Cagliari ​​Statement on Linguistic​ and Conceptual Justice in AI​

Drafted in Cagliari at the Linguistic Justice Society Satellite Workshop on “Interdisciplinary Perspectives on Linguistic Justice” and 6th European Experimental Philosophy Conference

Please fill out this form to add your signature: https://forms.gle/27BGwnW3CmLe49XE7

Cross-linguistic inconsistency in AI is the point where linguistic and conceptual justice become visible. Driven by the de facto dominance of English, AI multilingual alignment shows a form of cross-linguistic deference, a systematic tendency to resolve cross-linguistic disagreement by treating the English judgment as authoritative, rather than acknowledging the natural cross-linguistic inconsistency of human natural language.

Cross-linguistic inconsistency in AI models, therefore, should not be treated as a defect to be eliminated. Alignment that targets cross-linguistic consistency turns linguistic difference into error by presupposing an interlingual semantic common ground that, at the moment, mostly reflects English reference meanings, categories, and annotations. Correcting “inconsistencies” there can suppress legitimate semantic plurality and progressively establish English (spoken natively by only about 6% of the world’s population) as the de facto standard for what meanings should remain equivalent across languages. Rather, language models require language-specific alignment, grounded in rich data on how ordinary speakers actually use their words in each language.

Background:

While the capabilities of current frontier AI models have shown remarkable progress in areas such as mathematical ability, their semantic competence in natural languages—including, in particular, lower-resource languages—fails to accurately reflect basic human semantic intuitions (i.e., how ordinary speakers actually use their words in diverse contexts). This gap is what this statement addresses. 

Multilingual AI systems now answer questions about foreign facts, reports, opinions, etc. in local languages without explicitly indicating that translation mediates the response. Moreover, because these systems acquire meaning from patterns of usage, they inherit the semantic norms of the larger corpora they are trained on—which are predominantly English. We therefore hold that decisions about how such systems behave across languages are not merely technical, but normative.  

We begin from the following facts about language, which current AI systems often fail to respect: 

– Concepts do not align perfectly across languages. Languages carve the world in slightly different ways. The closest translation of a word does not always refer to the exact same concept. “The same but different” is not a contradiction but a general fact about concepts captured in languages, and therefore the best translation does not necessarily assure the same scope, application conditions, or implications. 

– As a result, speakers of different languages will sometimes reach divergent judgments about the same case. For example, in some contexts, they even give contradictory answers (yes/no) to a question (whether something is C, or someone has the attribute C, etc.) concerning or using divergent concepts. Given the number of concepts, contexts, and language pairs, such cross-linguistic disagreement is widespread, not exceptional.  

– Such disagreements often go unnoticed. Since multilingualism does not necessarily come with reliable native-speaker intuitions, even multilingual researchers may fail to acknowledge cross-linguistic disagreement. Cases must be discovered and established empirically. 

Against these facts, examination of current frontier models finds that they are not yet faithful to human semantic intuitions: they are often internally inconsistent about linguistic usage within a single language and respond differently to structurally identical cases. Such intra-linguistic inconsistency is a genuine technological defect. Instead, this Statement concerns cross-linguistic inconsistency (i.e., responding differently in different languages). If this type of “inconsistency” is regarded as a defect to be corrected, that is more than a technical alignment decision. There are potentially two conflicting alignment targets for aligning models here: (a) cross-linguistic consistency, which requires a system to return the same judgment across languages; and (b) folk-linguistic consistency, which requires the system to remain faithful to the ordinary usage of each language. (This choice arises only for judgments that turn on the concepts of a language, and consistency across languages is appropriate and expected in science and elsewhere.) If the aim of alignment is cross-linguistic consistency, uniformity can easily coalesce around English usage, because the huge asymmetry in the data used to train current AI models makes cross-linguistic consistency in practice lean toward English-based criteria—whether or not this is intended. 

This latter possibility worries us. Where divergence is (prima facie) appropriate yet unnoticed, and where uniformity defaults to English, the result is cross-linguistic deference: the systematic tendency to resolve cross-linguistic disagreement by treating the ​​​​​​English (or any dominant language’s) judgment as authoritative. 

