Module 08Frontiers & Integration

Learning Outcomes

# Module 8: Frontiers and Integration

Estimated time: 3–4 hours

Type: Conceptual + synthesis (workbook + portfolio builder)

Prerequisites: Modules 1–7 (or at least Modules 1 and 7, plus your chosen application modules)

What you'll produce: A complete portfolio prototype: task, baseline, experiment, evidence, decision, disclosure

1. Local and Open-Weight Models

Throughout this course, you have used AI through chat interfaces — models running on servers controlled by companies. There is an alternative: running models on your own hardware.

What are open-weight models?

An open-weight model is one whose trained parameters (the "weights" — the numerical result of training) are publicly available for download. Anyone with sufficient hardware can run the model. Examples include:

• Llama (Meta)
• Mistral (Mistral AI)
• Qwen (Alibaba)
• DeepSeek (DeepSeek AI)
• OLMo (Allen Institute for AI)

These are distinct from "open source" in the traditional sense — the training data and code may not be public — but the model itself can be downloaded, inspected, modified, and run locally.

Why run models locally?

1. Data stays on your device. No text you process is sent to an external server. This matters for confidential, sensitive, or restricted data (interview transcripts, administrative records, unpublished drafts).

1. Reproducibility. You control the model version. A proprietary model accessed through an API can be updated, modified, or deprecated without notice. A file on your hard drive does not change.

1. Cost. No per-token charges. The cost is your hardware and electricity.

1. Customisation. You can fine-tune an open-weight model on your own data, adapting it to your domain, terminology, or task.

The trade-offs

• Hardware requirements. Running a capable model requires a reasonably powerful computer, ideally with a dedicated GPU. Smaller models (1–8 billion parameters) run on consumer laptops. Larger models (70B+) need specialised hardware or cloud GPU rental.

• Capability gap. The largest open-weight models are competitive with proprietary models on many tasks, but the most capable proprietary models (GPT-5, Claude 4, Gemini) still lead on complex reasoning, long-context tasks, and multilingual performance.

• Setup overhead. Installing and running a local model requires some technical comfort — downloading model files, configuring software (LM Studio, Ollama, vLLM), and managing storage. It is not as simple as opening a website.

When local models make sense for research

• Your data cannot leave your device (ethics, GDPR, institutional rules)
• You need reproducible results over months or years
• You are processing a large volume of text and API costs would be prohibitive
• You want to fine-tune a model on domain-specific data

When proprietary models make sense

• You need the highest available capability for a complex task
• Your data is not sensitive and can be transmitted
• You are doing exploratory work where speed of iteration matters more than long-term reproducibility
• You lack the hardware or technical comfort to run models locally

The course position: Neither choice is "better." The choice should be justified in your disclosure, with the trade-offs stated. "Convenience" is not a methodological justification — but "the task required multilingual capability that available open-weight models do not provide at sufficient quality" is.

2. Multimodal AI

Everything in Modules 1–7 has been about text. But AI systems increasingly handle images, audio, and video as well.

What multimodal models can do

• Image analysis: Describe what is in a photograph, extract text from a scanned document (OCR), identify objects, read charts and graphs
• Image generation: Create images from text descriptions (DALL-E, Midjourney, Stable Diffusion)
• Audio transcription and analysis: Convert speech to text, identify speakers, detect emotion or sentiment in voice
• Video understanding: Summarise video content, detect scene changes, track objects across frames

Relevance to social science research

Multimodal AI opens research possibilities that were previously impractical at scale:

• Visual content analysis: Classify images in social media posts, news photographs, or political advertisements — not just their captions
• Protest event detection: Scan video footage for crowd size estimation, sign text extraction, or police-protester interaction coding
• Archival research: Transcribe and translate handwritten documents, historical audio recordings, or microfilm scans
• Accessibility: Generate image descriptions for visually impaired researchers or study participants

The same principles apply

Multimodal models are still prediction engines. They can miscaption an image, hallucinate text that is not in a document, or misidentify an object. Every validation principle from Module 7 applies: verify against human judgment, test on known cases, check whether errors are systematic, and disclose what was used.

The boundaries

• Consent and privacy: Processing images or audio of people raises privacy concerns that text processing does not. Do not upload identifiable images or recordings without consent.
• Copyright: Generated images may reproduce elements of copyrighted training data. Using AI-generated images in publication may raise rights issues.
• Deepfake concern: The same technology that analyses media can synthesise it. Researchers using multimodal AI should be aware of how their work could be mischaracterised or misused.

3. Ethics, Governance, and Institutional Boundaries

AI use in research operates within institutional, legal, and ethical frameworks. You are responsible for knowing the ones that apply to your work.

Key frameworks

European Research Area (ERA). Living Guidelines on the Responsible Use of Generative AI in Research (updated May 2026). The factsheet version covers: researchers remain accountable; AI use must be transparent; peer review cannot be delegated; data protection applies. Read it — it is short and directly applicable.

University of Zurich. Recommendations on the Use of Generative Artificial Intelligence (September 2023). Implements seven guiding principles: responsibility, transparency, data protection, quality assurance, fairness, sustainability, and human oversight. Available at uzh.ch/en/explore/basics/ai/recommendations.html.

Swiss National Science Foundation (SNSF). Statement on AI in funding applications (February 2024). The key paragraph: applicants are responsible for content produced with AI assistance; reviewers cannot use AI to evaluate proposals; confidentiality of third-party data must be preserved.

Journal policies. American Journal of Political Science (AJPS) AI Policy (June 2025) and American Political Science Review (APSR) author instructions: AI use must be disclosed; AI cannot be an author; AI-generated references must be verified; reviewers cannot upload manuscripts to AI tools.

