Digital Technologies, Artificial Intelligence and Peace: risks and opportunities for conflict-affected Mediterranean societies

1. Context and why it matters now

Digital technologies are now part of the infrastructure of conflict and of peace. Messaging applications carry rumours and evacuation notices; satellite imagery and open-source data feed analysts; generative AI lowers the cost of producing persuasive text, images and audio; and states can switch off connectivity in minutes. The International Committee of the Red Cross (ICRC) notes that harmful information is not new, but that technology has greatly increased its speed and scale, which is especially concerning in crisis settings [17]. The United Nations Secretary-General's policy brief on information integrity states that disinformation and hate speech have been used to incite violence, prolong conflict and thwart peacemaking, and that generative AI has added to the risks [18].

The Mediterranean and its neighbourhood combine several of the conditions that make these questions pressing. The region includes active armed conflict and occupation, fragile transitions, contested elections and protests, large displaced populations and very uneven connectivity. ACLED recorded 204,605 conflict events between 1 December 2024 and 28 November 2025 worldwide, with Palestine ranked as the third most deadly conflict and conflict-related deaths in Syria rising from just over 6,000 to over 9,000 over the preceding twelve months [19]. In 2025 Access Now recorded at least 313 internet shutdowns worldwide, more than in any earlier year it has tracked, and called conflict the leading trigger for the third year running [5].

Governance has not kept pace. A binding European regime for AI is being phased in, with its timetable revised in 2026 [20, 21]; a Council of Europe treaty has opened for signature [13]; and the United Nations has created a scientific panel and a global dialogue on AI [22]. Most of the thirteen countries in which TIPCS has offices sit outside these binding regimes, and the risks and opportunities discussed in this report fall largely in the gaps between them.

This report is a desk review of public sources. It has three aims: to summarise what the evidence supports and what it does not; to map connectivity, shutdowns and online repression across the thirteen countries where comparable data exist; and to set out, in a section clearly labelled as a proposal, how TIPCS would approach research and practice in this field with a do-no-harm protocol. It makes no claim about TIPCS activities beyond those stated in that section.

2. What the evidence says

Harmful information, hate speech and violence

Terminology matters. The ICRC groups misinformation, disinformation and hate speech under the umbrella term 'harmful information' and focuses on harm rather than on whether content is true [17, 23]. The United Nations working definition of hate speech covers 'any kind of communication in speech, writing or behaviour, that attacks or uses pejorative or discriminatory language' toward a person or group on the basis of identity; the UN itself notes that this is not a legal definition and that there is no universal definition in international human rights law, whereas incitement to discrimination, hostility or violence is prohibited [24]. The UN Strategy and Plan of Action on Hate Speech, launched on 18 June 2019, sets 13 commitments for the UN system and rests on respect for freedom of opinion and expression, shared responsibility, digital citizenship and better data [25].

The causal evidence is narrower than the policy rhetoric. The most cited study is an analysis of the 1994 Rwandan genocide, which exploits variation in radio reception across villages. It estimates that about 51,000 perpetrators, or approximately 10% of the overall violence, can be attributed to one station, with spillovers to neighbouring villages larger than direct effects [1]. A study of Germany finds that anti-refugee sentiment on a political party's Facebook page predicts violent crime against refugees in municipalities with higher social media use, and that internet and Facebook outages remove the correlation, which supports a causal reading [2]. Both studies concern specific media, specific audiences and, in the Rwandan case, a state-sponsored campaign. Neither shows that online content alone causes violence; both point to a role as a propagation mechanism where organised actors, grievances and local social networks are already present.

The ICRC identifies four families of harm in armed conflict: direct harm to people, including by misdirecting them away from life-saving information; fuelling hatred and polarisation; weakening compliance with international humanitarian law through dehumanisation; and undermining humanitarian workers through false information about their organisations [23]. It adds that harmful information is not always unlawful, and that international humanitarian law applies online as it does offline [23, 26]. The ICRC statement to a United Nations Security Council Arria-formula meeting in December 2023 refers to a Global Advisory Board report with four principles and 25 recommendations, including that belligerents should avoid internet shutdowns [27]. The United Nations Global Principles for Information Integrity, published in June 2024, set out five principles: societal trust and resilience, healthy incentives, public empowerment, independent, free and pluralistic media, and transparency and research [28]. The Secretary-General's 2023 brief cites a 2022 survey in which 70 per cent of UN peacekeepers said mis- and disinformation had a severe, critical or moderate impact on their work [18].

On remedies, the evidence is thinner still. A set of six experiments and a YouTube field study reported in Science Advances found that short 'inoculation' videos improved recognition of manipulation techniques, by about 5% on average in the field study, but the field test used fictional posts, US users and a voluntary test about 18 hours after viewing, and measured recognition rather than sharing behaviour [29]. The ICRC cautions that watermarking alone will not overcome people's tendency to believe content that matches existing views, and that public detection tools can help creators of fakes to train better ones [26]. SIPRI's 2025 review of civilian AI risks lists influence operations and the spread of inaccurate AI-generated information, which erodes public trust, among the main misuse and accidental risks, and notes that experts disagree more about accidental and structural risks than about misuse [30].

