moncefem/cnil-sanctions-2011-2025

CNIL Sanctions Dataset — 2011 → 2025

A typed, statistics-ready dataset of 374 sanctions issued by the French data protection authority (CNIL — Commission Nationale de l'Informatique et des Libertés) between 2011 and 2025, structured into 34 analytical variables for quantitative research on GDPR / French privacy enforcement.

Generated end-to-end (scrape → typed schema → per-row extraction → export) by Gemma Miner — an autonomous text-to-dataset agent that turns any website into a research-grade dataset in minutes.

TL;DR — the big takeaways

• The CNIL became roughly 5× more active between 2014 and 2024 (18 → 86 decisions per year). 2025 is on the same trajectory (83 decisions through Q3).
• The simplified procedure, introduced in 2022, now drives 80 % of yearly volume — it has transformed CNIL enforcement from "a few high-profile cases per year" into a continuous stream of mid-size decisions.
• Volume and money decoupled in 2024. That year had the highest decision count on record (86) but the lowest aggregate disclosed fines in seven years (€55 M). The simplified procedure trades severity for throughput.
• 2025 then exploded to €487 M in disclosed fines — on just two mega-decisions (€325 M and €150 M, both Sept 1, both cookies & consent). Total fines across 2011-2025: €1.14 B, dominated by a handful of adtech/big-tech rulings.
• The single most common breach theme is security of processing (Art. 32 GDPR — 36 % of all sanctions), followed by information / transparency obligations (34 %) and data minimisation (30 %). Cookie/consent appears in only 14 % of decisions but dominates the highest-value decisions.
• Private companies account for 74 % of sanctioned entities; the public sector and associations together make up only ~14 %.

Quick start

📥 Load with 🤗 datasets (click to expand)

from datasets import load_dataset

ds = load_dataset("moncefem/cnil-sanctions-2011-2025", split="train")
print(ds[0])
print(ds.features)

🐼 Load with pandas (no `datasets` install needed)

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/moncefem/cnil-sanctions-2011-2025/cnil_sanctions_analysis.parquet"
)
print(df.shape)         # (374, 34)
print(df.dtypes)

🦆 Load with DuckDB (in-process SQL)

import duckdb

con = duckdb.connect()
con.execute("""
    CREATE VIEW cnil AS
    SELECT * FROM read_parquet(
      'hf://datasets/moncefem/cnil-sanctions-2011-2025/cnil_sanctions_analysis.parquet'
    )
""")
print(con.execute("SELECT n_sanction_year, COUNT(*) AS n, SUM(amount_fine_eur)/1e6 AS fines_meur "
                  "FROM cnil GROUP BY 1 ORDER BY 1").df())

---

Charts at a glance

1. Enforcement volume is growing fast — and the simplified procedure drives it

Yearly counts were stable at ~10–18 from 2011 to 2021, then exploded after the simplified procedure was introduced (loi du 24 janvier 2022). 2024 saw 86 sanctions, of which 80 % were simplified — a regime change, not just a trend line.

🔬 Reproduce these charts

import pandas as pd
import matplotlib.pyplot as plt

df = pd.read_parquet("cnil_sanctions_analysis.parquet")
yearly = df["n_sanction_year"].dropna().astype(int).value_counts().sort_index()
simp = df[df["is_simplified_procedure"] == True]["n_sanction_year"] \
        .dropna().astype(int).value_counts().reindex(yearly.index, fill_value=0)
std  = yearly - simp

fig, ax = plt.subplots(figsize=(10, 5))
ax.bar(yearly.index, std,  label="standard")
ax.bar(yearly.index, simp, bottom=std, label="simplified")
ax.set_title("CNIL sanctions per year, by procedure type")
ax.legend(); plt.show()

2. Volume and money decoupled — and 2025 broke the trend

Two observations the chart makes obvious:

1. 2024 is the inversion year. 86 decisions (highest ever) but only €55 M in disclosed fines (lowest since 2019). The simplified procedure produces mid-size sanctions; the CNIL went broad rather than deep.
2. 2025 is the rebound. €487 M in nine months — almost all of it from two cases on Sept 1 (€325 M + €150 M, both consent / cookies). On a per-decision basis, 2025 is by far the most expensive year on record.

The €1.14 B aggregate hides a fat-tailed distribution: the median fine sits at €10 000, but the top decile drives essentially all of the monetary value. Practically every year past 2019 has one or two decisions ≥ €40 M.

Top 5 fines in the dataset:

Date	Organisation type	Fine	Main breaches
2025-09-01	Société développant plusieurs services en ligne	€325 M	Cookies (information & consent) + unsolicited commercial messaging
2025-09-01	Société de vente en ligne (vêtements, chaussures, accessoires)	€150 M	Cookies (information & consent)
2021-12-31	Services internet (moteur de recherche, plateforme de vidéos)	€150 M	Cookie-refusal UX
2020-12-07	Société de services technologiques	€100 M	Cookies + information + consent + opposition
2022-12-19	Vendeur OS / logiciels / matériels	€60 M	Cookies & trackers consent

Cookies / consent under the ePrivacy directive (Art. 82 LIL) — not GDPR fines per se — produce the most expensive decisions in the dataset.

