SNB speeches under AI scrutiny: shorter is not better policy
A new working paper examines 1,063 German-language speeches from 1928 to 2024. The comparison with classic computational methods also shows why AI results need to be checked.
The Swiss National Bank talks about monetary policy differently today than it did decades ago. Its speeches have become shorter over the long term, more accessible in their language and more focused on inflation. That is the conclusion reached by Severin Joho and Enzo Rossi in an SNB working paper published on 23 October 2026. To analyse 1,063 German-language speeches given between 1928 and 2024, they compared classic computer-assisted language analysis with ChatGPT.
The paper describes a historical shift, not the current monetary policy situation. It neither predicts the next policy rate decision nor proves that simpler speeches steer inflation better. As a working paper it also reflects the authors' views, not necessarily those of the National Bank. It is precisely these limits that make the comparison useful for public debate.
What becomes visible across almost a hundred years
From the mid-1950s, the number of speeches per year initially rose. After the move to flexible exchange rates, the trend accelerated. But the historical movement is not an uninterrupted increase up to the present: after the global financial crisis, the frequency of speeches fell again. Changes in the communication format also affect how many individual documents are counted.
The study therefore distinguishes, among other things, between addresses to general meetings, introductory remarks on monetary policy assessments and other speeches. That is more than an archival question. A short joint statement can produce fewer documents than three separate presentations, without the public receiving correspondingly less information. The sheer number is not a measure of transparency or quality.
The collection examined contains only German texts. The authors point to additional French, English and Italian speeches that are not part of the main analysis. The material is therefore broad, but not every statement by the SNB since its founding is captured. A language statistic must disclose its selection as openly as its calculation method.
Why the computer comparison matters
The researchers first cleaned the documents of headers, page numbers and other layout elements. They then calculated word counts and readability measures with Python and had corresponding analyses carried out by ChatGPT. The two approaches produced similar temporal patterns for several indicators. They did not, however, deliver identical results.
This is particularly clear with the Vienna text formula, a measure of linguistic difficulty. In the published comparison table, the correlation between the two methods here is 0.728, compared with 0.996 for the German-language Amstad index. A high correlation means similar movements, not necessarily identical measurements. Word counts also differed.
The authors also document concrete errors: ChatGPT initially applied a constant in a formula incorrectly; batch processing produced outliers. But the classic analysis was not error-free from the start either. Older PDF formats produced implausible values until the files were converted. The methodological conclusion is therefore not human versus machine, but controlled data preparation and verifiable calculation.
Readable is not automatically understood
Readability indices take into account, for example, sentence length or word features. They can thus make structural changes visible. They do not directly test whether someone understands the reasoning behind a rate decision. The SNB paper explicitly distinguishes these questions. A speech can become shorter and still presuppose economic prior knowledge.
An independent study of the Australian central bank from 2021 underlines the importance of different target groups. Joan Huang and John Simon found that perceived readability and comprehensible reasoning are only weakly linked. Simpler writing is not necessarily more understandable. A specialist and a household without economic training can judge the same text differently.
An archival finding is not a rate signal
For Swiss households, companies and media, this makes a practical difference. Frequent inflation terms show what is being talked about. They do not on their own prove how large a future decision will be or what effect a speech has on expectations. That would require additional research that actually captures the understanding and behaviour of the addressees. For assessing an individual speech, the occasion, the target audience and the full line of argument therefore remain important; a single frequently used word cannot replace that context.
The new study makes the development of SNB language more verifiable. Its strongest contribution is precisely not that it outsources monetary policy to ChatGPT. It shows how a large body of text can be opened up – and why comprehensible reasoning remains more important for the Swiss public than a good readability score.
Sources
- Joho and Rossi: SNB Working Paper 13/2026, publication and original study
- Reserve Bank of Australia: Central Bank Communication – One Size Does Not Fit All, 2021
- Lead image: Hpschaefer / Wikimedia Commons
- Licence of the lead image: Creative Commons BY 3.0
- Article image: Myriam Thyes / Wikimedia Commons
- Licence of the article image: Creative Commons BY-SA 3.0
