AI in Conflict Settings

Giuseppe Samo

IDIAP Research Institute

Beijing BLCU

​

​

EMDM, CRIMEDIM, Turin,

28.05.2026

​

​

Who am I

Research Associate @ Idiap, Martigny�Distinguished Professor @ BLCU, Beijing�Collaborator @ University of Geneva

My research focuses on AI interpretability in multilingual contexts, as well as its applied dimensions in geolinguistic data and emergency settings.

​

I have published, inter alia, in Nature Communications, Findings of EMNLP, Spatial Cognition and Computation, and Studies in Health Technology and Informatics.

​

Selected References

Samo, G., & Parker, R. (2025). Exploring AI-Driven Decisions in Disaster Simulation for Emergency Medical Teams. Studies in health technology and informatics, 328, 151–152. https://doi.org/10.3233/SHTI250691

​

Ursini, FA., Samo, G. (2025) Extracting toponyms from OpenStreetMap and other gazetteers: comparing representational accuracy in multilingual contexts. Humanit Soc Sci Commun 12, 798 (2025). https://doi.org/10.1057/s41599-025-05025-1

2

Roadmap

1. Introduction

2. Core concepts: AI & LLMs

On synthetic data: the output of the conversationalAI based on LLMs

3. Interactive Exercise 1

​

4. Operational applications

AI with natural and structured data

​

5. Risks, ethics & decision-making

​

6. Interactive Exercise 2

​

7. Wrap-up & key messages

3

Using “LLMs” in deep field conditions

Task 1: Fill a form in English

Task 2: Translation of some messages from English to Mapese (the language spoken in Mapi region)

4

Fill a form in English

5

Template

Notes

Consumer-facing

conversational AI

based on a LLM

User

Final result!

Translate some notes from English

English → Mapese

Ʃøň ǥlørbɘn qʋåƶȝıx ħʉɳɖʀêþ fŷrƈȝåx. Ɫɨqʋɵƥ ʈȿɨɳʞɐɹ mɵɳɵƌɵƞ ʑɨƞƈʈɵƥ ʝɵƞɢɬɵƞ. Æɸʉ ɞɵƥŧɵƞ ʈɵẍɵƥ ɖɨƞɢɳɵƥ, ɞɵƥ ħłɵƥɵƞ ʈɵƥŧɵŧ ʈɨƞɢɬɨƞ ǥłørbɘn ʄøƞȝɵƞ.

​

Ⱬɨƞʈɵƥ ɖɨƞɢɳɵƥ qʋåƶȝɨƞ ŋɵɱɵƞɵƞ ɵƥŧɵƞ ɞɵƥ fŷrƈȝɐx. Ɏɵɱɵɵƞ ʄłɵƥ ʈɵƥŧɵƞ ʈɨƞǥłɵƞ ɵƥ ɭɨƞǥɵƞ ʄɵɱɵƞɵƞ. Ṩɵɱɵɵƞ ɖɨƞɢɳɵƞ ʝɵƞɢɬɵƞ ɭɵƥɵƞ ʄɵɱɵƞ, ɵƥ ɞɵƥ ħłɵƥɵƞ ǥłɵƥɵƞ ȝɵŧɵƞ ʄɵɱɵɵƞ.

​

Ʃøň ǥlørbɘn ʄŷrƈȝåx ɞɵƥ ʄɵɱɵƞɵƞ ɢłɵƥɵƞ ʈȿɨƞʞɐɹ. Ɏɵɱɵɵƞ ʄłɵƥ ɵƥŧɵƞ ɨƞ ɵƞɢɬɵƞ ʄɵɱɵɞɵŧ. Ṩɵɱɵɵƞ ɖɨƞɢɳɵƞ ɵƞɢɬɵƞ ʄɵɱɵɞɵŧ, ɵƥ ʄɵɱɵɵƞ ʄłɵƥ ȝɵƞɵƞ ʄɵɱɵɞɵŧ ɭɵƥɵƞ.

​

6

Large Language Model

7

Large Language Model

8

Model

In its most common usage, a model is an abstract (and reduced) representation of an item or a concept.

