in roles where tasks are more routine and easier to automate.

This analysis does not examine recruitment outcomes and it does not establish any causal relationship but it does sit in a similar part of the system. Early-career roles, as described in job adverts, appear to occupy a more ambiguous position, producing weaker signals for classification and remaining below the threshold required for a decision.

What becomes visible when behaviour is observed

Most discussions of AI evaluation still focus on whether a single output looks reasonable. Seen once, these outputs do look reasonable. Seen repeatedly, something else becomes visible. Some cases fluctuate, particularly where interpretation is required and others remain fixed. Some never move at all, not because they are stable in the sense of being certain but because they remain unresolved.

Closing observation

Not everything that sits at the boundary behaves in the same way. Some judgements change when repeated. Others do not change, but they do not resolve either. That difference is easy to miss when you only look once.

This is not instability, but it is not neutral either. What that leads to in practice is less obvious and I explore that separately in the next article on this topic.

References

Brynjolfsson, E., Chandar, B. and Chen, R. (2025) Canaries in the Coal Mine? Stanford Digital Economy Lab.

Further Reading & Resources

This article forms part of an ongoing research series exploring AI behaviour, operational decision systems and behavioural reliability in practice.

The wider work examines how AI-supported judgements behave under repetition, ambiguity, comparison and operational constraint across real-world environments.

Applied guides, downloadable reports and research editions are available via Payhip.

Explore Applied Guides & Reports

Subscribe now

Over the past several months, this work has been developed as a series of linked studies examining how AI systems detect, interpret and evaluate age-related bias in recruitment language.

Today, the fourth phase of this work is now complete.

This latest stage moves beyond observing variation to examining how AI judgements behave when tested under repeated evaluation, constraint and comparison.

The series as a whole traces a progression in focus.

It begins with identifying patterns in data, moves through interpretation and explanation, and arrives at the point where those judgements can be tested as behaviour over time.

For ease of reference, the full series is set out below.

Phase 4 represents the point at which this work moves from explanation to structured testing.

Phase 1 — Detection
Analysis of age-adjacent language across UK job adverts at scale.
→ [Phase 1 Report]

Phase 2 — Interpretation
Examination of how an AI system identifies and explains age-related signals.
→ [Phase 2 Report]

Phase 3 — Behavioural Audit
Introduction of a behavioural evaluation framework, testing stability, confidence and explanation under repetition and ambiguity.
→ [Phase 3 Report]

Phase 4 — Testing Under Repetition and Constraint
A structured behavioural audit examining how AI judgements hold under repeated evaluation, cross-model comparison and reduced context.
→ [Phase 4 Report]

Phases 1 - 4

Complete research arc of the full progression from detection to behavioural audit. Includes ebook and guide.

→ [Phases 1 to 4]

Together, these phases trace the development of a behavioural approach to AI evaluation, moving from outputs to observed system behaviour over time.

Subscribe now

I’ve published a new piece on Medium reflecting on how my work developed from analysing age bias in job adverts to examining how AI systems behave under repeated evaluation.

The turning point came when I began repeating the same evaluative task under identical conditions and observed that the same judgement does not always remain the same.

The article sets out how this led from output-based evaluation towards a behavioural perspective, and connects to the working paper on behavioural evaluation of AI judgement systems.

You can read it here:

Subscribe now

This working paper introduces a framework for evaluating the behavioural reliability of AI judgement systems.

Summary

The paper brings together observations from Phase 1–3 of the Agents at Work study, which examined how AI systems interpret age-coded language in job adverts.

While the initial work focused on bias detection, repeated evaluation revealed a different issue. AI systems do not always produce stable judgements when the same task is performed multiple times.

The framework proposed in this paper focuses on three ways of observing system behaviour:

• repeated execution of the same evaluative task

• examination of internal signals such as confidence or agreement

• comparison across independent systems

Together, these perspectives make it possible to observe patterns such as judgement stability, convergence, drift, and fragmentation.

Access

📄 Working paper (Zenodo) : [A Behavioural Evaluation of AI Judgement Systems]

📄 Summary note: Why a Single AI Answer Tells You Very Little

This paper forms part of a broader programme of work examining how AI systems behave under repeated evaluation and controlled variation.

Part of the Agents at Work series exploring behavioural evaluation of AI systems.

If you’re working with AI systems and want to explore how this type of behavioural evaluation might apply in practice, this work can also be used to support independent audit and advisory discussions.

Subscribe now

That report examines how an agentic AI system behaves when asked to make the same judgement repeatedly, under constraint and ambiguity. It focuses on stability, explanation drift, confidence behaviour and internal self-review.

