Vibe coding and generative AI

Autonomy, sensitive data and sustainability in research settings

María Cristina Nanton

September 7, 2026

About me

María Cristina Nanton

Developer
Health Ministry @ Buenos Aires | The Global Health Network

Data science in the public sector
Data-driven projects in a global health research network

Vibe coding…

… vibe coding?

There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. […]

I ask for the dumbest things like “decrease the padding on the sidebar by half” because I’m too lazy to find it. I “Accept All” always, I don’t read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. […]

I’m building a project or webapp, but it’s not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

— Andrej Karpathy, Twitter, February 2025 (emphasis mine)

1 - “We built something and it works”

AI within everyone’s reach

Generative AI platforms that give access to advanced language models


Claude - Anthropic
(Claude Opus 5, Sonnet 5, Fable 5.1)


ChatGPT - OpenAI
(GPT-5.5 Instant, GPT-5.6 Sol)


Copilot - Microsoft
(GPT-5.6, Claude Opus 5)

Conversational interface + specific ways to integrate them into code-based projects
Claude Code | Codex | GitHub Copilot | Etc

AI within everyone’s reach

And also open models: downloadable models you can run on your own infrastructure (a laptop, an institute server, an already-approved cloud)

Who publishes them

  • US: Llama 4 (Meta) · Gemma 4 (Google) · gpt-oss (OpenAI)
  • Europe: Mistral Small 4 (Mistral AI)
  • China: Qwen 3.x (Alibaba) · DeepSeek V4 · GLM-5.1 (Z.ai) · Kimi K2 (Moonshot)

What you need to run them

  • Small ones: a laptop
  • Frontier ones: servers with several GPUs

How they run

  • On your own computer or on a server

What you gain

  • Privacy: the data doesn’t travel to a provider
  • No changing terms of service
  • Predictable cost: hardware, not tokens

If it runs on your own infrastructure, the data never leaves the institution.

Platforms: describe the app and it appears (it even gets published!)

  • A chat and a live preview: you describe what you want in plain language and the platform generates the frontend and, depending on the platform, also the backend, database and login, and publishes it in one click
  • Nothing to install and nothing to program: the result is a usable product, accessible via a public URL
  • Examples: Lovable | Base44 | Bolt.new | Replit |Claude (artifacts) | ….

What kind of products come out of this?

Products such as

  • Institutional websites and landing pages
  • Forms with validation and login
  • Dashboards
  • Prototypes to show a committee or funder

Some examples I’ve seen

  • Digital publication prototype for an academic organization
  • Student training site with a built-in LLM to answer questions
  • Tool for managing students and research projects

In short: autonomy and speed

  • Prototypes in days or hours
  • Faster iterations
  • Fewer (or no) intermediaries

Between prototype and production (1/2)

Some questions that often come up:

  • How do I integrate this into my institution’s platform?
  • How do I integrate specific tools: institutional login, scheduling, other platforms?
  • How do I pay for it once I outgrow the free plan, or when the terms of service change?

Between prototype and production (2/2)

  • How do I know that the data my users enter (or that I entered) is secure?
  • How hard would it be to build this from scratch in my organization?

2 - Making the most of autonomy without creating problems (for ourselves)

Working with sensitive data

The regulatory floor

  • By region: from very detailed frameworks to contexts with general regulation but no AI-specific guidance
  • Institutional: it may or may not exist
  • The absence of a rule that explicitly forbids something doesn’t make it safe or ethical

Working with sensitive data

Your own floor — three questions before you start

  • Could someone’s identity be reconstructed from the data I’m using?
  • Who will access the data, and in what environment?
  • Are there expectations about how the data will be handled?

Old problems, amplified: a Google Doc stays where you left it; a prompt can end up as training data, and an app generated with Lovable can publish its database to the internet.

A recent example (1/2): AI credits for rare disease research

Anthropic, AI for Science program — thematic call (Jul 2026)

  • Up to USD 50,000 in Claude credits for six months; applications open until August 2, 2026
  • Two tracks: basic science and early-stage biotechs (starting doses, biomarkers, regulatory documentation)
  • The rationale: ~400 million people living with one of more than 7,000 rare diseases; small populations, scattered data, shared mechanisms that are hard to detect

Source: anthropic.com/news/rare-disease-research-grants

(2/2): “…need a safer framework for patient data”

Julien Gagneur (Technical University of Munich)

  • The terms are not fit for clinical or genomic patient data: by accepting the credits, you guarantee that Anthropic may collect, analyze and train models on all inputs and outputs, under a perpetual, irrevocable license
  • Having permission to analyze data is not the same as having permission to transfer it to a third party: the governance burden and legal risk shift to the researcher and their institution

(2/2): “…need a safer framework for patient data”

  • An agent connected to your local environment reads files, runs commands and sends outputs using your credentials: data can leave without anyone uploading it, especially if code is approved without being reviewed
  • Proposal: a protected research mode — no training on or licensing of patient data, processing inside the approved environment, auditable controls and a clear data processing agreement. Open code, methods and evaluations

Source: medium.com/@gagneur — “…need a safer framework for patient data”

Shadow AI: what are we talking about?

