Has Artificial Intelligence Depressed Graduate Wages in the United Kingdom?

Generative artificial intelligence has rapidly spread across graduate-heavy professions, representing one of the most significant technological shocks to the modern labour market.1 Large language models (LLMs) such as OpenAI’s ChatGPT and Microsoft Copilot are now directly integrated into core operations in finance, law, consulting and software development – occupations traditionally characterised by a high proportion of human capital. Previous automation waves mainly displaced routine manual labour, but the introduction of AI targets cognitive and analytical tasks that were once considered uniquely able to be completed by university graduates.

 

As a result, the economic implications are ambiguous. Human capital theory suggests that if the marginal productivity of skilled workers improves with the use of AI, graduate wages should increase. However, task-based models of automation suggest that if routine analytical work is offloaded to AI, especially in junior positions, wage growth will stagnate or even decline.2 Both these opposing forces can be seen by anecdotal evidence: AI-assisted ‘vibecoding’ has lowered the barriers to entry for startup creation, yet firms simultaneously report reduced demand for entry-level roles.6/7

 

The question that arises from this is whether AI exposure is associated with higher real wage growth for graduates.

 

Coupled with an AI exposure index, occupational wage data from the Office for National Statistics (ONS) can be used to empirically assess whether a complementary or substitutive effect dominates in practice.4

 

If AI primarily complements graduate labour, occupations with higher AI exposure should experience faster real wage growth. This means that AI sufficiently enhances productivity to increase the returns to human capital, so there would be a positive association between AI exposure and real wage changes.

 

Conversely, if AI substitutes the routine cognitive tasks within graduate roles, wage growth may weaken in higher exposure fields. Entry-level analytical tasks, such as coding and data processing, are particularly vulnerable. Higher AI-exposed occupations would therefore exhibit slower real wage growth relative to less exposed ones.

 

There is a third possibility: there is no measurable effect on real wages in the short run. There may be a time lag in the diffusion of the technology, with firms gradually adjusting. The rigid nature of employment contracts may also mean delays in observable wage changes.

 

Quantitative analysis is required to distinguish between these outcomes. Occupational data was drawn from the ONS Annual Survey of Hours and Earnings (ASHE).4 This provides information on gross weekly earnings across occupations in the UK classified by Standard Occupational Classification (SOC) codes.

 

The sample covers 2018-2024. This captures data before and after the widespread diffusion of generative AI tools post 2022. Nominal wages have been converted into real terms using CPI inflation data to ensure that recorded changes reflect actual purchasing power.5

 

AI exposure has been measured using an occupational exposure index constructed from task-level assessments of susceptibility to LLM automation.3 Different occupations were assigned a continuous ‘exposure score’. This is based on the proportion of tasks that overlap with capabilities seen by AI. Higher values correspond to greater theoretical exposure to automation. This enables comparison between exposure levels.

 

The dependent variable, constructed from the ASHE weekly earnings data4, is real wage growth by occupation. Nominal wages are adjusted for inflation and then converted into logarithmic changes to measure proportional change. This means that extreme values have less of an influence. This also allows coefficients to be interpreted in approximate percentage terms rather than numeric values, which is not useful for change analysis.

​

​​

​

​

The primary independent variable is an occupational AI exposure index, mapped to SOC occupations. As exposure is defined by task composition, it is time invariant at the occupation level (AIExposurei). To facilitate descriptive analysis, occupations are grouped into exposure groups (e.g. low, medium, high) to compare average wage patterns across groups, both over time and between pre and post LLM periods.

The following equation can be written:

​

 

​

The constant term ‘α’ represents the average wage change across all occupations in the base period, adjusting for included variables.

 

‘AIExposurei’ is a time-invariant index that measures the degree to which occupation ‘i' is exposed to generative AI based on its task composition. The higher the value, the higher the proportion of tasks that overlap with the capabilities shown by LLMs. Since exposure is based on occupational composition instead of year-specific events, it will be constant over time in the regression.

 

‘Postt’ is a binary variable, which equals 1 for the post-LLM period, i.e., 2023-2024, and 0 otherwise.

 

The cross-term ‘AIExposurei * Postt’ measures the effect of having more AI-exposed occupations experiencing differential wage growth in the post-LLM period relative to less exposed ones.

 

The coefficient ‘β’ therefore measures the differential wage growth in the post-LLM period for highly exposed occupations relative to less exposed ones.
 

·       β>0: Complementarity – The effect of AI exposure on wage growth is positive, indicating that AI-exposed occupations experienced faster wage growth in the post-LLM period relative to less exposed ones.

·       β<0: Substitution – The effect of AI exposure on wage growth is negative, indicating that AI-exposed occupations experienced slower wage growth in the post-LLM period relative to less exposed ones.

·       β≈0: No differential effect – The effect of AI exposure on wage growth is zero, indicating no differential effect on wage growth in the post-LLM period relative to less exposed ones.

 

The control variables, in vector Xit, include occupational characteristics that can independently affect wage growth, e.g. education levels, sectoral composition and gender composition, among others. 

 

The year fixed effects variable (λt) captures economy-wide factors affecting all occupations equally each year. These include inflationary pressures, macro-economic fluctuations, and general labour market conditions, among others.

 

The error term (εit) represents the unexplained wage growth not captured by the model.

 

Running the regression with the data, the following results, which are summarised in the figure below, are obtained.

 

OLS Regression Results

===================================================================

Dep. Variable:          dln_real_wage   R-squared:                       0.474

Model:                            OLS   Adj. R-squared:                  0.254

Method:                 Least Squares   F-statistic:                     50.40

Date:                Thu, 26 Feb 2026   Prob (F-statistic):           7.79e-06

Time:                        13:50:11   Log-Likelihood:                 94.647

No. Observations:                  45   AIC:                            -161.3

Df Residuals:                      31   BIC:                            -136.0

Df Model:                          13

Covariance Type:              cluster

===================================================================

coef    std err          z      P>|z|      [0.025      0.975]

-----------------------------------------------------------------------------------

Intercept           0.0071      0.003      2.482      0.013       0.002       0.013

C(year)[T.2020]    -0.0544      0.016     -3.500      0.000      -0.085      -0.024

C(year)[T.2021]     0.0167      0.022      0.775      0.438      -0.026       0.059

C(year)[T.2022]    -0.0377      0.009     -4.367      0.000      -0.055      -0.021

C(year)[T.2023]    -0.0161      0.012     -1.368      0.171      -0.039       0.007

C(soc)[T.2]         0.0020      0.000      7.010      0.000       0.001       0.003

C(soc)[T.3]        -0.0147      0.001    -24.012      0.000      -0.016      -0.014

C(soc)[T.4]         0.0166      0.001     18.485      0.000       0.015       0.018

C(soc)[T.5]         0.0059      0.001      4.980      0.000       0.004       0.008

C(soc)[T.6]         0.0303      0.001     20.771      0.000       0.027       0.033

C(soc)[T.7]         0.0097      0.002      5.397      0.000       0.006       0.013

C(soc)[T.8]         0.0131      0.002      6.297      0.000       0.009       0.017

C(soc)[T.9]         0.0101      0.002      4.207      0.000       0.005       0.015

exposure_x_post     0.0005      0.024      0.020      0.984      -0.046       0.047

===================================================================

Omnibus:                       10.165   Durbin-Watson:                   2.889

Prob(Omnibus):                  0.006   Jarque-Bera (JB):               16.652

Skew:                           0.534   Prob(JB):                     0.000242

Kurtosis:                       5.782   Cond. No.                         19.5

===================================================================

​

Notes:

[1] Standard Errors are robust to cluster correlation (cluster)

Figure 1. Regression data from the ASHE

 

From Figure 1, the point estimate for the Beta is 0.0005 with a standard error of 0.024 and a p-value of 0.984. This estimate is extremely close to zero, rendering it statistically insignificant. The constructed 95% confidence interval ranges from -0.046 to 0.047, including economically meaningful negative values. The central estimate, nonetheless, implies no measurable relationship. A 10% increase in the AI exposure index is associated with a 0.005% increase in wages.