9

Model

A typical technique is to represent the object of study as a collection of relevant features and their values. We can represent an object of study as a pair (x, y), where

x = (a₁, a₂, ..., a_n) and y = a_n+1

10

Hand-defined attributes

​

​

x = (a₁, a₂, ..., a_n) and y = a_n+1

​

11

apple

X = fruit, books, colours

X = colours, shapes, size

hand-defined

Inductively defined attributes

​

x = (a₁, a₂, ..., a_n) and y = a_n+1

​

12

inductively defined

An example: human vs. machine

13

An example: human vs. machine

14

Linguistic Data

Linguistic data are more complex than standard “numerical” data.

​

There is no 1-to-1 mapping between the graphemes and what they represent, across languages and alphabets/logograms.

​

E.g. English cook (verb or noun), pesca in Italian (fish or peach), etc.

15

Large Language Model

16

Language model

Language modeling is the task of predicting what word comes next in a text.

​

​

17

Flamingoes are ________

beautiful

pink

birds

apple

Language models (everyday, in the recent past)

18

Language model

19

Flamingoes are _______

beautiful

pink

birds

apple

x¹ x² y

P (y | x¹, x², …, xⁿ)

where y can be a word in a vocabulary V = {w₁ , w₂ , …, w_|V| }

A "stochastic parrot" (Bender et al. 2021)

A language model will chose the most “probable” word. The probability is calculated on observations, so we need a corpus (a database, cf. the training database). → e.g., weather forecast

20

📖 Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021, March). On the dangers of stochastic parrots: Can language models be too big?🦜. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency (pp. 610-623).

Language models

The probability vary according to the training data.

21

Flamingoes are ________

beautiful

pink

birds

apple

​

​

0.02 0.35 0.20

​

0.05 0.15 0.43

​

0.39 0.10 0.06

​

0.00 0.00 0.00

encyclopedic

entries

tourists’

reviews

kids’

books

Surrounding text and bidirectionality

Transformers (Vaswani et al. 2017) and BERT (Devlin et al. 2019):

22

Flamingoes are ________ that fly in large flocks

pink

birds

apple

beautiful

📖 Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems, 2017, 30.

📖 Devlin, J., M.-W. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. NAACL HLT.

Words as numbers

23

Training text

Tokenization

Embeddings

Turin is a fantastic city

#Tur #in #is #a #fan #tast# #i #c #cit# #y

[0.021, -0.183, 0.447, 0.009, -0.56, 0.302, 0.118]�[-0.112, 0.334, -0.221, 0.678, 0.045, -0.390, 0.201]�[0.512, -0.045, 0.119, -0.842, 0.273, 0.066, -0.155]�[-0.301, 0.208, 0.764, -0.014, -0.433, 0.590, 0.037]�[0.089, -0.671, 0.153, 0.402, -0.119, 0.255, -0.488]�[-0.250, 0.143, -0.398, 0.511, 0.322, -0.077, 0.610]

​

Exploring health-related content in transformers

24

25

Comparing “true” and “false” pairs: WHO mythbusters

We tested several models pretrained on these languages.

26

Large Language Model

27

Size of training data

28

(from https://babylm.github.io/)

Large Language Models

Many large language models are multilingual, but…

​

Not every language has the same amount of resources! English is overrepresented also in monolingual-LLMs (e.g. Italian, Orlando et al. to 2024).

​

29

📖 Riccardo Orlando etal. 2024. Minerva LLMs: The first family of large language models trained from scratch on Italian data. In Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), pages 707–719, Pisa, Italy. CEUR Workshop Proceedings.

Conversational AI based on LLMs

Current conversational AI tools are based on these pretrained models.

​

However, there might be human-in-the-loop!

​

30

Testing on ChatGPT

31

ChatGPT. (2024, October 3). Response generated by artificial intelligence (Version 2). OpenAI.

Testing on DeepSeek

32

DeepSeek (15.03.2026)

Are these conversational AI really good across languages?

​

Testing snippets of disasters simulations for deep field (Samo & Parker, in prep.)

33

Hallucinations and sycophancy

34

Sycophantic AI (ChatGPT, 28.05.2026)

35

Sycophantic AI (ChatGPT, 28.05.2026)

36

LLMs and types of Conversational AI

37

• Much of AI interpretability research is done on the internal representations of pretrained models (the vectors we have seen before).