Shadow AI is the unsanctioned use of any artificial intelligence (AI) tool or application by employees or end users without the formal approval or oversight of the information technology (IT) department.

— IBM, “What is shadow AI?” (2024)

New risks: what data gets sent to commercial platforms and models, both voluntarily and involuntarily, and how a model’s outputs get absorbed into institutional assets.

Shadow AI: the Samsung case (note: 2023)

  • 2022: ChatGPT launches
  • March 2023: Samsung lifts an earlier restriction and allows ChatGPT
  • March–April 2023: one engineer pastes the full source code of a program from their semiconductor database to ask for a fix; another does the same with code for faulty equipment; a third uploads an entire meeting to have it write up the minutes
  • Emergency measure: a cap on prompt length
  • May: the ban is back, now covering every generative AI tool. The reason, according to an internal memo: whatever is sent is stored on external servers, is hard to retrieve or delete, and could end up exposed to other users

Nobody meant to leak anything: each of them was solving a work problem with the tool at hand

Sources: The Register, Apr 6, 2023 · Gizmodo, Apr 2023 · Bloomberg, May 2, 2023

Why Shadow AI shows up in 2026

  • Organizations without a dedicated IT team (or whose IT team has no capacity for groups interested in AI)
  • No funding for institutional AI
  • Lack of technical know-how to deploy AI within an institutional governance framework

Where data can leak

What data?

Confidential

  • Credentials and keys: in a .env file, in the code, in a screenshot
  • Databases and spreadsheets with personal data (patients, participants, staff)
  • Medical records, results, samples

Internal

  • Details about my projects or organization
  • Internal code, protocols and documents

Through where?

  • The prompt: “let me paste the table so you can help me”

  • The shared folder: an AI with access to a folder reads everything in it

  • The terms of service: the provider stores inputs and outputs and, depending on the plan, trains on them

  • What the AI generates: apps with a public URL and a database with no access rules; repos with credentials inside

  • Integrations: connectors to Drive, email or calendar; browser extensions

Examples of university policies

Harvard

  • Nothing confidential in public tools: research data, medical information, personnel records (not even for meeting minutes!)
  • With default settings, whatever you share is not private
  • Approved tools, licensed by the university, specifically for sensitive data

Penn State

  • De-identification is not foolproof: AI can infer identities
  • Metadata identifies too: file names, geolocation, dates
  • Nothing identifiable without IRB approval
  • Researchers must obtain participants’ consent if generative AI is used to analyze or process their data

Sheffield

Minimum criteria before using AI with sensitive data (1/2)

  1. What kind of data is it?
    Classify it before opening the chat: public, internal or confidential
  2. With which tool?
    One approved by your institution, under a contract that prohibits training on your data
  3. Who authorized it?
    Patient or participant data → ethics committee + data agreement, before you start

Minimum criteria before using AI with sensitive data (2/2)

  1. Can it be done with less?
    Anonymized or synthetic samples; no identifying metadata
  2. What else can the tool do?
    If it reads your data, consumes external content and can send things out → take away at least one of those
  3. Who reviews what comes out?
    Before publishing: who can read the database? Are any keys exposed?

Sources:
“Cuidado con lo que le confIAs” (in Spanish) — AEPD, 2026
“The lethal trifecta for AI agents” — S. Willison, 2025
“Common security risks in vibe-coded apps” — Wiz Research, 2025

Reflections and key questions

If what you do with AI will have real impact

Any digital asset that affects people, projects or products in a research setting calls for assessing:

  • Sustainability in the medium (or short!) term
  • Security, sensitivity and institutional risk
  • Your own knowledge gaps about the problem you’re tackling with AI

But also..

Don’t be a “meat proxy”

By all means, prompt AI. But don’t just relay the output. Read it, understand it, validate it, and then write a response in your own words (a decent certificate that you’ve done the prior steps). Making that effort is value you can add.

— Niklas Gruhn, “Don’t be a meat proxy”, August 3, 2026

On this, 3 key questions

What value are we expected to bring
to the projects we work on?

What is worth investing time in?

Where data is stored is a decision

Is it your decision?

Thank you!

María Cristina Nanton
mcnanton@gmail.com