14.3 Problem Solving, Technical Troubleshooting & Artificial Intelligence (DigComp 5.1-5.4)

14.3 Problem Solving, Technical Troubleshooting & Artificial Intelligence (DigComp 5.1-5.4)

Official Reference: European Commission Joint Research Centre (JRC) DigComp 2.2. Area 5 (Problem Solving) covers Competences 5.1 (Solving technical problems), 5.2 (Identifying needs and technological responses), 5.3 (Creatively using digital technologies), and 5.4 (Identifying digital competence gaps). The 2022 update also added more than 250 new examples, including an appendix on citizens interacting with AI systems, which runs across all five areas rather than belonging to a single competence.

Under Notice EPSO/AD/427/26, the Digital Skills module requires administrators to demonstrate cognitive adaptability, technical problem-solving acumen, and modern technological literacy. An AD5 administrator must not only troubleshoot day-to-day digital roadblocks but must also critically navigate emerging digital technologies—most notably Artificial Intelligence (AI). With the formal adoption of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689), European civil servants are called upon to act as informed guardians of fundamental rights, algorithmic transparency, and ethical public administration.

---

Competence 5.1: Solving Technical Problems — Systematic Troubleshooting

When digital tools malfunction, unstructured or haphazard intervention frequently exacerbates errors. Public administrators must employ an empirical, step-by-step diagnostic triage protocol to isolate failures across distinct functional layers.

The Empirical Diagnostic Protocol

1. Define the Scope of the Failure: Establish whether the issue is isolated to a single local document, a specific user account, a particular workstation, or the wider institutional network. If a colleague sitting nearby can access the intranet while your machine cannot, the issue is local; if all users lose access simultaneously, the failure is centralized (e.g., an institutional Single Sign-On outage).
2. Physical and Hardware Layer Verification: Inspect physical connections (Ethernet cables, docking station USB-C interfaces, power supplies). Verify hardware indicator lights and confirm that wireless switches or airplane mode toggles are not inadvertently activated.
3. Power Cycling and Volatile Memory Clearance: A standard restart ("reboot") resolves a substantial percentage of software errors. Rebooting clears volatile Random Access Memory (RAM), terminates rogue orphan background processes, flushes memory leaks, and resets deadlocked operating system thread states.
4. Application and Process Isolation: When an application freezes, administrators open the Task Manager (Windows) or Activity Monitor (macOS) to identify runaway processes exhibiting abnormal CPU or memory consumption. Forcing a process termination is preferred over a hard system power cut, which can corrupt open disk files.
5. Safe Mode and Clean Boot Diagnostics: If an operating system fails to boot or exhibits unstable behavior, booting into Safe Mode loads only the minimal core operating system drivers and kernel services, disabling third-party startup programs and shell extensions. If the fault disappears in Safe Mode, a recently installed third-party driver or application is the root cause.
6. Cache and Temporary Data Remediation: Corrupted browser caches, expired authentication cookies, and orphaned temporary files frequently cause rendering bugs or login loops in web applications (e.g., Ares or EU Learn). Performing a hard refresh (Ctrl + F5 or Cmd + Shift + R) or clearing browser site cookies resets session state without requiring software reinstallation.

Network Stack Troubleshooting

When network connectivity fails, administrators diagnose the network stack hierarchically from local interfaces to remote servers:

Competence 5.2: Identifying Needs and Technological Responses

Competence 5.2 means assessing a need, then choosing, evaluating and adjusting the right digital tool, including personalising your own environment (accessibility settings, keyboard layouts, notifications). Test items typically describe a need and ask for the most appropriate response. Screen the options against five questions:

Competence 5.3: Creatively Using Digital Technologies

Competence 5.3 means using digital tools to create knowledge and to innovate processes and products, alone or with others. Examples include automating a recurring report instead of rebuilding it each month (see Section 13.2), prototyping a dashboard to test a policy idea, or running an online co-creation session with stakeholders.

Competence 5.4: Identifying Digital Competence Gaps

Competence 5.4 means recognising where your own digital competence needs improving, helping others with theirs, seeking self-development opportunities and keeping up with digital change. DigComp describes progression in eight proficiency levels grouped in pairs: Foundation (1–2), Intermediate (3–4), Advanced (5–6) and Highly specialised (7–8). The EU's free Europass digital skills self-assessment is built on DigComp. The EU's Digital Decade target is for at least 80% of adults to have basic digital skills by 2030.

AI Literacy Across DigComp 2.2

The DigComp 2.2 update added examples on Artificial Intelligence (AI) and automated systems across all five areas: recognising when you interact with AI, evaluating AI outputs critically (Area 1), and understanding data use and rights (Area 4). The AI Act adds a legal duty of AI literacy for providers and deployers of AI systems (Article 4), applicable since 2 February 2025.

How Artificial Intelligence Operates: From Algorithms to Neural Networks

Unlike traditional deterministic software—which executes explicit "if-then" rules scripted by human software engineers—modern AI systems learn mathematical representations from empirical training data:

• Machine Learning (ML): Algorithms optimize a mathematical objective function to uncover patterns in data:
  • Supervised Learning: Models train on labeled input-output pairs (e.g., thousands of historical trade declarations labeled "compliant" or "fraudulent"), learning a mapping function to predict labels on unseen data.
  • Unsupervised Learning: Models cluster unlabeled data into natural groupings or discover anomalous outliers without predefined target categories.
  • Reinforcement Learning: Autonomous agents learn optimal behavior policies through trial-and-error interactions with an environment, receiving scalar rewards or penalties.
• Artificial Neural Networks and Deep Learning: Modeled loosely on biological neural pathways, deep neural networks comprise an input layer, multiple hidden processing layers, and an output layer. Nodes (artificial neurons) apply weighted linear transformations followed by non-linear activation functions. During training, the network adjusts millions or billions of internal parameters (weights and biases) using backpropagation and gradient descent to minimize output error.
• Inference vs. Training: Training is a computationally intensive, one-off phase requiring clusters of high-performance GPUs to process massive datasets. Inference is the operational phase where the trained, frozen model evaluates a new user input (a prompt or image) and computes a prediction or output in real time.

Generative AI and Large Language Models (LLMs)

Generative AI refers to models capable of synthesizing novel text, imagery, code, or audio based on user prompts. Large Language Models (LLMs) (such as GPT-4, Gemini, or open-weight models like Llama) are built upon the Transformer architecture, utilizing self-attention mechanisms to process input sequences in parallel, computing contextual relationships between distant words.

• Tokenization and Next-Token Prediction: LLMs do not read words as conceptual symbols; they break text into numerical fragments called tokens. At its core, an LLM functions as a sophisticated probabilistic engine that calculates a probability distribution over vocabulary tokens to predict the most likely next token in a sequence.
• Administrative Capabilities: Within European institutions, authorized enterprise LLMs assist officials by summarizing multi-hundred-page public consultation filings, translating technical administrative dossiers across official EU languages, drafting initial routine correspondence, and formatting raw tables into structured briefing memos.

Fundamental Cognitive Limitations of Generative AI

Administrators must never confuse linguistic fluency with factual truth or genuine cognitive comprehension:

• Hallucinations: LLMs produce text based on statistical plausibility, not factual reality. Consequently, they can generate completely fabricated legal precedents, non-existent CJEU judgments, fictitious treaty articles, or false biographical details with the same confident, authoritative tone as factual statements.
• The "Stochastic Parrot" Reality: LLMs lack an internal world model, moral agency, or symbolic semantic understanding. They recognize syntactic correlations and linguistic patterns, not underlying physical, legal, or logical truths.
• Absence of Real-Time Factual Grounding: Standard foundation models are bounded by a fixed knowledge cutoff date. They possess no awareness of legislative developments, geopolitical events, or administrative appointments occurring after their training, unless augmented via Retrieval-Augmented Generation (RAG) systems connected to live document databases.
• Context Window Limits: Models possess a finite context window (measured in thousands of tokens). When processing massive legislative dossiers, information situated in the middle of long prompts can suffer from "attention degradation" (the "lost in the middle" effect).

Algorithmic Bias, Fairness, and Feedback Loops

Automated decision systems reflect and amplify the biases present in their training corpora:

• Historical Bias in Training Data: If historical administrative recruitment or loan approval data reflects societal discrimination, machine learning models treat those historical inequalities as predictive rules, systematically disadvantaging specific protected demographic groups.
• Representation Bias: Datasets heavily skewed toward dominant official languages or major Member States lead models to perform poorly when processing minority languages or specialized regional administrative contexts.
• Feedback Loops: Deploying biased algorithmic systems in public administration creates self-fulfilling prophecies, where automated risk scoring increases scrutiny on particular demographics, generating more enforcement actions that falsely "validate" the model's original biased predictions.

The EU Artificial Intelligence Act (Regulation (EU) 2024/1689)

The EU AI Act is the first comprehensive horizontal law on artificial intelligence. It entered into force on 1 August 2024 and applies in stages. Prohibited practices and the AI-literacy duty apply from 2 February 2025, and obligations for general-purpose AI models from 2 August 2025. Most other obligations are phased in from 2026, but check the current timetable, because in November 2025 the Commission proposed to delay parts of the high-risk regime. The Act uses a graduated, risk-based architecture:

Mandatory Requirements for High-Risk AI Systems

Before deploying a high-risk AI system within the EU, providers and administrative deployers must ensure:

1. Continuous Risk Management System: Structured risk identification, estimation, and mitigation throughout the entire system lifecycle.
2. High-Quality Data Governance: Training, validation, and testing datasets must be relevant, representative, free of systematic errors, and vetted for potential biases.
3. Detailed Technical Documentation and Automatic Logging: Maintenance of extensive records enabling supervisory authorities to verify regulatory compliance, with automated event logging tracking operational traceability.
4. Transparency and User Instructions: Clear explanations of capabilities, system limitations, and accuracy metrics provided to institutional deployers.
5. Human Oversight (Human-in-the-Loop): System architecture must enable natural persons to oversee operation, prevent automation bias, override system outputs, and immediately halt execution via an operational "stop" button.
6. Accuracy, Robustness, and Cybersecurity: High resilience against errors, adversarial attacks, and data poisoning.
7. Fundamental Rights Impact Assessment (FRIA): Public sector bodies must conduct a formal FRIA assessing the specific impact on affected individuals prior to putting a high-risk AI system into service.

EU Institutional Governance: The European AI Office

Governance of the AI Act is centered in the European AI Office, established within the European Commission. The AI Office enforces rules on general-purpose AI (GPAI) foundation models, monitors technological advancements, fosters innovation via regulatory sandboxes, and coordinates enforcement alongside the European Artificial Intelligence Board (EAIB), which comprises representatives from each Member State.