​

• A lot of research is conducted using only pretrained representations through outputs and APIs (which can be very costly).

​

• End users interact with consumer-facing LLMs that are accessible for free or through subscriptions.

Is sharing "sensitive" data possible?

38

• Users’ inputs are not directly incorporated into model training, and any future training use is aggregated and heavily diluted across enormous tokens.
• The impact of isolated and/or sensitive data (e.g., proper names) is typically minimal, especially when such elements are rare and lack statistical relevance at scale. Also even a proper name will be tokenized into small forms. E.g. s# am #o. Still, protect identities is important.
• Data handling is subject to jurisdiction-specific regulations across countries and territories; where retrieval or post-processing is involved, using anonymized or fictional “variables” is recommended to ensure privacy, with controlled decoding if needed.

Interactive session

39

Imagine you are lucky and you still have internet connection in the Mapi region.

​

You need to send a report to a higher health authority.

​

You would like to save time and you need to write something that is a little bit more "verbose" and not just some cells to be filled.

Interactive session: Example

40

Interactive session: filling the evaluation form

41

https://l1nk.dev/gz6jtq7

QR code

If you use your laptop,

you can use this short link to

access and edit the evaluation form

Operational applications: structured natural data

42

Structured data

• Structured data are data that are organized and annotated in a predefined format and schemata, making it easily searchable, machine-readable, and suitable for storage in relational databases.
• Machine Learning algorithms (the inductively-defined one of before) can help in find solutions and patterns.

43

Pre-AI era

44

Pre-AI era

45

Applications of structured data (visual, textual)

• AI-assisted medical image screening and diagnostic triage, allowing remote clinical assessment, but also telemedicine workflows.
• Geospatial intelligence (GIS) integration for real-time mapping and monitoring of road networks or healthcare infrastructure.

​

46

Applications of structured data (visual, textual)

AI-assisted medical image screening and diagnostic triage, enabling remote clinical assessment and telemedicine workflows.

47

https://anjos.ai/research/radiology/

The problem of biases

We have observed bias in LLMs due to the overrepresentation of English.

Similarly, also in vision data, there are issues with population representativeness in datasets (e.g., in skin cancer detection; see Behara et al., 2024; Buolamwini & Gebru, 2018).

📖 Behara K, Bhero E, Agee JT. AI in dermatology: a comprehensive review into skin cancer detection. PeerJ Comput Sci. 2024 Dec 5;10:e2530. doi: 10.7717/peerj-cs.2530. PMID: 39896358; PMCID: PMC11784784.

📖 Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. Conference on Fairness, Accountability and Transparency, 77–91. http://proceedings.mlr.press/v81/buolamwini18a.html?mod=article_inline&ref=akusion-ci-shidai-bizinesumedeia

48

Structured data in GIS:

49

Volunteered Geographic Information

vs.

Authoritative Geographic Information

Structured data

50

Volunteer-data =/= authoratative data

Humanitarian OSM

51

Digital data collection and monitoring

• AI ethics is shaped by diverse regulations and cultural frameworks, which vary across countries and jurisdictions.

​

• Between Ethos and Oikos.

​

• A central guiding principle remain the “do no harm” perspective.

​

• It is important to keep in mind, that the very same structured data (e.g. GIS) and technologies can be used for military purposes and automatization of actions (for example, drones or other technologies).

52

Scenario 2: structured data in Galactica

53

Take home message & conclusions (from Samo & Parker 2025: 152)

​

​

• While AI offers promising tools for risk decision-making, its limitations necessitate careful validation before deployment in real-world applications.

​

• AI can complement human decision-making but should not replace expert judgment, particularly in high-stakes scenarios.

54

📖 Samo, G., & Parker, R. (2025). Exploring AI-Driven Decisions in Disaster Simulation for Emergency Medical Teams. Studies in health technology and informatics, 328, 151–152. https://doi.org/10.3233/SHTI250691

Thank you!

​

Grazie!

55

If you have questions do not hesitate to contact me:�giuseppe.samo@idiap.ch

giuseppe.samo@unige.ch

samo@blcu.edu.cn