Is generative AI really changing public sector work? Evidence from a multi-country survey
summary
The gist
Artificial intelligence has the potential to transform how public services are delivered, with important implications for public service productivity Our survey of 3,362 public sector workers across…
In short
A survey of over 3,300 public sector workers across multiple countries reveals widespread, bottom-up adoption of generative AI. While users report significant productivity gains (around 80%), this benefit is often offset by rising work expectations and concerns about security and bias. The findings suggest that without proper guidance reaching frontline staff, the technology's transformative potential remains limited.
Key concepts
- Bottom-up Adoption
- This refers to how workers adopt AI tools individually rather than having institutions mandate their use. For example, in the EU, many individuals use AI tools even if their organization has not formally adopted them. This decentralized approach can speed up initial use but relies heavily on individual willingness.
- Productivity vs. Work Scope
- Workers generally feel more productive using AI (around 80% report this). However, a significant portion of workers also feel the scope of their job has increased because they are now expected to do more with the time saved by AI. This creates a tension where efficiency gains are absorbed by higher demands.
- Implicit Bias
- This is the risk that AI systems might unfairly favor or disadvantage certain groups due to biases present in the data used to train them. The survey highlighted this as a serious concern across both users and non-users, suggesting that unchecked bias could negatively impact public service delivery.
Terminology used across episodes
This episode discusses
- Is generative AI really changing public sector work? Evidence from a multi-country survey · Paper Radio
- The Impact of AI on Developer Productivity: Evidence from GitHub Copilot
The paper
Is generative AI really changing public sector work? Evidence from a multi-country survey · Read on arXiv
Public Policy Programme, The Alan Turing Institute, British Library, UK · Institute for Global Prosperity, University College London, UK · pattrn.ai, UK · Centre for Advanced Spatial Analysis, University College London
Artificial intelligence has the potential to transform how public services are delivered, with important implications for public service productivity. But how widely is the technology actually being used - and is it making any difference? Our survey of 3,362 public sector workers across the UK, the USA, South Africa and the EU finds widespread use of AI in every region. Our findings suggest that this spread is being driven by bottom-up individual adoption rather than top-down institutional adoption. In the EU, for example, 65% of respondents use AI, while only 19% say their organisation has formally adopted it. Most users are positive about its effects, with 79% saying AI has improved their productivity. However, only 29% say it has increased what is expected of them at work, raising questions about how truly transformative the technology is so far. Across regions, South Africa is a clear outlier, with the highest levels of use, organisational adoption, optimism and awareness of guidelines.
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Is generative AI really changing public sector work? Evidence from a multi-country survey".
Jane: Here's the gist of the paper. Artificial intelligence has the potential to transform how public services are delivered, with important implications for public service productivity Our survey of 3,362 public sector workers across the UK, the USA, South Africa and the EU finds widespread use of AI in every region Our findings suggest that this spread is being driven by bottom-up individual adoption rather than top-down institutional adoption Most users are positive about its effects, with 79% saying AI has improved their productivity However, only 29% say it has increased what is expected of them at work, raising questions about how truly transformative the technology is so far Adoption Dynamics and Spread The widespread use of AI in the public sector is driven by bottom-up individual adoption rather than top-down institutional adoption For example, in the EU, 65% of respondents use AI, while only 19% say their organisation has formally adopted it This decentralized mode of adoption may lower barriers to use and accelerate diffusion The success of such bottom-up adoption likely depends on the attitudes of those expected to use these technologies in their day-to-day work Impact on Productivity and Work Scope Workers report that around 80% of users say it makes them more productive at work However, around 30% mentioned that generative AI has increased how much is expected from them at work South African users are generally the most positive about generative AI and its effects on their work Yet they are also by far the most likely to report rising expectations More than half of South African users say generative AI has changed the scope of their day-to-day work This points to a broader tension where time saved is absorbed by higher demands rather than returned to workers Concerns Regarding Governance and Bias More than 70% of respondents, both among AI users and non-AI users, showed concerns around security and privacy of AI systems More than half of respondents stated that there is not enough guidance available on how to use these systems In addition, more than half of respondents stated that any implicit bias in the AI systems could be a serious concern in their field of work There are significant differences in opinions between AI users vs non-AI users, with 75% of AI users being optimistic about use of AI in their field (compared to below 40% for non-AI users) Guideline Awareness and Regional Variation There is growing discussion around the need for guidelines governing AI use to ensure such tools are applied safely and without bias Awareness of specific AI guidelines varies significantly across regions, with South Africa reporting the highest levels of awareness overall For instance, nearly 50% of South Africa respondents were aware of the SA National Artificial Intelligence Policy Framework, but awareness across other regions was significantly low Supervisors tend to be more aware of guidelines compared to non-supervisors Implications for Policy and Practice The findings suggest that governments and employers should disseminate guidance through supervisors, embed it in induction and training, and attach it to the tools workers actually use Targeted training and support in areas where risk is greatest would address this Furthermore, measures of output alone will overstate the benefits if productivity gains are being absorbed by rising expectations The low rate of organisational adoption compounds the problem where workers use consumer tools under consumer terms The paper concludes that publishing frameworks is not enough if they do not reach the frontline workers who are already using these tools The dominance of generative AI through individual adoption suggests that future technology provision may rely on large technology companies rather than institutional models References
Bick et al., 2024: Bick, A., Blandin, A., and Deming, D.