AI for early warning and conflict analysis

Quantitative conflict forecasting has a track record that is better than often assumed, with clear limits. ViEWS, developed at Uppsala University, produces monthly probability forecasts for Africa at country and sub-national level for three types of organised violence, using public data. For state-based conflict in the 2015 to 2017 test period, the ensemble reached an AUROC of 0.9555 at country level and 0.9484 at sub-national level, but the sub-national AUPR was only 0.277; accuracy fell with the forecast horizon, and the country-level ensemble had a worse Brier score (0.0932) than a simpler demography-based model (0.0642) [3]. A nine-year review of forecasts made in 2011 concluded that they were meaningfully accurate but predicted low-level conflict less well than major armed conflict and missed some regional shifts [4]. A forecasting competition for state-based violence in Africa found that a simple weighted ensemble beat the average individual model [31]. ViEWS covers Africa, so these results speak to Kenya only and cannot be read across to the Levant or the Maghreb without testing.

ACLED's Conflict Alert System (CAST) forecasts event counts for battles, explosions and remote violence, and violence against civilians for the next six four-week periods, for every country and at first administrative level, using a gradient-boosted model trained on ACLED data from 2018 together with population, infant mortality and political-institution indicators [32]. ACLED describes itself as an independent, impartial global monitor funded in part by the UN Complex Risk Analytics Fund [33], and its Conflict Index scores countries on deadliness, danger to civilians, geographic diffusion and group fragmentation [19]. CAST forecasts are counts of recorded events, not confirmed outcomes, and the methodology page does not describe a dedicated limitations section [32].

Within the United Nations, the Department of Political and Peacebuilding Affairs (DPPA) lists AI-powered Digital Dialogues, which allow senior officials to consult up to 1,000 participants in their own languages and dialects, a generative AI training programme for political affairs officers aligned with UN guidance on ethical AI use, strategic foresight work, and data resources including a Peace and Security Data Hub [34]. A 2025 review by the Geneva Science and Diplomacy Anticipator (GESDA) reports that DPPA, with academic and private partners, has tested AI tools in Iraq, Libya and Yemen, including AI-assisted digital dialogues, machine learning to gauge public opinion and geospatial analysis for early warning, and records limits: current models have limited ability to infer causes, data quality is uneven across space and time, and data capture raises security and privacy concerns [35]. The CSIS Futures Lab, after a workshop with more than 45 mediators and technology experts, concluded in April 2026 that AI's most credible role is to augment mediator-controlled workflows rather than replace mediator judgement, and warned of risks to neutrality, confidentiality and mediator control [36]. This review found no independent evaluation of the Libya, Iraq or Yemen pilots; the public descriptions do not report results.

PeaceTech Lab describes its mission as using technology, media and data to reduce violent conflict and highlights online hate speech as a focus, notably in South Sudan, where its project combines a lexicon of hateful terms, tools to visualise influence networks and validation with local groups [37, 38]. The public pages reviewed describe methods, not measured effects on violence; this report found no published impact evaluation of such lexicon-based monitoring. The IPIE (International Panel on the Information Environment) scientific panel on AI and peacebuilding, drawing on about 600 documents, describes AI as 'fundamentally a dual-use technology' that can support conflict-risk monitoring, humanitarian coordination and peace dialogues, but can equally expand repression, surveillance and disinformation [16].

Shutdowns, repression and the legal standard

The Office of the UN High Commissioner for Human Rights defines a shutdown as a measure taken by or for a government to intentionally disrupt access to and use of online information and communications systems, including throttling, platform blocking and VPN blocking. It finds that shutdowns 'very rarely meet' the tests of necessity and proportionality under Article 19(3) of the International Covenant on Civil and Political Rights, and that shutdowns during armed operations restrict reporting and human rights monitoring and may conceal abuses [39]. Access Now's 2025 report records 125 conflict-related shutdowns in 14 countries, 40% of the global total, and 70 shutdowns linked to grave human rights abuses, and notes that 14 shutdowns of satellite internet systems occurred in seven countries, including Syria, up from four in 2024 [5]. Freedom House's Freedom on the Net 2025 found the fifteenth consecutive year of global decline, with at least 57 countries seeing arrests or imprisonment for online expression, a record high [40]; the report also warns that sovereign AI projects could embed censorship and surveillance in authoritarian settings [40].

3. The regional picture across the thirteen countries

Connectivity

The ITU estimates that about 6 billion people, roughly three-quarters of the world's population, were online in 2025, with 2.2 billion offline, 96% of them in low- and middle-income countries [41]. Urban use is 85% against 58% in rural areas [41]. Table 1 gives the latest World Bank figures, which are compiled from ITU data, together with start-of-2025 estimates from DataReportal (Kepios) where the World Bank series is old or missing. The two sources use different methods and do not always agree, as the Libya and Lebanon rows show. Mobile subscriptions can exceed 100 per 100 people because one person may hold several SIM cards.

Two features stand out. First, the range is wide: Kenya and Mauritania are well below the others, Syria's most recent World Bank observation dates from 2018 and DataReportal's estimate of 35.8% suggests similarly low use [43], while Jordan, Türkiye, Italy and Lebanon are above or close to 90% depending on the source. Second, the national averages conceal the damage that conflict does to infrastructure and the uneven access within countries; in Gaza, Access Now reports that many local internet service providers have been offline since October 2023 [5]. The ITU's 2025 Global Connectivity Report places Internet use in Africa at just over one in three people, consistent with the World Bank figure for Kenya [58].

On gender, the ITU reports that in 2025 77% of men and 71% of women use the Internet globally, a parity score of 0.92 unchanged since 2019; the score for the Arab States is 0.86, unchanged between 2019 and 2025, for Africa it rose from 0.70 to 0.78, and Europe has reached parity [7]. The GSMA reports that in low- and middle-income countries women were 14% less likely than men to use mobile internet in 2024 [47]. No verified country-level gender gap figures for the thirteen countries could be retrieved in this review; the regional scores should not be assigned to individual countries.

Shutdowns and online freedom

Table 2 combines Access Now's counts for 2024 and 2025 with Freedom House's Freedom on the Net 2025 scores for the countries it assesses and for which a country page was reviewed. Counts are minimum figures, because Access Now notes that some shutdowns go unreported and that its numbers can change as new information emerges [5]. The 2024 report counted 296 shutdowns in 54 countries [48]; the 2025 report revised the 2024 total to 304 [5], so 2024 country counts may also have been revised.

Three patterns appear. Exam-related disruptions are the most routine: Access Now counted 11 exam-related shutdowns in six countries in 2025, including Algeria, Iraq, Jordan, Kenya and Syria, with Iraq alone imposing about two hours of daily disruption throughout exam periods [5]. Protest-related disruptions appear in Libya, where authorities imposed shutdowns in Tripoli on 16 May 2025 during clashes between armed groups and protests against the Government of National Unity, with reports of live ammunition used against protesters [5]; Freedom House describes the Tripoli shutdown as lasting about 12 hours [50]. Conflict-related shutdowns affect Palestine and Syria, where in 2025 the authorities also confiscated Starlink devices and prohibited unlicensed satellite internet [5].

Absence from the shutdown count does not mean an open environment. Freedom House scores Egypt at 28 out of 100, Not Free, and describes targeted censorship and surveillance rather than wholesale shutdowns, with the blocking of news outlets and the detention of more than 100 people who used social media to call for protests over power cuts and rising prices [40, 49]. Türkiye is scored 31, with throttling of social media for 42 hours in March 2025 and a nine-day block of Instagram in August 2024 [51]. Jordan is scored 47, with enforcement of its 2023 cybercrime law and the blocking of 12 online news outlets in May 2025 [52]. Iraq is scored 41, with exam shutdowns and blocking of sites on grounds of 'immoral content' [53]. Kenya's score fell from 64 to 58, the largest fall in the survey, after a roughly seven-hour disruption on 25 June 2024 during protests against the Finance Bill [40, 54]. Tunisia is scored 59, a fall of one point, with prosecutions under Decree-Law 2022-54 [55]. These are findings of one monitoring organisation, reported here with attribution; governments cited in the sources frequently give security, public order or anti-cheating reasons for the measures.

4. Normative and regulatory frameworks

Five frameworks matter most for TIPCS countries, and their reach differs sharply.

• EU AI Act. The Regulation applies to providers placing AI systems on the EU market and to third-country providers and deployers whose output is used in the EU, but not to systems used exclusively for military, defence or national security purposes, nor to systems developed solely for scientific research [12]. Article 5 prohibits, among other practices, harmful manipulation, social scoring, untargeted scraping of facial images, emotion recognition in workplaces and schools, and most real-time remote biometric identification for law enforcement, with exceptions [56]. In 2026 the EU adopted an 'AI Omnibus' which, according to law-firm summaries, moved obligations for stand-alone high-risk systems to 2 December 2027 and for systems embedded in regulated products to 2 August 2028, while general-purpose AI obligations and most transparency obligations were not delayed [20, 21]. Italy is directly bound; other TIPCS countries are affected only through exports to, or use in, the EU.
• Council of Europe Framework Convention on Artificial Intelligence. Opened for signature on 5 September 2024, it covers public authorities and private actors acting on their behalf, with parties choosing how to apply it to the private sector; parties need not apply it to national security activities but must ensure those respect international law and democratic institutions, and it does not cover national defence [13]. The European Union ratified it on 15 May 2026 [57]. Apart from the EU, none of the TIPCS countries appears among the parties or signatories listed on the Council's page [13].
• UN Global Digital Compact. Opened for endorsement in September 2024, it commits governments to an international scientific panel on AI and a global AI policy dialogue, and to promote access to independent, fact-based information against mis- and disinformation, and describes the internet as 'open, global, stable and secure' [14]. The Compact is a political commitment and does not create binding obligations. Its panel, the Independent International Scientific Panel on AI, was created by General Assembly resolution A/RES/79/325 of 26 August 2025, has 40 members, is limited to the non-military domain, published a preliminary report on 1 July 2026, and met the first Global Dialogue on AI Governance in Geneva on 6 and 7 July 2026 [22].
• UNESCO Recommendation on the Ethics of Artificial Intelligence. Adopted in November 2021 by 193 member states, it sets ten principles including proportionality and do no harm, human oversight, transparency and fairness and non-discrimination [15]. It is a non-binding standard.
• Information integrity instruments. The UN Strategy on Hate Speech [25], the Secretary-General's 2023 policy brief [18] and the 2024 Global Principles for Information Integrity [28] are non-binding. Binding law relevant to shutdowns is Article 19 of the International Covenant on Civil and Political Rights as interpreted by the High Commissioner's 2022 report [39].

The practical consequence is a gap. The most detailed binding AI rules apply in the EU, exclude national security and defence, and do not directly regulate how governments in the rest of the region use surveillance, forecasting or content-filtering tools. The instruments that do reach all thirteen countries are non-binding, and their implementation in conflict settings has not been systematically assessed in the sources reviewed.

5. Arabic, dialects and low-resource language bias

Early warning, moderation and dialogue tools depend on language. The evidence available indicates that Arabic, and in particular its dialects, is poorly served.

• Content moderation. An independent human rights due diligence review commissioned by Meta and published in September 2022 found 'unintentional bias' in Meta's policy and practice during the May 2021 escalation in Israel and Palestine, including that routing Arabic content to reviewers by dialect or regional expertise was insufficient; it found both over-enforcement and under-enforcement and recommended changing the routing [8]. Findings of a review commissioned by a platform are not independent verification of platform behaviour, but the report was published in Arabic, English and Hebrew.
• Maghrebi Arabic. A September 2024 report by the Center for Democracy and Technology reports that models trained on Moroccan or Algerian Arabic perform worse than those on Modern Standard Arabic because of limited data, that training sets lack examples of Arabizi and code-switching, that annotators who do not speak the dialect produce inconsistent labels, and that moderators are often assigned content from any Arab country regardless of dialect; 57% of surveyed users did not believe platforms do enough to moderate Maghreb content [9].
• Cultural bias in language models. A study that tested 16 models in Arabic on a benchmark of 628 prompts and 20,368 entities reports that multilingual and Arabic monolingual models favour entities associated with Western culture, with cases of stereotyping, and that widely used pre-training sources such as Wikipedia may be poorly suited to culturally aware models [10]. This is a preprint, last revised in March 2024, and the models tested are now dated.
• Dialect and cultural benchmarks. The AraDiCE benchmarks add seven synthetic dialect datasets alongside Modern Standard Arabic and a cultural benchmark covering Gulf, Egyptian and Levantine contexts; Arabic-specific models such as Jais and AceGPT outperformed multilingual models on dialectal tasks, while dialect identification, generation and translation remained difficult [11]. The cultural benchmark covers Gulf, Egyptian and Levantine regions only, and this review found no equivalent for Maghrebi or Hassaniya varieties.

These results matter for peace practice in specific ways. A hate-speech classifier that misreads a dialect can over-remove legitimate speech or miss incitement, and either error can deepen mistrust of institutions [8, 9]. A summarisation tool used in a mediation process that favours Western framings may distort the positions of parties [36, 10]. The IPIE panel recommends backing smaller language models tailored to local languages and contexts and making data on both successes and failures openly available [16]. This review found no published evaluation of the performance of general-purpose language models on conflict-related Arabic text for Libyan, Syrian, Iraqi or Palestinian dialects.

6. Analysis: mechanisms, openings and counter-evidence

Four mechanisms recur across the evidence. First, harmful information acts through amplification of existing grievances and organised networks rather than as an independent cause [1, 2]. Second, trust is both target and shield: the UN principles identify societal trust and resilience and independent, pluralistic media as foundations of information integrity [28], and the ICRC notes that false information about humanitarian actors erodes the acceptance on which access depends [23]. Third, connectivity creates exposure and protection at once. The German study found that outages removed the link between hate speech and violence [2], but the High Commissioner finds that shutdowns rarely meet legal tests and that they restrict monitoring and may conceal abuses [39]; the ICRC likewise calls restrictions counterproductive [23]. Fourth, analytic tools inherit the limits of their data: ACLED and ViEWS rely on recorded events [3, 32], and where reporting is restricted, whether by shutdowns or by journalists' safety, recorded violence understates real violence and the models learn from the gap.

Openings for cooperation follow from these mechanisms. Measures that do not depend on government control of speech command the broadest support: media and digital literacy, access to reliable information, protection of independent media and transparency and researcher access to platform data are named by the ICRC and the UN [18, 28, 23]. Because shutdowns during exams are a recurring, low-politics practice recorded in Algeria, Iraq, Jordan, Kenya and Syria [5], they offer a concrete entry point for dialogue on proportionate alternatives, although this review did not assess the evidence on the effectiveness of alternatives in the region. Forecasting offers cooperation opportunities where data are shared transparently and results are used for preventive dialogue rather than security targeting [35, 36].

Counter-evidence and caveats must be stated. Forecasting performance reported for Africa [3, 4, 31] cannot be assumed for other regions. Evidence on inoculation and literacy comes largely from Western samples and short-term outcomes [29]. Lexicon-based hate-speech monitoring has no verified impact evaluation [37, 38]. Some of the strongest evidence of harm from social media comes from Germany and Rwanda, with different media and political contexts [1, 2]. The same tools that support peace can support repression: the IPIE panel stresses dual use [16], and Freedom House warns of AI-enabled surveillance and censorship [40]. Finally, much of the data on shutdowns and online freedom comes from monitoring organisations whose methods rely on technical measurement plus news and personal accounts [5]; they are credible but not neutral arbiters, and governments often dispute their characterisations.

7. Implications for policy and practice

The following implications are drawn from the evidence reviewed. They are addressed to kinds of actors and do not assess any specific government.

• Governments: refrain from blanket shutdowns and meet the six conditions set out by the High Commissioner (clear public law, necessity, proportionality, prior independent authorisation, advance notice, redress) where any restriction is contemplated [39]; publish information on orders and their scope; examine alternatives to exam-related shutdowns [5, 39]; and build civilian AI governance that covers public-sector use, including security uses, which the EU and Council of Europe regimes largely leave to national law [12, 13].
• Donors: fund independent evaluation of peacebuilding AI pilots, with open publication of failures [16]; support language resources and smaller local models for Arabic dialects, Amazigh, Kurdish, Swahili and other languages of the region, designed with native-speaker annotators [16, 9]; condition connectivity programmes on shutdown risk assessments, as the High Commissioner recommends for development organisations [39]; and avoid funding tools whose data governance is unclear.
• Platforms and technology companies: assign Arabic content to reviewers who speak the relevant dialect, invest in training data for Maghrebi and other dialects, publish enforcement data, and conduct human rights due diligence and conflict-sensitive moderation [26, 8, 9, 39].
• Municipalities and public service providers: plan for connectivity failure, since shutdowns and infrastructure damage interrupt health, education and emergency communication [39]; communicate through trusted community channels and not only through social platforms [26].
• Civil society and media: strengthen detection and documentation of shutdowns and of harmful information targeting communities, in cooperation with measurement networks, and protect staff data [5, 39]; adopt a do-no-harm protocol before using AI tools (see below).
• Researchers: test forecasting and classification tools outside the regions where they were validated, report error rates by dialect, and release evaluation data under ethical safeguards [3, 9, 11].

8. What TIPCS proposes

Research questions

1. How do harmful information and hate speech circulate in Arabic dialects, French and English around specific triggers (elections, displacement, protests), and what local response practices reduce harm?
2. How accurately do publicly available language models and classifiers handle conflict-related text in Tunisian, Algerian, Libyan, Egyptian, Hassaniya, Levantine and Iraqi Arabic compared with Modern Standard Arabic?
3. What are the documented effects of exam-related and protest-related shutdowns on service delivery and local peace processes, and what alternatives exist?
4. Under what conditions can open-source early-warning outputs be used in dialogue with local actors without exposing communities to harm?

Methods and data

TIPCS would conduct desk reviews and secondary analysis of public datasets (ACLED, Access Now, Freedom House, ITU, World Bank), and build small, documented evaluation sets of publicly posted text with native-speaker annotators, working only with content that is public and aggregating before reporting. Where it tests classifiers or language models, TIPCS would publish error rates by dialect and would not release datasets containing identifiable content. Any qualitative component would use informed consent, option to withdraw and secure storage, under its research ethics and editorial independence policy.

Do-no-harm protocol for AI in peacebuilding

Drawing on the recommendations of the IPIE panel [16], the UNESCO principles of proportionality and do no harm and human oversight [15], and CSIS findings on mediator control [36], TIPCS would apply the following before using any AI tool in its work: a written purpose and necessity test; a data protection assessment that excludes personal data of conflict-affected people unless essential and consented; human review of every output used in analysis or communication; bias testing for the specific dialects involved; a check that the tool's vendor does not retain or train on submitted data; a ban on using tools for surveillance, profiling or targeting of individuals; and an incident log with a route for complaint under its whistleblowing and complaints policy. Its safeguarding and PSEA policy and its conflict of interest and sanctions and due diligence policies would apply to technology partners.

Partnership model and measurement

TIPCS would work with national universities, language technology groups, digital rights monitors and measurement networks, and would seek to consult affected communities through its country offices. Success would be assessed by whether the outputs are used: published evaluation datasets and error-rate reports, a public protocol others can adopt, training modules in Arabic, French and English, and independent peer review. TIPCS would not claim effects on violence from this work, because the evidence reviewed above shows that such claims require designs that desk research cannot provide.

9. Risks, limitations and what we do not know

Several evidence gaps limit what can be concluded. There is no verified, comparable country-level data on the gender gap in internet use for most of the thirteen countries. There is no published impact evaluation found for AI-assisted dialogue pilots in Libya, Iraq and Yemen, nor for lexicon-based hate-speech monitoring. ViEWS and its evaluation cover Africa only. We found no peer-reviewed assessment of general-purpose language-model performance on conflict-related Libyan, Syrian, Iraqi or Palestinian Arabic. Data on shutdowns rely on technical measurement and reporting, so counts are minima. The status of EU AI Act obligations changed during 2026 and was reviewed here through law-firm summaries rather than the Official Journal text, so dates should be verified before use [20, 21]. Finally, risks of the work itself should be acknowledged: research on harmful information can itself expose communities, tools built for early warning can be used for targeting, and the publication of evaluation sets can aid those who wish to evade detection. These risks are the reason for the do-no-harm protocol above and are not eliminated by it.

10. Conclusion

The evidence supports a measured position. Harmful information matters most when it meets organised violence, grievance and weak trust; AI-assisted analysis can add value when it is transparent, locally validated and under human control; shutdowns and online repression are common in the region and are harmful for monitoring and protection; and the language gap for Arabic dialects is a practical obstacle to using current tools safely. The normative frameworks that exist are uneven in reach and mostly exclude security uses. Progress will depend less on adopting new tools than on evaluating the tools that exist, on sharing evidence of what fails, and on protecting the connectivity and independent media that peace processes rely on. TIPCS proposes to contribute to that evidence with a research programme designed for independence, transparency and do-no-harm.

Method and limits of this brief

This report is a desk review of public sources accessed on 26 October 2026; it does not draw on primary fieldwork, interviews or non-public data. Several documents were read through web summaries or abstracts rather than in full text, including the Rwanda and Germany studies, the ViEWS article, the inoculation study and the EU AI Act omnibus (reviewed through law-firm summaries). Country figures from the World Bank and DataReportal use different methods and years, and a few series are old (Syria 2018, Libya and Lebanon mobile 2022). Shutdown counts are minimum figures from one monitoring coalition and were revised between reports. Freedom on the Net scores were retrieved for seven of the thirteen countries only. No verified country-level data on the gender gap in internet use were retrieved. Author lists for the AraDiCE and ViEWS papers were not confirmed from the pages reviewed and are given as 'et al.' or by title.

References

1. [1]Yanagizawa-Drott, D. (2014). Propaganda and Conflict: Evidence from the Rwandan Genocide. The Quarterly Journal of Economics, 129(4), 1947-1994. https://ideas.repec.org/a/oup/qjecon/v129y2014i4p1947-1994.html (accessed 2026-10-26)
2. [2]Müller, K. and Schwarz, C. (2021). Fanning the Flames of Hate: Social Media and Hate Crime. Journal of the European Economic Association, 19(4), 2131-2167 (working paper: CAGE Online Working Paper 373, 2018). https://ideas.repec.org/p/cge/wacage/373.html (accessed 2026-10-26)
3. [3]Hegre, H. et al. (2019). ViEWS: A political violence early-warning system. Journal of Peace Research. https://journals.sagepub.com/doi/10.1177/0022343319823860 (accessed 2026-10-26)
4. [4]Hegre, H., Nygård, H. M. and Landsverk, P. (2021). Can We Predict Armed Conflict? How the First 9 Years of Published Forecasts Stand Up to Reality. International Studies Quarterly, 65(3), 660-668. PRIO. https://www.prio.org/publications/12592 (accessed 2026-10-26)
5. [5]Access Now and #KeepItOn coalition (2026). Rising repression meets global resistance: Internet shutdowns in 2025 (KeepItOn annual report, March 2026). Access Now. https://www.accessnow.org/wp-content/uploads/2026/03/KeepItOn-Internet-Shutdowns-2025-Annual-Report.pdf (accessed 2026-10-26)
6. [6]World Bank (2026). Individuals using the Internet (% of population), indicator IT.NET.USER.ZS. World Development Indicators (source: ITU), API data last updated 8 October 2026. https://data.worldbank.org/indicator/IT.NET.USER.ZS (accessed 2026-10-26)
7. [7]ITU (2025). Data show a lack of progress in closing the gender gap in Internet use (Facts and Figures 2025). International Telecommunication Union. https://www.itu.int/itu-d/reports/statistics/2025/10/15/ff25-the-gender-digital-divide/ (accessed 2026-10-26)
8. [8]BSR (2022). Human Rights Due Diligence of Meta's Impacts in Israel and Palestine in May 2021. Business for Social Responsibility, 22 September 2022 (commissioned by Meta). https://www.bsr.org/en/blog/human-rights-due-diligence-of-meta-impacts-in-israel-and-palestine-may-2021 (accessed 2026-10-26)
9. [9]Elswah, M. (2024). Moderating Maghrebi Arabic Content on Social Media. Center for Democracy and Technology, September 2024. https://cdt.org/wp-content/uploads/2024/09/2024-09-26-CDT-Research-Global-South-Moderating-Report-English-Arabic-final.pdf (accessed 2026-10-26)
10. [10]Naous, T., Ryan, M. J., Ritter, A. and Xu, W. (2024). Having Beer after Prayer? Measuring Cultural Bias in Large Language Models. arXiv:2305.14456 (v4, 20 March 2024). https://arxiv.org/abs/2305.14456 (accessed 2026-10-26)
11. [11]AraDiCE authors (2025). AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs. Proceedings of COLING 2025 (arXiv:2409.11404). https://arxiv.org/abs/2409.11404v2 (accessed 2026-10-26)
12. [12]European Union (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 2: Scope (text as presented by the Future of Life Institute AI Act Explorer). https://artificialintelligenceact.eu/article/2/ (accessed 2026-10-26)
13. [13]Council of Europe (2024). Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225). Council of Europe. https://www.coe.int/en/web/artificial-intelligence/the-framework-convention-on-artificial-intelligence (accessed 2026-10-26)
14. [14]United Nations (2024). Global Digital Compact. United Nations. https://www.un.org/global-digital-compact/en (accessed 2026-10-26)
15. [15]UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence. United Nations Educational, Scientific and Cultural Organization. https://www.unesco.org/en/artificial-intelligence/recommendation-ethics (accessed 2026-10-26)
16. [16]International Panel on the Information Environment (2025). Artificial Intelligence and Peacebuilding: Opportunities and Challenges (technical paper of the IPIE Scientific Panel on AI and Peacebuilding), press release of 22 September 2025. IPIE. https://www.ipie.info/news/press-release-the-double-edged-sword-of-ai-in-peacebuilding-why-human-rights-must-guide-its-design (accessed 2026-10-26)
17. [17]ICRC (2021). Harmful information: Misinformation, disinformation and hate speech in armed conflict and other situations of violence. International Committee of the Red Cross. https://www.icrc.org/en/publication/4556-harmful-information-misinformation-disinformation-and-hate-speech-armed-conflict (accessed 2026-10-26)
18. [18]United Nations (2023). Our Common Agenda Policy Brief 8: Information Integrity on Digital Platforms. United Nations. https://www.un.org/sites/un2.un.org/files/our-common-agenda-policy-brief-information-integrity-en.pdf (accessed 2026-10-26)
19. [19]ACLED (2025). Conflict Index. Armed Conflict Location and Event Data Project. https://acleddata.com/conflict-index/ (accessed 2026-10-26)
20. [20]National Law Review (2026). EU Digital Omnibus on AI enters into force (law-firm summary: Official Journal publication 24 July 2026, entry into force 27 July 2026). https://natlawreview.com/article/eu-digital-omnibus-ai-enters-force (accessed 2026-10-26)
21. [21]Sidley Austin (2026). EU Lawmakers Reach Provisional Agreement to Delay Key EU AI Act Obligations. Data Matters blog, 22 June 2026. https://datamatters.sidley.com/2026/06/22/eu-lawmakers-reach-provisional-agreement-to-delay-key-eu-ai-act-obligations/ (accessed 2026-10-26)
22. [22]United Nations (2025-2026). Independent International Scientific Panel on AI: FAQ (General Assembly resolution A/RES/79/325 of 26 August 2025). United Nations. https://www.un.org/independent-international-scientific-panel-ai/en/faq (accessed 2026-10-26)
23. [23]ICRC (2024). Why is the ICRC concerned by harmful information in war? Law and Policy Blog, 10 September 2024. International Committee of the Red Cross. https://blogs.icrc.org/law-and-policy/2024/09/10/why-is-the-icrc-concerned-by-harmful-information-in-war/ (accessed 2026-10-26)
24. [24]United Nations (n.d.). Understanding hate speech: What is hate speech? United Nations. https://www.un.org/en/hate-speech/understanding-hate-speech/what-is-hate-speech (accessed 2026-10-26)
25. [25]United Nations (2019). United Nations Strategy and Plan of Action on Hate Speech. United Nations. https://www.un.org/en/hate-speech/un-strategy-and-plan-of-action-on-hate-speech (accessed 2026-10-26)
26. [26]Bonessi, D. M. (2025). Digital Dilemmas 2.0: Deepfakes. Intercross blog, 20 March 2025. International Committee of the Red Cross. https://blogs.icrc.org/intercross/2025/03/20/digital-dilemmas-2-0-deepfakes/ (accessed 2026-10-26)
27. [27]ICRC (2023). Digitalization of armed conflict brings new threats for civilians: statement to the Arria-formula meeting on AI and its impact on hate speech, disinformation and misinformation, 19 December 2023. International Committee of the Red Cross. https://www.icrc.org/en/arria-formula-statement-ai-mdh (accessed 2026-10-26)
28. [28]United Nations (2024). United Nations Global Principles for Information Integrity: Recommendations for Multi-stakeholder Action. United Nations, June 2024. https://digitallibrary.un.org/record/4052524 (accessed 2026-10-26)
29. [29]Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S. and Lewandowsky, S. (2022). Psychological inoculation improves resilience against misinformation on social media. Science Advances, 8(34) (as summarised by ScienceDaily, 24 August 2022). https://www.sciencedaily.com/releases/2022/08/220824152220.htm (accessed 2026-10-26)
30. [30]Boulanin, V., Palayer, J. and Ovink, C. (2025). Addressing the Risks that Civilian AI Poses to International Peace and Security: The Role of Responsible Innovation. SIPRI, November 2025. https://www.sipri.org/publications/2025/other-publications/addressing-risks-civilian-ai-poses-international-peace-and-security-role-responsible-innovation (accessed 2026-10-26)
31. [31]Vesco, P. et al. (2022). United they stand: Findings from an escalation prediction competition. International Interactions, 48(4). PRIO. https://www.prio.org/publications/13337 (accessed 2026-10-26)
32. [32]ACLED (n.d.). CAST methodology (Conflict Alert System). Armed Conflict Location and Event Data Project. https://acleddata.com/methodology/cast-methodology (accessed 2026-10-26)
33. [33]ACLED (n.d.). About ACLED. Armed Conflict Location and Event Data Project. https://acleddata.com/about-acled/ (accessed 2026-10-26)
34. [34]UN Department of Political and Peacebuilding Affairs (n.d.). Innovation. United Nations DPPA. https://dppa.un.org/en/dppa/innovation (accessed 2026-10-26)
35. [35]Heilprin, J. (2025). Radar Spotlight: Peering into the future of conflict prevention with predictive peacebuilding. GESDA, 22 January 2025. https://gesda.global/the-future-of-conflict-prevention-is-predictive-peacebuilding/ (accessed 2026-10-26)
36. [36]Atalan, Y., Jensen, B. and Reynolds, I. (2026). AI and the Future of Mediation. CSIS Futures Lab, 27 April 2026. Center for Strategic and International Studies. https://www.csis.org/analysis/ai-and-future-mediation (accessed 2026-10-26)
37. [37]PeaceTech Lab (n.d.). Home page and programme descriptions. PeaceTech Lab. https://www.peacetechlab.org/ (accessed 2026-10-26)
38. [38]Dangerous Speech Project (n.d.). Analyzing and monitoring hate speech in South Sudan (PeaceTech Lab project description). https://dangerousspeech.org/south-sudan (accessed 2026-10-26)
39. [39]OHCHR (2022). Internet shutdowns: trends, causes, legal implications and impacts on a range of human rights, report of the United Nations High Commissioner for Human Rights, A/HRC/50/55, 13 May 2022. United Nations. https://documents.un.org/doc/undoc/gen/g22/341/55/pdf/g2234155.pdf (accessed 2026-10-26)
40. [40]Freedom House (2025). Freedom on the Net 2025: An Uncertain Future for the Global Internet. Freedom House. https://freedomhouse.org/report/freedom-net/2025/uncertain-future-global-internet (accessed 2026-10-26)
41. [41]ITU (2025). Measuring Digital Development: Facts and Figures 2025, press release of 17 November 2025. International Telecommunication Union. https://www.itu.int/en/mediacentre/Pages/PR-2025-11-17-Facts-and-Figures.aspx (accessed 2026-10-26)
42. [42]World Bank (2026). Mobile cellular subscriptions (per 100 people), indicator IT.CEL.SETS.P2. World Development Indicators (source: ITU), API data last updated 8 October 2026. https://data.worldbank.org/indicator/IT.CEL.SETS.P2 (accessed 2026-10-26)
43. [43]Kemp, S. (2025). Digital 2025: Syria. DataReportal (Kepios), 3 March 2025. https://datareportal.com/reports/digital-2025-syria (accessed 2026-10-26)
44. [44]Kemp, S. (2025). Digital 2025: Libya. DataReportal (Kepios). https://datareportal.com/reports/digital-2025-libya (accessed 2026-10-26)
45. [45]Kemp, S. (2025). Digital 2025: Lebanon. DataReportal (Kepios). https://datareportal.com/reports/digital-2025-lebanon (accessed 2026-10-26)
46. [46]Kemp, S. (2025). Digital 2025: Palestine. DataReportal (Kepios). https://datareportal.com/reports/digital-2025-palestine (accessed 2026-10-26)
47. [47]GSMA (2025). The Mobile Gender Gap Report 2025: Progress closing the mobile internet gender gap stalls in LMICs. GSMA, 14 May 2025. https://www.gsma.com/newsroom/press-release/progress-closing-the-mobile-internet-gender-gap-stalls-in-lmics-gsma-mobile-gender-gap-report-2025 (accessed 2026-10-26)
48. [48]Access Now and #KeepItOn coalition (2025). Emboldened offenders, endangered communities: Internet shutdowns in 2024 (KeepItOn annual report). Access Now. https://www.accessnow.org/press-release/keepiton-internet-shutdowns-2024-en/ (accessed 2026-10-26)
49. [49]Freedom House (2025). Egypt: Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/egypt/freedom-net/2025 (accessed 2026-10-26)
50. [50]Freedom House (2025). Libya: Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/libya/freedom-net/2025 (accessed 2026-10-26)
51. [51]Freedom House (2025). Türkiye (Turkey): Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/turkey/freedom-net/2025 (accessed 2026-10-26)
52. [52]Freedom House (2025). Jordan: Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/jordan/freedom-net/2025 (accessed 2026-10-26)
53. [53]Freedom House (2025). Iraq: Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/iraq/freedom-net/2025 (accessed 2026-10-26)
54. [54]Freedom House (2025). Kenya: Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/kenya/freedom-net/2025 (accessed 2026-10-26)
55. [55]Freedom House (2025). Tunisia: Freedom on the Net 2025 Country Report. Freedom House. https://freedomhouse.org/country/tunisia/freedom-net/2025 (accessed 2026-10-26)
56. [56]European Union (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5: Prohibited AI practices (text as presented by the Future of Life Institute AI Act Explorer). https://artificialintelligenceact.eu/article/5/ (accessed 2026-10-26)
57. [57]Council of Europe (2026). European Union ratifies the Council of Europe Framework Convention on Artificial Intelligence, 15 May 2026. Council of Europe. https://www.coe.int/en/web/artificial-intelligence/-/european-union-ratifies-the-council-of-europe-framework-convention-on-artificial-intelligence (accessed 2026-10-26)
58. [58]ITU (2025). Global Connectivity Report 2025, chapter 2. International Telecommunication Union. https://www.itu.int/itu-d/reports/statistics/2025/11/17/gcr-2025-chapter-2/ (accessed 2026-10-26)

How to cite

Tunisian Institute for Peace and Conflict Studies (TIPCS) (2026). Digital Technologies, Artificial Intelligence and Peace: risks and opportunities for conflict-affected Mediterranean societies. Monastir: TIPCS. www.tipcs.org/publications/digital-peace-ai