 

The overall model fit is moderate (R² = 0.474), which means that 47% of the variation in occupation-level wage growth is explained by occupation and year fixed effects. The year fixed effects are reasonable since wage growth declined sharply in 2020 (coefficient -0.0544, p < 0.001), which corresponds to the pandemic shock, and again in 2022 (coefficient -0.0377, p < 0.001), which corresponds to the inflationary squeeze.

The 2023 dummy has a negative but statistically insignificant coefficient (-0.0161, p = 0.171). The occupation fixed effects are statistically significant, which means that there are occupation-level differences in the structural wage growth patterns. But, conditional on these controls, AI exposure does not explain differential post-period wage growth. 

 

Regardless, there are several limitations that affect the results. Firstly, the exposure index shows the theoretical overlap in tasks rather than the actual adoption of AI.8 Therefore, there could be a wide range of actual exposure in occupations that the model classifies as being highly exposed to AI. Secondly, the post-treatment period is relatively short. Wage changes tend to happen annually with some lag. Moreover, generative AI was only adopted in the economy in late 2022. Therefore, the model can only reflect the very beginning of the wage response. The model also estimates the effect at the occupation level. This may mean if the effect of generative AI is mainly on entry-level jobs while wages in higher positions tend to stay the same, the overall wage response might not change, while the internal structure of wages changes.

 

Generative AI has entered graduate-heavy professions at an unprecedented rate, even in areas thought to be safe from automation. Though the short-term effects on wages are statistically indistinguishable from zero at the occupation level, technological progress is rarely visible in the immediate macroeconomic earnings figures.

If productivity is indeed driven by AI in the long run, the returns to complementary skills such as judgment, strategy, and coordination may increase. On the other hand, if analytical work is increasingly automated, potential pressure may arise first through hiring patterns and internal wage differentials before it is visible across a much larger horizon.9

​

References

1. Department for Science, Innovation and Technology. Assessment of AI Capabilities and the Impact on the UK Labour Market. London: UK Government, 2023.
   https://www.gov.uk/government/publications/assessment-of-ai-capabilities-and-the-impact-on-the-uk-labour-market/assessment-of-ai-capabilities-and-the-impact-on-the-uk-labour-market

2. House of Commons Parliamentary Office of Science and Technology. Artificial Intelligence and Employment. POSTbrief 757. London: UK Parliament, 2024.
   https://post.parliament.uk/research-briefings/post-pn-0757/

3. Felten, Edward, Manav Raj, and Robert Seamans. “Occupational, Industry, and Geographic Exposure to Artificial Intelligence: A Novel Dataset and Its Potential Uses.” 2021.
   https://arxiv.org/abs/2104.10723

4. Office for National Statistics. Annual Survey of Hours and Earnings (ASHE). London: ONS.
   https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/earningsandworkinghours

5. Office for National Statistics. Consumer Price Inflation Tables. London: ONS.
   https://www.ons.gov.uk/economy/inflationandpriceindices

6. Institute for Public Policy Research (IPPR). Up to 8 Million UK Jobs at Risk from AI, Finds IPPR. London: IPPR, 2024.
   https://www.ippr.org/media-office/up-to-8-million-uk-jobs-at-risk-from-ai-unless-government-acts-finds-ippr

7. McKinsey & Company. “AI’s Uneven Effects on UK Jobs and Talent.” 2023.
   https://www.mckinsey.com/uk/our-insights/the-mckinsey-uk-blog/ai-uneven-effects-on-uk-jobs-and-talent

8. Institute for the Future of Work. AI and Employee Pay in the UK: New Evidence and Implications. London: IFOW, 2023.
   https://www.ifow.org/publications/policy-briefing---ai-and-employee-pay-in-the-uk-new-evidence-and-implications

9. National Foundation for Educational Research. Up to Three Million UK Jobs at Risk Over the Next Decade. Slough: NFER, 2023.
   https://www.nfer.ac.uk/press-releases/up-to-three-million-uk-jobs-at-risk-over-the-next-decade-says-report/