Tom: J.
Jane: (2024).
Tom: The rapid adoption of generative AI
Bright et al., 2025: Bright, J., Enock, F., Esnaashari, S., Francis, J., Hashem, Y., and Morgan, D.
Paper summary: Tom: So to recap, this paper "Is generative AI really changing public sector work? Evidence from a multi-country survey" is looking at a big set of workers across the UK, USA, South Africa, and the EU to see what's actually happening with AI in their jobs.
Jane: The core idea they are exploring is whether this technology is truly changing things or just causing some superficial changes. They found that while there's widespread use of AI everywhere these regions, it’s largely driven by individuals picking it up on their own rather than big organizations mandating it all at once.
Lu: The paper claims that this decentralized way of adoption means tools can be deployed directly by individual workers to experiment with and automate parts of their own workflows, which is different from the large scale, top-down programs we've seen in earlier public sector technology introductions
Bright et al., two thousand twenty-five: <ref:2611.16102#pg2>.
Meng: They show that this bottom-up mode of adoption might lower the barriers to entry for using these tools and allow productivity gains to emerge organically from how people actually work on the ground.
Lalam: The survey itself included three thousand three hundred sixty-two public sector workers across those four countries, which is a pretty substantial sample size for getting these kinds of insights <ref:2611.16102#pg1,3,362 public sector workers across>.
Tom: And looking at the results, they found that a significant portion of users are positive; specifically seventy-nine percent of respondents said AI has improved their productivity at work <ref:2611.16102#pg1,AI has improved their productivity>.
Jane: But the crucial part they pull out is that only twenty-nine percent say AI has increased what is expected of them at work, which points to a major area for concern regarding the real impact of this technology <ref:2611.16102#pg1,has increased what is expected of>.
Lu: That tension between time saved and higher demands seems to be a central point; it suggests that the time saved by using AI isn't necessarily being returned to the workers.
Meng: If senior management keeps piling more work onto people, then measuring just how much faster they can do things doesn't tell the whole story about their workload.
Lalam: The paper also delves into specific concerns around security and privacy, noting that over seventy percent of respondents, both users and non-users, are worried about these aspects of AI systems <ref:2611.16102#pg1>.
Tom: And on top of that, they found that more than half the people surveyed felt that any implicit bias in the AI systems could be a serious problem in their specific line of work.
Jane: They also looked at guideline awareness, and they found it varies by region; South Africa reported the highest level of awareness overall for those specific AI guidelines mentioned in the study.
Lu: And they pointed out that supervisors generally seem to be more aware of these guidelines than non-supervisors, which is a structural observation about who knows what.
Meng: So, if you're looking at deployment from an engineering standpoint, it sounds like the tools are being adopted quickly by the users first before any formal rollout happens.
Tom: And they conclude that because of this individual adoption pattern, the success of these AI tools really depends on those attitudes of the people who are actually expected to use them day-to-day.
Jane: It’s a complex picture where productivity gains are happening, but only if the workers themselves are okay with what that means for their job scope.
Conclusion: Tom: So wrapping up this paper "Is generative AI really changing public sector work? Evidence from a multi-country survey," the main thing is that widespread AI use is happening bottom-up, driven by individuals testing tools on their own.
Jane: And the paper isn't making a strong claim about whether it's fundamentally changed the job structure yet; it shows productivity gains are happening, but only if those gains aren't being absorbed by increased job demands.
Lu: The authors are emphasizing that just creating guidelines isn't enough; they have to get that guidance out to the frontline workers who are using these tools every day.
Meng: And they suggest that output alone doesn't tell the whole story if those productivity gains are actually being absorbed by rising expectations from management.
Lalam: The future direction suggested is that because individual adoption is so dominant, we might see a reliance on large technology companies providing the tools rather than purely institutional models for deployment.
Tom: That means the focus needs to be on targeted training and support in areas where risk is greatest, not just blanket policies.
Jane: It’s about making sure that when people are using these AI tools, they understand the risks and what's expected of them so it actually works for them.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates