TNS0013

Written evidence submitted by Dr Waddington

Unaddressed Vulnerabilities in NSS 2025: Generative AI as a Threat to UK National Security and Democratic Integrity

Executive Summary

• GenAI and large language models (LLMs) present serious national security risks through embedded bias, hallucinations, deepfakes, and systematic marginalisation—vulnerabilities that hostile actors can exploit to undermine UK democracy and social cohesion.
• The NSS 2025 recognises frontier technologies but fails to address how GenAI bias creates exploitable weaknesses, how AI hallucinations threaten intelligence integrity, or how deepfakes enable unprecedented influence operations.
• Current governance arrangements lack the technical expertise and humanities perspective essential for understanding how AI systems marginalise communities and enable censorship overreach.
• The Foreign Influence Registration Scheme (FIRS) cannot capture sophisticated AI-mediated manipulation campaigns that exploit bias patterns or deploy deepfakes at scale.
• Without immediate action, the UK faces a critical period where adversaries can weaponise AI bias and synthetic content generation against democratic institutions.

Key Recommendations

1. Establish mandatory bias auditing for all AI systems used in security-sensitive contexts
2. Create hallucination detection and flagging requirements for AI systems processing official information
3. Implement technical standards for deepfake identification with regular capability assessments
4. Expand FIRS to explicitly cover AI-generated manipulation campaigns
5. Establish an independent AI oversight body with mandatory humanities and civil society representation

Introduction

I submit this evidence as an Associate Professor of International History, specialising in genocides and international relations. My disciplinary background provides particular expertise in recognising patterns of systemic harm, manipulation, and the erosion of democratic norms. Alongside this, my recent research has examined how artificial intelligence systems perpetuate bias, enable censorship overreach, and contribute to the marginalisation of vulnerable groups. While my work does not primarily focus on national security studies, these findings have led me to consider the serious ramifications if such technologies were to be abused by actors hostile to the UK’s interests. I therefore bring a humanities perspective that highlights how AI systems can generate risks to democratic integrity and social cohesion.

The NSS 2025's treatment of AI risks focuses primarily on technological capacity and cyber threats whilst overlooking the profound vulnerabilities created by AI bias, hallucinations, and deepfake generation. These gaps represent immediate exploitable weaknesses in UK democratic resilience.

1. Critical Gaps: The NSS Underestimates GenAI Exploitation Vectors

AI Bias as a National Security Vulnerability

The NSS fails to recognise that biased AI systems create systematic vulnerabilities that hostile actors can exploit. Current GenAI systems exhibit documented bias patterns that can be weaponised through targeted disinformation campaigns, algorithmic marginalisation of specific communities, and deliberate amplification of social divisions to erode national cohesion.

The DeepSeek Case Study

Recent analysis of China's DeepSeek AI system demonstrates how large language models function as instruments of information power. LLMs inevitably reflect the biases embedded in their training data and development choices, but the critical security distinction lies in whether these biases are systematic and intentional.

State-directed AI development programs deliberately embed political narratives and ideological positions aligned with national objectives. These systems can subtly shape public discourse in ways that support geopolitical strategy, creating exploitable fault lines for foreign influence operations. Commercial models, while less directly tied to national policy, carry different risks—prioritising market imperatives and user retention in ways that may amplify harmful biases or create strategic dependencies.

Strategic Implications

Both pathways pose security risks, but state-directed models represent a qualitatively different threat. They transform AI systems from tools with incidental biases into deliberate instruments of soft power projection. When deployed at scale, these systems can systematically exclude marginalised voices from public discourse, manipulate information environments, and undermine democratic representation.

The persistence of bias across all AI systems underscores that LLMs are not neutral technologies but socio-technical systems whose outputs advance the priorities of their creators. The NSS must acknowledge this reality and develop frameworks to identify, assess, and counter AI-enabled information warfare before these vulnerabilities are fully exploited.

AI Hallucinations: A Critical Intelligence Integrity Threat

AI hallucinations—when systems generate false but plausible information—pose unprecedented risks that the NSS does not address:

• False Intelligence: AI systems providing fabricated "evidence" or analysis to government officials
• Synthetic Corroboration: Multiple AI systems producing consistent but false information, creating illusion of verified intelligence
• Public Trust Erosion: Exposure of AI hallucinations in official contexts undermines confidence in government communications

Deepfakes: Beyond Detection to Democratic Disruption

Whilst the NSS mentions synthetic media, it underestimates the sophistication of current deepfake capabilities:

• Political Destabilisation: High-quality deepfakes of political figures can create international incidents or domestic crises within hours
• Evidence Manipulation: Deepfake audio/video evidence that overwhelms current detection capabilities
• Scale and Speed: AI-generated deepfakes can flood information ecosystems faster than human fact-checkers can respond

2. Governance Failures: Why Technical Expertise Alone Is Insufficient

The Censorship Trap

As I have argued in my published research, addressing AI-enabled disinformation risks creating dangerous censorship overreach. The NSS's approach risks concentrating content moderation decisions in technical and security communities without sufficient democratic oversight.

Critical Issue: Decisions about harmful AI content cannot be made solely by technologists or security officials, as they lack the humanities expertise necessary to understand cultural context, power dynamics, and democratic values at stake.

Solution: Mandatory inclusion of humanities scholars, ethicists, and civil society representatives in all AI content governance decisions, as outlined in my previous work on AI censorship oversight for Times Higher Education.

Marginalisation Through AI Governance

Current proposals for AI regulation risk further marginalising already vulnerable communities:

• Algorithmic Discrimination: AI systems trained on biased data systematically disadvantage minority communities
• Voice Suppression: Over-broad content moderation disproportionately affects marginalised speakers
• Democratic Exclusion: Technical governance structures exclude communities most affected by AI bias

3. FIRS Inadequacy: Missing AI-Mediated Influence Operations

Current Scope Limitations

FIRS focuses on traditional influence activities but cannot capture sophisticated AI-mediated campaigns:

• Deepfake Loophole: No explicit coverage of AI-generated deepfake campaigns
• Bias Exploitation: No mechanism to identify hostile actors exploiting AI bias patterns
• Scale Mismatch: Registration requirements unsuited for automated, scalable GenAI campaigns

Technical Evidence for Expansion

Deepfake Sophistication: Current deepfake technology can produce convincing synthetic content at scale, overwhelming traditional detection methods. Recent analysis shows detection accuracy declining as generation quality improves.

Bias Exploitation Capabilities: Hostile actors can systematically exploit known bias patterns in UK AI systems without creating detectable "influence activity" under current FIRS definitions.

Hallucination Weaponisation: AI hallucinations can be deliberately triggered to create false but plausible information that appears to come from legitimate sources.

4. Funding and Implementation Priorities

Ring-Fenced Resources Needed

The NSS's ambitious technology spending fails to specify funding for:

• Bias Detection Infrastructure: Technical capabilities for identifying and monitoring AI bias in security contexts
• Hallucination Monitoring: Systems for detecting and flagging AI-generated misinformation
• Deepfake Detection: Rapidly evolving technical capabilities to identify synthetic content
• Humanities Expertise: Essential non-technical oversight capabilities

Cross-Departmental Coordination Gaps

Current departmental divisions create dangerous blind spots:

• Cabinet Office: Policy coordination without technical AI expertise
• DCMS: Digital policy without security context
• Home Office: Security focus without AI bias understanding
• DSIT: Technical focus without democratic governance expertise

Recommendation: Create dedicated cross-departmental AI democratic resilience unit with mandatory humanities representation.

5.’Blue Sky’ Implementation Requirements

Mandatory AI Auditing Standards

All AI systems used in security-sensitive contexts must undergo:

• Bias Testing: Systematic evaluation for discriminatory patterns
• Hallucination Assessment: Testing for false information generation
• Manipulation Vulnerability: Assessment of hostile exploitation risks

Technical Standards for Democratic Protection

• Deepfake Detection: Regular capability assessments and mandatory reporting
• Synthetic Content Labelling: Technical requirements for AI-generated content identification
• Bias Monitoring: Ongoing surveillance of AI system discriminatory impacts

Humanities-Led Oversight

Based on my research demonstrating the critical need for non-technical expertise in AI governance:

• Mandatory Representation: Humanities scholars required on all AI oversight bodies
• Democratic Impact Assessment: Cultural and social analysis required for all AI security measures
• Community Protection: Specific safeguards for marginalised communities affected by AI bias

‘Early Warning System’ and responsive safeguards

To ensure risks are identified in real-world use and addressed rapidly:

• User Reporting Channels: All AI systems must provide clear mechanisms for users to flag concerns such as bias, hallucinations, or suspected manipulation.
• Centralised Monitoring Hubs: Reports must be aggregated and reviewed by independent oversight bodies, enabling cross-system pattern detection and response.
• Escalation Protocols: Issues must be triaged by severity, with high-risk cases (e.g., discriminatory impacts or disinformation campaigns) requiring mandatory regulator notification and corrective action.
• Community Watchdog Participation: Civil society organisations and academic experts must be granted access to anonymised case data, ensuring transparency and democratic oversight.

Conclusion

The NSS 2025's treatment of AI risks fundamentally misunderstands the threat landscape. Whilst focusing on technological capacity and traditional cyber threats, it ignores the immediate vulnerabilities created by AI bias, hallucinations, and deepfake generation.

Hostile actors are already exploiting these weaknesses. China's AI development, as seen in DeepSeek, demonstrates how state-controlled AI can embed systematic biases serving geopolitical interests. Without immediate action addressing bias auditing, hallucination detection, deepfake identification, and humanities-led oversight, the UK faces a critical vulnerability window.

The integration of technical AI expertise with humanities scholarship is not optional—it is essential for democratic survival in the age of generative AI. The NSS must immediately address these gaps or risk fundamental democratic disruption through AI-mediated influence operations that current frameworks cannot detect or counter. Recent research demonstrates that these are not theoretical concerns but immediate threats requiring urgent policy response. The UK's democratic resilience depends on recognising and addressing these vulnerabilities before they are fully exploited.

Selected References

Sonja Bjelobaba, Lorna Waddington et al, ‘Maintaining research integrity in the age of GenAI: an analysis of ethical challenges and recommendations to researchers’, International Journal for Educational Integrity, (June 2025) https://edintegrity.biomedcentral.com/articles/10.1007/s40979-025-00191-w

Joy Buolamwini, Unmasking AI: My Mission to Protect What is Human in a World of Machines (New York, 2024)

Emma McClure & Benjamin Walk, ‘Algorithmic Microaggressions’, Feminist Philosophy Quarterly 8 (3/4). Article 5 (2022)

Sarah Mercer et al, ‘China’s AI Evolution: DeepSeek and National Security’,  https://cetas.turing.ac.uk/publications/chinas-ai-evolution-deepseek-and-national-security

John Naughton, ‘DeepSeek: cheap, powerful Chinese AI for all. What could possibly go wrong?’ Guardian Newspaper, (January 2025)

UNESCO, Guidance for generative AI in education and research (2023)

Lorna Waddington, ‘Navigating Academic Integrity in the Age of GenAI: A Historian's Perspective on Censorship’, https://academicintegrity.org/resources/blog/536-navigating-academic-integrity-in-the-age-of-genai-a-historian-s-perspective-on-censorship (November, 2024)

Lorna Waddington & Richard de Blacquiere-Clarkson, (July 2025) ‘The humanities must have a role in overseeing AI ‘censorship’’, Times Higher Education https://www.timeshighereducation.com/opinion/humanities-must-have-role-overseeing-ai-censorship

Juliette Rowsell, ‘Lecturers fear impact of DeepSeek ‘censorship’ on students’ work’ (Feb 2025) https://www.timeshighereducation.com/news/lecturers-fear-impact-deepseek-censorship-students-work (Feb 2025)

https://www.gov.uk/government/publications/foreign-influence-registration-scheme-introduction (2025)

https://www.gov.uk/government/publications/national-security-strategy-2025-security-for-the-british-people-in-a-dangerous-world#:~:text=The%20purpose%20of%20the%20National,and%20to%20explain%20the%20actions (2025)

18 September 2025

The views expressed here are those of the author alone. They may not represent the views of their affiliated organisations and should not be taken as an official statement or position of the affiliated organisations.