Alongside that work, I’ve produced a short ebook, From Explainability to Behaviour, which sets out the behavioural perspective that underpins the report.

The companion ebook brings together the full essay sequence for the series and is intended as an orientation rather than a methods guide, ahead of the individual essays being published in full. It looks at why familiar signals such as confidence, explanation, and agreement can quietly lose meaning once behaviour over time becomes the object of attention.

Readers can refer to it alongside the report, or read it independently as a way into the work and you can get this book either:

• Free for my subscribers from the Resources Library - subscribing is free (see note below)

• As a companion book with the Phase 3 Report - Agents At Work

• Available separately

Subscribers - access details are on the subscriber download page. If you subscribed before the library page was updated or the link does not work and you cannot access it, please get in touch and I’ll send the link to you directly by email. Subscribing to my Substack is free.

Subscribe now

Introduction

Over the past several months, my writing here has been exploring a change in how AI systems are evaluated.

Much of the current conversation still treats explainability as the primary route to trust. If a system can articulate why it reached a decision, that explanation is often taken as evidence that the judgement itself is sound.

That assumption becomes less stable once systems move beyond single outputs.

In agentic workflows that plan, revise, reflect and sometimes disagree with themselves, explanation alone provides only a partial account. What matters increasingly is not only what a system says about a decision, but how it behaves when the same judgement task is repeated under similar conditions.

Phase 3 is released into that context.

From Interpretation to Behaviour

The Agents at Work research series has approached this problem in stages.

Phase 1 examined whether age-adjacent language appears in UK job adverts at scale.

Phase 2 analysed how an AI agent reasoned about those signals once detected – how it interpreted language, justified classifications and framed its conclusions.

Phase 3 changes the emphasis again.

Instead of asking what the system sees or how it explains itself, it asks a different question:

When the same judgement is made repeatedly, does the system behave in a stable, interpretable, and meaningfully overseen way?

This phase does not assess bias prevalence. Instead, it introduces a behavioural reliability and oversight framework for evaluating agentic AI systems, focused on how judgement unfolds over time, under repetition, variation, and constraint.

The report demonstrates this framework through an applied audit of age-bias detection in recruitment text. The case study is not the contribution on its own; it provides a concrete setting in which the framework’s behavioural lenses can be observed in practice.

The Phase 3 Framework

The Phase 3 framework is designed to examine how judgement systems behave under conditions that more closely resemble real operational use.

Rather than treating variation as noise to be averaged away, the framework treats it as a behavioural signal – something to be observed, characterised, and interpreted over time.

It examines system behaviour across seven dimensions:

• stability of verdicts across repeated runs

• changes in explanations over time

• confidence behaviour under ambiguity

• consistency in linguistic cue identification

• agreement and disagreement across models

• responsiveness of internal self-review and agreement signals

• sensitivity to truncated or degraded input context

Together, these dimensions form a structured way of examining whether a system’s behaviour remains dependable when judgement is repeated, constrained, or placed under mild stress.

Although Phase 3 is demonstrated through job advert analysis, the framework itself is not specific to recruitment or age-related language. The behavioural dimensions it examines are properties of judgement systems rather than of any particular domain. Domain-specific elements, such as input type or cue taxonomy, change. The behavioural structure does not.

For this reason, the framework can be applied to other settings in which AI systems evaluate people or content, including moderation, risk assessment and decision support.

Why This Matters

AI systems are increasingly used to evaluate people.

In these settings, fluent explanations and confident outputs are often treated as sufficient signals of trustworthiness. Yet explanation alone does not establish reliability.

Behavioural variation can appear across runs even when explanations remain fluent, and oversight mechanisms can fail to respond when instability emerges elsewhere in the system.

The Phase 3 framework provides a reference point for examining these behaviours at a moment when explainability, on its own, is no longer enough.

About the Report

Agents at Work – Phase 3: How AI Agents Change Under Repetition, Ambiguity and Context Loss is now available as a full 50 page research report.

The report introduces and applies a behavioural reliability and oversight framework for evaluating agentic AI systems, demonstrated through an applied study of age-bias detection in recruitment text.

Its focus is on observable system behaviour rather than outcomes. It does not assess employer intent or real-world discrimination. Instead, it examines how judgement systems behave when decisions are repeated, constrained, or subject to partial context loss – conditions that more closely resemble real operational use.

The report is intended for researchers, auditors, policymakers, and practitioners working on AI reliability, governance, and ethical evaluation.

The abstract and table of contents are available in the preview. The complete report can be accessed on the link below, together with From Explanation to Behaviour, a short orientation ebook that sets out the behavioural perspective behind the report.