The principles that cut across all frameworks

1. You are accountable. A tool cannot be a co-author, cannot be cited as a source of fact, and cannot absorb responsibility for an error.
2. Disclose. Name the tool, the purpose, and the location in the work. The disclosure must be specific enough that a reviewer can evaluate it.
3. Protect data. Do not upload material you lack the authority to share: unpublished co-authored drafts, interview transcripts, administrative data, identifiable records.
4. Verify. AI-generated references, claims, and analyses must be checked by a human before submission.
5. Preserve human judgment. Peer review, grading, and final editorial decisions must be made by people.

4. Building Your Workflow: Integration

The eight modules of this course are designed to be combined. A real research workflow typically involves several of them:

• Module 2 (Writing) + Module 7 (Validation): Draft a paper section with AI assistance, then verify claims, citations, and argument structure against the evidence framework.
• Module 3 (Literature) + Module 4 (Annotation): Search for and screen papers, then classify them by method, finding, or relevance using a codebook.
• Module 4 (Annotation) + Module 5 (Code/Data): Generate labels with an LLM, then analyse the labelled dataset using AI-assisted code — and verify every step.
• Module 6 (Agents) + Module 7 (Validation): Design an agent workflow, run it, and audit the output against the seven validation checks.
• Module 3 (Literature) + Module 6 (Agents): Build a literature-search agent with persistent instructions, then inspect the screening transcript.

The portfolio asks you to integrate at least two modules into a coherent workflow for a task from your own research.

The integration checklist

Before you finalise your workflow, verify:

• [ ] Every AI-assisted step has a named human check
• [ ] The evidence for each claim is traceable to a verified source or a documented computation
• [ ] The disclosure names every tool, model, date, and purpose
• [ ] The four questions are answerable for every delegated task
• [ ] The validation plan uses checks from the menu (Module 7) appropriate to the claims
• [ ] Data boundaries are respected — nothing uploaded that you lack authority to share
• [ ] The workflow can be described in a methods section a reviewer could evaluate

5. The Portfolio: Your Final Integration

The portfolio is the course's capstone. It answers one question:

Your answer takes one of four forms:

• Integrate the use into your workflow
• Integrate with limits — specify the conditions, checks, and boundaries
• Continue experimenting — the evidence is promising but not yet sufficient
• Do not integrate — the evidence does not support adoption for this task

All four conclusions can be equally strong. The grade depends on the quality of the evidence and reasoning, not on whether you adopted AI.

Portfolio structure

1. Research task and baseline (~1 page): What is the task? How do you currently do it? What does success and failure mean for this task?

1. Proposed AI use (~1 page): What exactly would the AI do? What judgment is being delegated and what is retained? What data, ethics, or institutional constraints apply?

1. Experiment (~2 pages): What claim did you test? What was your comparison? What tool, model, date, instructions, and settings? What did you predict before running? What evidence did you preserve?

1. Evidence and failures (~1 page): What did the experiment show? What checks did you run? What failed, was uncertain, or was limited? Why do the failures matter for the research claim?

1. Decision (~1 page): "On the evidence I have, I will…" and "I would change this decision if…" Be specific about the condition that would change your mind.

1. Revised workflow and disclosure (~1 page): The resulting workflow in any readable form, with who/what performs each step, what check applies, what record remains, and what happens when a check fails. Plus a disclosure paragraph suitable for a methods section.

What earns credit

• A bounded task with a meaningful baseline
• A credible comparison or experiment
• External evidence — not just the model's output
• Attention to failure, variation, and boundary cases
• A decision supported by evidence, with a stated condition for changing it
• Transparent documentation

What does not earn credit

• Using many tools rather than one well-evaluated one
• Long or technically complex workflows
• Polished model output presented as evidence
• The fact of adoption (integrate is not better than do not integrate)
• Claiming a human was "in the loop" without showing what the human could reject or change

6. The Four Questions, One Last Time

Apply them to your completed portfolio:

1. What is being delegated? In your final workflow, name every task and every decision — and who or what performs each.
2. What would make the result believable? For each delegated step, name the external evidence that supports it. If you cannot name it, you have not verified it.
3. What could be lost or flattened? What variation, minority perspective, ambiguity, or tacit knowledge does your AI-assisted workflow handle differently — or not at all — compared to your baseline?
4. What must remain visible or human? What decisions, checks, and records survive outside the model's context window? If a reviewer cannot inspect them from your disclosure, they do not survive.

7. Summary

• Open-weight models offer data locality, reproducibility, and cost advantages — at the cost of hardware requirements and some capability gap. The choice requires justification.
• Multimodal AI extends the same principles (and the same validation requirements) to images, audio, and video.
• Institutional frameworks (ERA, UZH, SNSF, journals) converge on five principles: accountability, disclosure, data protection, verification, and human judgment.
• Integration means combining modules into a coherent workflow where every AI-assisted step has a named human check and every claim traces to verifiable evidence.
• The portfolio is your answer to the course's governing question. The quality of your evidence and reasoning determines the answer's strength — not whether you adopted AI.

8. The Course, In One Paragraph

AI can make research faster, larger, and more ambitious. It can also make it sloppier, less transparent, and harder to reproduce — and the same tool can do both, depending on how it is used. The difference is not the model, the interface, or the prompt. It is the researcher's willingness to ask what is being delegated, to demand evidence from outside the system, to watch for what is lost or flattened, and to keep visible what must remain human. That is the method. The tool is just the tool.