In this regard, we affirm the UNESCO Recommendation on the Ethics of Artificial Intelligence (2021) and the work of the UNESCO Coalition for Linguistic Diversity in AI, which recommend addressing the exclusion of under-represented languages from AI systems through data, infrastructure, and evaluation. This Statement’s concern here is complementary as well as distinct: inclusion in a system, by expanding coverage, is not enough, and a language may be fully supported (well-resourced, fluently generated, benchmarked) but still have its concepts overruled if the system’s judgments in that language are calibrated to English semantic norms. 

Principles:

We therefore propose the following principles in multilingual alignment:

1. Cross-linguistic consistency is a choice, not a specification. The relevant question is not whether a system displays consistency, but which consistency it displays and by whose semantic norms. 

2. Divergence is not an error by default. A judgment that diverges from the English one because of the semantic norms of another language must not be treated as a defect to be corrected.  

3. Equal semantic standing. Semantic norms of no particular language carry inherent authority over those of any other. The de facto dominance of English must not become semantic and conceptual privilege. 

4. Legitimate authority. Those people whose concepts are at stake—linguistic communities, rather than only developers—have a claim to a voice in how such divergences are resolved.  

5. Transparency where it matters. Where a judgment is known to be language-dependent, and the stakes are not obviously trivial, users should be informed of that dependence. 

Calls to the Parties Concerned:

We therefore issue the following calls to concerned parties:  

• To AI developers and providers (AI companies, public entities, etc.): treat cross-linguistic alignment policies, especially whether they aim for cross-linguistic consistency (the same judgment across languages) or folk-linguistic consistency (truthfulness to the ordinary usage of each language), as a published, reviewable normative commitment; do not suppress cross-linguistic divergence as mere contradiction; document which consistency norm a model implements. 

• To researchers in AI, linguistics, and the empirical study of concepts (including experimental philosophy and cognitive science): build open, empirical, multilingual evidence on where conceptual divergence actually occurs, and, when evaluating models, report not only cross-linguistic agreement but also within-language consistency. 

• To institutions and funders: support not only work that extends language coverage but also empirical work that documents the semantic norms and usage of languages other than English on their own terms, and, where AI outputs are used as evidence for policy (for example, where a model is used to classify, summarize, or judge multilingual responses, submissions, or cases) require, as a condition of funding or procurement, that the conceptual standard behind them be disclosed. 

• To standard-setting bodies and benchmark developers: do not treat agreement across languages as a measure of quality. A multilingual benchmark that positively evaluates a model if it returns the same answer in every language encodes cross-linguistic deference as a metric, and rewards the removal of legitimate divergence. Report within-language consistency and cross-linguistic agreement as separate quantities without aggregating them into a single score, and set only within-language consistency, not cross-linguistic consistency, as the target rewarded by a benchmark. 

• To all parties: recognize cross-linguistic inconsistency in AI not as a bug to be quietly removed, but as the point at which linguistic and conceptual justice become visible.

Signed, September 2026

1. Masaharu Mizumoto (Japan Advanced Institute of Science and Technology), co-creator
2. Filippo Contesi (Cagliari), co-creator
3. Çağla Çimendereli (Old Dominion, Virginia), co-creator
4. Daniela Alderuccio (ENEA Italy), co-author
5. Maya Alkateb-Chami (Harvard), co-author
6. Rossana Cotroneo (ENEA Italy), co-author
7. Denisa Kera (Bar-Ilan), co-author
8. Giuseppina Calenda (Mediterranea Reggio Calabria)
9. Diana Camps (Glasgow)
10. Angelo Delliponti (Adam Mickiewicz)
11. Vilius Dranseika (Jagiellonian)
12. Eugen Fischer (East Anglia)
13. Jenny Jones (Edinburgh)
14. Markus Kneer (Graz)
15. Katarzyna Kuś (Warsaw)
16. Edouard Machery (Pittsburgh)
17. Sergi Morales-Gálvez (Barcelona)
18. Kevin Reuter (Gothenburg)
19. Maciej Tarnowski (Jagiellonian)
20. Bengt-Arne Wickström (Andrássy Budapest)
21. Katarzyna Zyń (Warsaw)

Additional Signatories:

19. Luca Marchetti (Genoa)
20. Simon Evnine (Miami)
21. Stephen May (Auckland)
22. Peter Finocchiaro (Wuhan)
23. Kyria Rebeca Finardi (UFES, Espírito Santo)
24. Jennifer Jenkins (Southampton)
25. Diana I. Pérez (Buenos Aires)
26. Shawna Shapiro (Middlebury College, Vermont)
27. Diego Lawler (CONICET, Argentina)
28. Karina Pedace (Buenos Aires)
29. John Harbord (Maastricht)
30. Nancy H. Hornberger (Penn)
31. Rosemary Salomone (St John’s School of Law)
32. Jennifer S. Wilson (Rice)
33. Andrea Wehrli (Bern University of Applied Sciences)
34. Yongyan Zheng (Fudan)
35. Stephan Meyer (Basel)
36. Claudia Finkbeiner (Kassel)
37. Hanada Taha (Zayed)
38. Elisa Paganini (Milan)
39. Christine Hélot (Strasbourg)
40. Aisling O’Boyle (Queen’s Belfast)
41. Gereon Wolters (Konstanz)
42. Alfredo Tomasetta (IUSS Pavia)
43. Jennie Snow (Loyola Marymount)
44. Meike Wernicke (British Columbia)
45. Silvana Pani (Milan)
46. Carla Bagna (University for Foreigners of Siena)
47. Maria Regina Mascia (Cagliari)
48. Josep Soler Carbonell (Stockholm)
49. Paloma Puljic (Graz)
50. Chiara Facciani (University for Foreigners of Siena)
51. Mireia Trenchs Parera (Pompeu Fabra)
52. Adriana Cristina Sambugaro de Mattos Brahim (UFPR Paraná)
53. Argyro Kanaki (Dundee)
54. Ljubisa Bojic (Institute for Artificial Intelligence Research and Development of Serbia)
55. Nestor Sebastian Salamanca Garcia
56. Gabriel González (Texas Rio Grande Valley)
57. Amidor Boaz (Bar-Ilan)
58. Zachy Hennessey
59. Brian Rosenblum (Kansas)
60. Jay Jordan (Utah)
61. Jon Giullian (Kansas)
62. Jenna Elizabeth Pollard (Kansas)
63. Bartholomew Dean (Kansas)
64. Susanne Treiber (Madison College)
65. Simbiat Ajao (Business School of AI)
66. Michael Beaney (Aberdeen/Peking)
67. Alessandro Corona Mendozza (Aarhus)
68. Irene Olivero
69. Mahendra K. Verma (York)
70. Brent Metz (Kansas)
71. Enrique Encinas Dolejsová (Oslo School of Architecture and Design)
72. Robert M. Maier (VdÜ/IGDORE)
73. Maria Verdaguer (Institut d’Estudis Catalans)
74. Thomas J. Spiegel (Miyazaki International)
75. Alejandro Jiménez Gijón (UNED Spain)
76. Joonas Pennanen (Jyväskylä)
77. Alexander Cohen (Rensselaer Polytechnic Institute)
78. Sandeep Mertia (Stevens Institute of Technology)
79. J.F. Blanc (Institut d’Estudis Occitans)
80. Daniel Devatman Hromada (Berlin University of the Arts)
81. Christoph Lumer (Siena)
82. Josie Valeri (Point Park)
83. James Andrew Stroble (Hawaii)
84. Dean Keith Simonton (UC Davis)
85. Ellen Head (Miyazaki International)
86. Sahara Al-Madi (Linguistic Security Institute)
87. Luke Blanchford
88. Lior Zalmanson (Tel Aviv)
89. Sarah E. Cornwell (York)
90. Marcovalerio Di Schiena
91. Giuseppina Sgueglia
92. Jad Kiadan (Forum Transregionale Studien)
93. Mikel Burley (Leeds)
94. María Luisa Carrió-Pastor (UPV Valencia)
95. Manuel García-Carpintero (Barcelona)
96. Andrew Kirton (Leeds)
97. Analeigh E. Horton (Fairleigh Dickinson)