👉 [Agents at Work Phase 3 Report plus Ebook]

What Follows

The essays in the Ethics and AI Tools sections will continue to develop the framework introduced here, examining individual behavioural dimensions – including stability, explanation drift, confidence behaviour and oversight responsiveness – in greater detail.

Phase 3 is not a conclusion. It is a methodological anchor for the work that follows.

Thank you for reading – and for staying with the research as it develops.

Imogen

If this work is useful or relevant to you, subscribing is a simple way to signal that it’s worth continuing.

Subscribe now

Phase 2

When I first built the agentic workflow for Phase 1 of this project, I expected the headline findings to be about adverts, which ones were youth-coded, which ones leaned on seniority, where mature-coded cues appeared. And we did learn a lot from that stage.

But something interesting emerged in the process:
every verdict the agent gave came with a short explanation, a little window into why the model thought an advert might contain age-coded language.

Those explanations turned out to be far more revealing than I expected.

Phase 2 of the research turns the lens inward. Instead of re-analysing the adverts, I analysed the agent’s reasoning itself.

I wanted to understand the patterns behind its decisions, the types of linguistic signals it consistently responded to, and how those signals varied across different job roles.

And the picture that came back is surprisingly human.

What the agent actually sees

Every explanation from Phase 1 was grouped into broad categories:

• youth-coded descriptors,

• mature-coded cues,

• experience-implies-age signals,

• dynamic/fast-paced language,

• exclusionary implications and more.

These categories were derived directly from the agent’s own patterns, not from any external lexicon.

Examples from the agent’s reasoning

To make these categories more tangible, here are a few examples taken directly from the agent’s own explanations (using the anonymised text summaries, not original advert text).

Youth-coded descriptors

Example:

“The job advertisement may be biased towards younger candidates based on wording.”

What the agent flagged:
The agent flagged tone suggesting pace, vibrancy or “young energy” even without explicit terms.

Experience-implies-age signals

Example:

“The job description assumes prior career progression and advancement.”

What the agent flagged:
This reflects an expectation of a long, linear career path, which the agent reads as an age-linked cue.

Mature-coded cues

Example:

“May imply a requirement for long experience or established professional maturity.”

What the agent flagged:
The agent associates steady, dependable or seasoned language with older applicants.

Exclusionary implication

Example:

“The wording suggests who might ‘fit’ the role more than what the role requires.”

What the agent flagged:
Subtle tone or framing, not specific words, hints at an unspoken preference.

We can argue about what the agent decided to flag but these examples help illustrate the kinds of signals the agent reacted to, which helps us understand the process. When we zoom out across all 11 job roles, clearer patterns start to emerge.

Summary of the Results

Across all 11 job roles, the results were remarkably consistent:

Exclusionary implication was the most common cue

One cue stood out above the rest: exclusionary implication. Across every role, it appeared more often than any other pattern, making it the most consistent form of implicit age signalling.

There was a steady presence of youth-coded cues across the dataset

Sales, customer-facing and growth roles displayed especially strong youth-coded signals, but elevated levels appeared across the entire dataset. Terms linked to pace, vibrancy and energy shaped a large proportion of perceived age signals.

Experience-implies-age remained steady across all roles

Often framed as “seasoned”, “established”, or requiring long tenure. It’s not inherently discriminatory, but it contributes to an unspoken age narrative around seniority.

Mature-coded cues appeared in care and stability-oriented work

These were lower in volume but consistent, calm, dependable, emotionally resilient and forming a small but recognisable cluster.

Neutral language was rare

Across 5,500 adverts, genuinely age-agnostic phrasing hardly appeared. Even completely ordinary descriptors tended to carry tone, texture or subtle expectations.

Tone, not terms

One theme came through strongly in this analysis:

age-coded signalling rarely comes from explicit wording. It comes from tone.

Pace, personality, cultural expectation. The “feel” of the advert and not the phrases we usually watch for.

This is why lexicon-based approaches often fall short, and why agent-based methods can offer richer insight. The agent is reacting to shading, implication and framing, much like a human reader does.

Why this matters

Age bias in recruitment isn’t usually deliberate. Most hiring managers don’t sit down to exclude people. But language carries cultural associations and those associations shape who feels they’re a good fit.

This analysis offers three important reminders:

• Bias hides in expectations, not only words
• Role norms carry their own age signals (e.g. fast-paced for sales, mature-coded for care roles)
• Ambiguous cues still shape confidence and sense of belonging

These patterns create a landscape of unspoken age signalling that affects people across the lifespan, from early-career applicants to those re-entering the workforce later in life.

The report

The full supplementary Phase 2 report is now available. It expands on the taxonomy, includes proportional cue charts, role-level distributions and extended notes on each category.

📄 Download: Agent at Work – Age-Coded Reasoning in UK Job Adverts (Phase 2)

Phase 1 and 2 Together

🧭 What’s next (Phase 3)

Phase 3 will examine the reliability, stability and explainability of the agent itself, in effect the development of an audit including:

• multi-run stability tests
• rationale and semantic drift
• confidence calibration
• consistency of cue assignment
• cross-model comparisons
• reviewer-agent oversight simulations

If Phase 2 tells us what the agent sees, Phase 3 will tell us whether we can trust how it sees it.

Bias is rarely loud, but its impact is huge. My hope is that this work nudges us a little further toward age-inclusive AI and age-inclusive recruitment.

This work is currently free and forms part of an ongoing programme of independent research into AI behaviour, reliability, and ethics.

If this piece has been useful, you can signal support by subscribing or leaving a comment. This helps me understand whether this research is worth sustaining and extending.

Subscribe now

When people think about bias in recruitment, they often imagine something overt, a phrase that clearly points to age, gender or ethnicity. But that’s not how bias works today. Most of it is woven into tone, implication and expectation. And unless someone deliberately goes looking for it, it sits in the background shaping who feels “right” for a role.

Phase One of my research project started with a simple question:

What signals do job adverts really send about age?

I wasn’t interested in catching anyone out. Most hiring managers never intend to exclude people. But language carries cultural assumptions, and subtle phrasing can nudge older applicants away long before they even think about clicking “Apply now”.

So I collected 5,500 job adverts from different sectors across the UK and ran them through an agentic analysis workflow that evaluates the language rather than the employer. The workflow behaves like a small analytical agent by reading, interpreting and summarising each advert in a structured, repeatable way.

What I found wasn’t shocking, but it was illuminating.

The patterns were everywhere and often unintentional

Age-coded language appeared in every sector I examined. In some cases it was obvious; in most cases it was not.

A few examples:

• “Fast-paced, energetic team” – common across sales, marketing, healthcare

• “Young, dynamic environment” – still surprisingly present

• “Digital native” – a quiet way of saying “under 40”

• “Career starter / Junior mindset / early in their journey” – fine when used appropriately, harmful when misapplied

• “High-pressure, always-on culture” – implicitly exclusionary to people with health conditions or caring responsibilities, who are disproportionately older

None of these phrases directly mention age. But they pull the centre of gravity toward youth and they do it repeatedly, consistently and often without anyone noticing.

Older-adult-friendly language was noticeably rarer

This was the most interesting part.

Across thousands of adverts, there were far fewer references to:

• experience

• judgement

• leadership maturity

• calm decision-making

• reliability

• transferable expertise

• career reinvention

• multigenerational teams

It’s not that employers don’t value these qualities, they clearly do. But recruitment language rarely reflects that value back to candidates.

This gap, more than anything else, can shape whether someone over 50 feels welcome before they even read the job description.

Borderline cases showed mixed signals

Not all adverts were exclusionary. Many sat in a middle ground:
positive on experience, but paired with phrases about speed, culture or energy that complicated the message.

These borderline adverts are important. They reveal the tension between two competing ideals:

• wanting experienced professionals

• wanting youthful energy, fresh thinking and cultural “fit”

The language reveals an implicit dilemma that organisations often haven’t yet named.

Why this matters (and why Phase Two became necessary)

At first, Phase One felt complete. I had enough evidence to show:

• how age-coded language is distributed

• which phrases tend to appear together

• which sectors lean more heavily on certain descriptors

• where ambiguity arises

• how older applicants might interpret these signals

However, as I worked through the data, I noticed something else:

The behaviour of the AI analysing the adverts was just as interesting as the adverts themselves.

Some ads produced consistent interpretations and others produced slight shifts depending on emphasis. A tiny number produced real divergence.

Not because the system was faulty but because the language itself was.

That realisation sparked Phase Two, where I turned the lens onto the AI and asked how stable, transparent and ethically cautious it was across repeated evaluations.

Phase Two, exploring reliability, explainability and ethics is coming early in 2025.

Phase One remains the foundation of the work. It shows that age-coded patterns are not rare, not isolated and not malicious, more that they’re structural.

Where this research goes next

This report triggered a whole new line of work for me:

• Phase Two: stability, explainability and ethics

• Phase Three: deeper drift analysis

• Sector-specific language clusters

• Human vs agent comparisons

• Practical guidance for inclusive hiring language

My goal is not to “score” adverts but to help organisations understand the subtle ways language shapes opportunity and to build AI systems that evaluate that language responsibly.

The agentic framework can be extended to analyse many more instances across a wider range of sectors and also adapted to different domains completely, so this is a starting point for much more research.

Download the Phase One Report

If you’d like to read the high-level findings and see how different sectors compared you can get my report here:

👉 Phase One — Age-Bias Analysis of UK Job Adverts

It’s written to be accessible, clear and practical, with no code, no jargon.

Thanks for reading, and thank you for supporting a research project that sits close to my heart. Please get in touch if you work in this important area and want to connect or collaborate.

Bias is rarely loud, but its impact is huge. My hope is that this work nudges us a little further toward age-inclusive AI and age-inclusive recruitment.

– Imy 🌿
Beyond the Average

Beyond the Average – Age-Inclusive AI Project. I don’t fill your inbox every time a post goes live, just drop by to see what is new. Subscribe to hear about new books, research and reports only.

Subscribe now

Today I’m releasing a new research report in the Age-Inclusive AI series, this time looking at how image-classification models handle age.

Most conversations about bias in computer vision focus on gender or race. Age sits in the shadows, even though older adults often face higher error rates, failed recognition, and systems that simply weren’t trained with them in mind.

So I went back to a very basic question:

If we balance the dataset, does fairness actually improve?

The Experiment

To explore this, I trained three versions of the same ResNet-18 model on the FairFace dataset:

• Unbalanced – reflecting the natural youth skew in the data

• Balanced – equal numbers of younger and older adults

• Balanced & Tuned – with some deeper layers unfrozen so the model can learn age-specific features

A quick note on the model (in plain English):

I used a model called ResNet-18, which is a standard image-classification model widely used in research. You don’t need to know the maths, but here’s the idea:

• Think of an image model as a system that learns to spot patterns

• Early layers learn very simple things (edges, shapes, textures)

• Deeper layers learn more meaningful features (eyes, wrinkles, hairlines, expressions)

• By the time you reach the end, the model has built up enough understanding to make a prediction about age

ResNet is popular because it’s lightweight, reliable and interpretable. It doesn’t try to do anything fancy. It simply learns patterns layer by layer.

What makes ResNet special is the “shortcut connections” inside it. These help the model avoid getting confused or “forgetting” what it learned earlier, which makes training much more stable.

So when I talk about “unfreezing deeper layers”, it just means letting the model learn more detailed age-related features instead of keeping those layers fixed.

Nothing exotic, just a solid, trusted model that helps show the effect of data balance very clearly.

What Did the Experiment Show?

The results were more dramatic than I expected.

The unbalanced model produced a 57-point fairness gap between younger and older adults.

• Simply balancing the data dropped this to 4 points.

• A small amount of fine-tuning pushed performance to near-parity.

No exotic architectures, no clever tricks, no heavy constraints, just better representation.

It’s a reminder that many fairness failures are, at their core, visibility failures.

• Who gets seen?

• Who gets included?

• Who is missing from the data that shapes modern AI systems?

This new FairFace study builds directly on my earlier UTKFace experiments and my image-generation work on Ageless Archetypes. Together, they’re forming a growing evidence base for why age must sit alongside gender, race, disability and other protected characteristics in responsible AI practice.

If you’d like to explore the methods, visuals, or results, you can get the full report here:

[Research Reports: Age Bias in Image Models — FairFace Edition + UTKFace Report]

Thank you, as always, for reading. And if you’re someone who doesn’t always feel “seen” by technology, I see you and this work is for you.

If you’re a researcher, educator or practitioner and want to discuss the methods or explore collaboration, I’d love to hear from you.

Note: If you’d like the first report in this series (UTKFace Edition), it is available FREE with the download of this new report and also to subscribers on the Resource page in the welcome email.

If you are already a subscriber and do not have the up-to-date link, or the link is not working, please email me at ageinclusive.ai@gmail.com if you are interested in this report and I can send it to you.

Beyond the Average – Age-Inclusive AI Project. I don’t fill your inbox every time a post goes live, just drop by to see what is new. Subscribe to hear about new books, research and reports only.

Subscribe now

This study continues my work on visual age bias, but focuses on a different stage of the AI pipeline. The earlier project ‘Ageless Archetypes’ examined bias in AI image generation and how models imagine age when prompted. This one looks deeper, at bias in recognition models and how they perceive and classify age from real data.

Free report available for subscribers.

Introduction

Most AI image systems “see” the world through data and like us, they learn what they see most often. That simple truth hides a quiet problem: most of the faces they’re trained on are young.

I wanted to see what difference that makes.

Would an AI model trained mostly on younger faces perform just as well when it encounters older ones? Or would its accuracy and fairness start to slip?

A Small Experiment with Big Implications

For this study, I used a public dataset of 24,000 facial images labelled by age. Around 70% of those faces were under 40.

To test the impact of balance, I trained two versions of the same model:

• one using the young-biased dataset (mostly under-40 faces)

• and one using a balanced dataset (equal numbers of younger and older adults).

Nothing else changed - same architecture, same training, same evaluation.
Only the data mix shifted.

A quick technical note
Both models used the same deep-learning architecture — a ResNet-18 convolutional neural network, built in PyTorch. It’s a standard image-classification model that analyses visual features layer by layer. I didn’t change the model at all between runs; only the data balance.

Performance was measured with common data-science metrics: recall, precision, and F1 score — all computed separately for younger and older age groups. In plain terms, recall tells us how many real examples the model recognised correctly, while precision checks how often it was right when it made a prediction.

In other words, this was a controlled datap-science test of how representation alone shapes model fairness.

What the Model Revealed

The results were interesting.

The young-biased version recognised younger faces correctly about 63 percentage points more often than older ones, illustrating a sharp gap that mirrored what we often see in real-world AI performance.

When I rebalanced the data, that gap dropped to under 5 points.

The overall accuracy stayed almost identical, meaning the model became fairer without becoming worse.

Fairer data didn’t hurt performance. It made it more equal.

That might sound obvious, but in machine learning it’s a powerful reminder that bias isn’t always about bad algorithms. Sometimes it’s about who’s visible in the first place.

Why it Matters

When AI vision systems are used in recruitment, healthcare, or security, performance gaps can turn into real-world exclusion.

If an older person’s face is more likely to be misclassified, the risk isn’t just a technical one, it’s social.

And because AI models are trained at scale, a hidden imbalance in the data can quietly become a statistical bias that repeats itself everywhere.

In short: what’s unseen becomes unrecognised.

The Ethical Takeaway

Balancing a dataset is not a glamorous fix. It doesn’t require complex algorithms or huge budgets.

But this small act of fairness by giving equal representation to older adults measurably improved the model’s reliability.

It’s proof that fairness can be engineered upstream, before the first line of code is written. And that’s the heart of age-inclusive AI: design choices that respect visibility, accuracy, and dignity for every age group.

What’s Next

This was a research edition using a non-commercial dataset. Next, I’ll be rebuilding the same experiment using FairFace, which is open for commercial use and includes richer diversity labels.

That version will become part of my broader Age-Inclusive AI Toolkit, a set of practical fairness and audit resources available for researchers and professionals.

Reflection

Age bias in AI isn’t a moral flaw, it’s an absence of presence. When data doesn’t include the full spectrum of human experience, systems learn to overlook it.

Fixing that starts with one simple step: making sure every age is seen.

If AI models learn from what they see, whose faces or stories do you think are still missing?

This report is available FREE to my subscribers from the Resource page made available on the welcome email. If you are already a subscriber and do not have the up-to-date link, or the link is not working, please email me at ageinclusive.ai@gmail.com if you are interested in this report and I can send it to you.

Serious researchers, educators, or practitioners interested in discussing the methods or interested in collaboration are welcome to contact me directly.

Beyond the Average – Age-Inclusive AI Project. I don’t fill your inbox every time a post goes live, just drop by to see what is new. Subscribe to hear about new books, research and reports only.

Subscribe now

This article builds on my earlier analysis, “Fair to Whom,” which mapped how often global AI policies explicitly name protected groups such as age, gender, disability, race, and ethnicity. This follow-up compares those results to the dominance of ethical principles - fairness, transparency, accountability, and safety - across the same texts.

Introduction to the Research

Across global AI governance frameworks, “fairness” is everywhere - but often in name only. When policymakers describe fair or trustworthy AI, they tend to speak the language of principles rather than people.

So while terms such as safety, accountability, and risk dominate the page, the very groups most affected by AI - by age, gender, disability, race, or ethnicity - barely appear at all.

What I Looked At

Building on Part I of the research ‘Fair to Whom?’, this phase compared how often the same eleven AI governance documents mention ethical principles versus protected groups.

For each text, I counted references to values such as fairness, transparency, accountability, safety, privacy, and human oversight, then compared them to the frequency of group-based words. Everything was normalised by document length to see proportional weight.

(Detailed ratios, per-document charts, and keyword dictionaries are included in the full report.)

What Emerged

The imbalance was striking across frameworks, ethical principles outnumber group references by roughly 95 to 1.

Procedural terms such as safety, risk management, and accountability occur dozens - sometimes hundreds - of times more often than any mention of age, disability, or ethnicity. Even “fairness,” when it appears, is usually framed as a system property to be managed, not a social promise to protect.

In short: AI policy has become excellent at governing machines, but not yet at governing fairness.

(Figure 1. How Often AI Policy Talks About Principles vs People — simplified summary; full charts in the report.)

Rights-based instruments give more space to protected groups, but even there, procedural language outweighs equality by a wide margin.

Why It Matters

This gap isn’t just rhetorical. When fairness is treated purely as procedure, it loses its human anchor.

Risk frameworks can ensure that systems are safe - yet remain silent about who is systematically disadvantaged.

Without explicit reference to protected groups, AI governance risks becoming ethically hollow: technically precise, socially blind.

Where to Next

Aligning fairness with equality law would mean:

• Explicitly naming all protected characteristics in audits and impact assessments.

• Balancing system-level principles (safety, accountability) with human-level protections (age, disability, race, gender).

• Funding deeper research into under-examined biases - especially around age and disability.

These steps move AI policy from abstraction to inclusion.

Using a proprietary text-mining and keyword-analysis method, this study analyses AI policy and governance frameworks to quantify the visibility of ethical principles and protected groups. The method can be adapted for custom audits or organisational benchmarking.

Update — October 2025: To support my research going forward, I’ve released the full methodology and all charts as a paid research report.

👉 You can read the complete research paper and download the charts here:
Protected Groups in AI Governance: A Text Analysis of Global Policy Documents

Closing Note

If you found this piece useful, consider subscribing or sharing it with someone interested in digital fairness and governance.

Advocating for age-inclusive AI, one story at a time.

Beyond the Average. Data. Insight. Life.

Subscribe now

---

In the UK, neutrality dominates. In the US and EU, low exposure and access play the same role - keeping older adults distant from innovation, not opposed to it.

Introduction: Challenging the Stereotype

The dominant story about older adults and technology is a simple one: they’re suspicious, resistant, even hostile. In the age of AI, this stereotype has hardened further - older workers are often portrayed as out of step with innovation.

Yet that perception tells us more about the design of technology than about older adults themselves.

A closer look at the data suggests something different. Across the UK, US, and Europe, the over-50s are not reflexively anti-AI. Instead, they are more likely to be neutral - neither positive nor negative - reflecting limited exposure, weaker digital skills, and patchier access.

Neutrality is not opposition. It’s a signal of untapped potential.

UK (ONS): Attitudes Lean Neutral, Not Negative

Data from the ONS Public Opinions and Social Trends survey reveals how attitudes to AI vary by age. When respondents were asked how much they agreed that AI will benefit them - one of the clearest indicators of general trust and optimism toward new technology:

• 50–69: around 36% are positive about AI, while many more are neutral (46%).

• 70+: positivity drops to 26% but neutrality dominates (53%).

• Negativity rises only modestly.

Source: ONS 2025

The lesson? Older Britons are not shouting “no” to AI. They’re sitting on the fence. The majority neither agree nor disagree with positive statements about AI - signalling uncertainty rather than entrenched mistrust.

According to the Ada Lovelace Institute’s How Do People Feel About AI? (2025) survey, older adults in the UK show a stronger preference for independent or public oversight of AI rather than company-led governance. Those aged 55 and over are most likely to trust regulators over companies and are less likely to have used large language models such as ChatGPT or Copilot. The report also highlights widespread concern that automation in areas like health, welfare, and policing could undermine fairness and human judgment. Taken together, these findings suggest that what appears as neutrality toward AI in the UK may stem from a digitally cautious older demographic and its expectation of public accountability.

US (Pew): Frequent Use Declines with Age

The picture is similar in the United States, though measured through behaviour rather than attitudes. Pew Research (2025) reports:

• 30% of 50–64 year olds say they interact with AI frequently.

• Among 65+, this drops to 19%.

That’s an almost identical ten point decline to the UK’s positivity gap.

Source: Pew 2025

Why does this matter? Because frequency of use is a strong proxy for positivity: few people use a tool regularly if they find it useless or threatening. Lower use among older groups helps explain their weaker positivity.

EU (Eurostat): The Skills Deficit

Across Europe, the underlying digital foundations are even clearer. Eurostat (2023) reports:

• Only about one-quarter to one-third of 65–74 year-olds in the EU have basic digital skills - 34% of men and 23% of women. Since older women outnumber men, the overall figure is likely closer to the lower end of that range.

• This means more than 70% of older Europeans lack even the basic competences needed to engage with AI.

Source: Eurostat – Digital Skills in 2023.

This sharp decline in skills (from ~70% among younger groups down to ~30% for older groups) shows that age inclusion in AI is not just about attitudes, but about capacity.

Internet Use: The Exposure Gap

Internet use data reinforces this structural gap:

• 97% of 16–29s in the EU are online daily (Eurostat 2025).

• Among 65–74s, only 61% had used the internet in the last 3 months (Eurostat 2020).

• In the UK, 86% of 65–74s were recent users in 2020, but this fell sharply to 54% among 75+ (ONS 2020).

• Middle-aged groups (30–64) fall in between, but the generational gap is most striking between the youngest and oldest cohorts.

The message is clear: older adults are less online, less often. And if you’re not online daily, you’re far less likely to bump into AI tools in the first place.

Pulling the Threads Together

Across all three regions, a pattern emerges: the more digital confidence and exposure people have, the more open they are to AI.

The real divide isn’t enthusiasm versus fear - it’s familiarity versus distance.

• UK: Older adults lean neutral about AI - not hostile, just undecided.

• US: Older adults use AI less frequently - exposure drives attitudes.

• EU: Older adults have weaker digital skills and lower internet usage - limiting both exposure and confidence.

Together, these datasets tell a coherent story:

“Older adults are not anti-AI - they are under-exposed. Their neutrality reflects low usage, which in turn reflects weak digital skills and less frequent internet use. Neutrality is a design and policy challenge, not a fixed opposition.”

While the Ada Lovelace Institute’s 2025 study offers a vital snapshot of attitudes and expectations, the analysis here takes those findings a step further. By linking attitudes, behaviour, and access data across the UK, US and EU, we can see how neutrality among older adults is shaped not just by preference or trust, but by deeper structural and experiential gaps - from uneven exposure to AI tools to the erosion of digital confidence. In other words, this analysis looks beyond what people say about AI to explore the ecosystem that makes engagement possible in the first place.

Implications for Design & Policy

So what can designers and policymakers do to bridge that distance?

Bridge the basics: Without basic digital skills, AI will remain out of reach for many over-50s. Digital inclusion programmes must be aligned with AI rollouts.

Build relevance: Older adults are more likely to engage with AI that solves problems they actually face (health, pensions, daily admin).

Design for trust: Neutrality is an opportunity. Clear explanations, safe defaults, and transparent AI systems can turn fence-sitters into supporters.

Acknowledge diversity: Within the 50+ group, gaps widen - especially by gender. Women 65–74 in the EU are the least digitally equipped. Age inclusion must also be gender-aware.

Conclusion: From Neutrality to Distance

In the UK, neutrality dominates older adults’ attitudes toward AI. Most over-50s are not rejecting AI outright - they’re uncertain. That hesitation reflects limited exposure and confidence rather than hostility.

In the United States, frequency of AI use falls sharply with age, suggesting that familiarity - not fear - drives engagement.

Across Europe, the barrier is more structural: low digital skills and patchy internet access leave many older adults outside the digital conversation altogether.

Taken together, these patterns reveal a common thread:

Where digital confidence and exposure are low, openness to AI is lower too - but not necessarily negative.

So while neutrality is the defining challenge in the UK, the broader issue across societies is digital distance - whether caused by skill gaps, access, or opportunity.

Older adults are not anti-AI. They are simply further from it.

📚 References

• Office for National Statistics (ONS). (2025) Public Opinions and Social Trends, Great Britain: Artificial Intelligence, January 2025.

• Office for National Statistics (ONS). (2020) Internet Users, UK: 2020.

• Pew Research Center. (2025) AI in Americans’ Lives: Awareness, Experiences and Attitudes.

• Eurostat. (2023) Digital Skills in the EU – 2023 Edition.

• Eurostat. (2020) How Popular is Internet Use Among Older People?

• Eurostat. (2025) The Young are Daily Internet Users – Older People Catching Up Slowly.

Ada Lovelace Institute. (2025) How Do People Feel About AI? Public attitudes to artificial intelligence in the UK. London: Ada Lovelace Institute.

I will be researching more in this area, working towards a more comprehensive report. Follow me here on Substack if you want to be notified about future articles.

If you found this piece useful, consider subscribing or sharing it with someone who cares about digital justice. I’m writing more on the ethics, data, and human impact of AI - especially where inclusion is concerned.

• Advocating for age-inclusive AI, one story at a time.

• Beyond the Average. Data. Insight. Life.

Subscribe now