🔬 Query the top fines yourself

import pandas as pd
df = pd.read_parquet("cnil_sanctions_analysis.parquet")
top = (
    df.dropna(subset=["amount_fine_eur"])
      .sort_values("amount_fine_eur", ascending=False)
      .head(10)
      [["dn_sanction", "organism_type_raw", "amount_fine_eur", "main_breaches_raw"]]
)
print(top.to_string(index=False))

3. Most sanctioned entities are companies

Sector	Share
Private company	74 %
Public administration	10 %
Professional individual	7 %
Association	5 %
Political party	4 %
Other	< 1 %

🔬 Compute the sector / breach-mix crosstab

import pandas as pd
df = pd.read_parquet("cnil_sanctions_analysis.parquet")
breach_cols = [c for c in df.columns if c.startswith("is_breach_")]
ct = df.groupby("cat_sector_group")[breach_cols].mean().round(2)
print(ct.to_string())

4. Security failures dominate the breach mix

Of the eight breach themes in the codebook, the top of the distribution is:

Rank	Theme	n	share
1	Security of processing (Art. 32)	134	36 %
2	Information / transparency	126	34 %
3	Data minimisation	111	30 %
4	Rights of data subjects	83	22 %
5	Lawful basis: consent	69	18 %
6	Cookies / trackers (Art. 82 LIL)	51	14 %
7	Special-category data (Art. 9)	38	10 %
8	Sub-processor obligations	25	7 %

Most decisions involve more than one theme — the median n_breaches per row is 2. Note the inversion between frequency and severity: cookies/consent appears in only 14 % of decisions but drives most of the monetary value (every fine ≥ €60 M in the dataset is a cookies/consent case).

(These flags are detected from each decision's public CNIL summary; null means the summary doesn't address that theme — it is not evidence the breach didn't occur.)

---

Suggested research questions

This dataset is sized for fast iteration on questions like:

1. Has the simplified procedure changed the WHO of enforcement? Compare sector mix in pre-2022 vs post-2022 decisions.
2. Does sanction severity predict the breach mix? Cookie/consent appears to over-index in the €40 M+ tail — is that statistically robust?
3. Are public-sector and private-sector breach profiles different? is_breach_security looks higher in the public set; quantify it.
4. What share of fines are GDPR vs ePrivacy/cookies? Many of the biggest "fines" in the dataset are L.82 cookies cases, not strict GDPR Art. 83. The decision_raw field lets you split.
5. Year-on-year severity: has the median (not just sum) of fines moved? Compute on amount_fine_eur excluding nulls.

🔬 Quick answer sketch for Q1 (pre-2022 vs post-2022 sector mix)

import pandas as pd
df = pd.read_parquet("cnil_sanctions_analysis.parquet")
df["era"] = df["n_sanction_year"].apply(lambda y: "pre_2022" if y < 2022 else "from_2022")
print(
    df.groupby(["era", "cat_sector_group"]).size()
      .unstack(fill_value=0)
      .apply(lambda r: (r / r.sum() * 100).round(1), axis=1)
)

🔬 Quick answer sketch for Q5 (year-on-year severity)

import pandas as pd
df = pd.read_parquet("cnil_sanctions_analysis.parquet")
agg = (
    df.dropna(subset=["amount_fine_eur"])
      .groupby(df["n_sanction_year"].astype(int))["amount_fine_eur"]
      .agg(["count", "median", "mean", "sum"])
)
agg[["median", "mean", "sum"]] = agg[["median", "mean", "sum"]] / 1e3  # €k
print(agg.round(1).to_string())

---

Codebook (34 columns)

All columns are explicitly typed. Where a row lacks evidence in the public summary, the value is null rather than a fabricated default (this is deliberate — see "Limitations" below).

Identification & metadata

Column	Type	Description
id	string	Deterministic content-hash id (stable across runs)
dn_sanction	date	Decision date, ISO YYYY-MM-DD
n_sanction_year	integer	Calendar year of the decision
sanction_date	string	Raw date string as scraped (DD/MM/YYYY)
sanction_year	integer	Same year, kept for cross-checks
organism_name	string	Name of the entity (often anonymised)
organism_type_raw	string	Raw CNIL category label (French)
main_breaches_raw	string	CNIL "principal failings / themes" text (French)
decision_raw	string	CNIL "decision adopted" text (French)
decision_title	string	Full title incl. délibération id where present
decision_url	string	Légifrance URL of the published decision (when published)

Sanction type & severity

Column	Type	Description
has_fine	boolean	Administrative fine present
has_injunction	boolean	Injunction (mise en demeure) accompanies the decision
has_astreinte	boolean	Periodic penalty payment (astreinte) attached
has_warning	boolean	Public warning (avertissement public) issued
amount_fine_eur	float	Fine amount in euros (NaN when none / undisclosed)
cat_fine_bucket	enum	none, under_10k, under_100k, under_1m, over_1m
n_decision_severity_score	integer	1–5 severity proxy synthesised from sanction type + amount

Procedure & sector

Column	Type	Description
cat_procedure_type	enum	simplified, standard (or null)
is_simplified_procedure	boolean	True if the decision used the simplified procedure
cat_sector_group	enum	private_company, public, association, professional_individual, political, other
is_public_sector	boolean	Sanctioned entity is a public body
is_health_related	boolean	Health context (hospital, médecin, mutual, …)
is_digital_platform	boolean	Online / platform context
has_external_decision_link	boolean	Decision links to a published Légifrance text

Breach / theme flags (extracted from each decision's summary)

Column	Theme
is_breach_security	Art. 32 GDPR — security of processing
is_breach_transparency	Information / transparency obligations
is_breach_consent	Lawful basis: consent
is_breach_data_rights	Rights of data subjects (access, erasure, opposition)
is_breach_cookies	Cookies / trackers (Art. 82 LIL)
is_breach_minimization	Data minimisation
is_breach_processor	Sub-processor / contractual obligations
is_involves_sensitive_data	Special-category data (Art. 9 GDPR)
n_breaches	Integer count of distinct breach themes in the decision

---

How this dataset was built

This file was produced by Gemma Miner in a single agent run — about 9 minutes end-to-end on openrouter/google/gemini-3.1-flash-lite:

1. Harvest — agent fetched the CNIL listing page, auto-generated a regex extractor over the HTML table, and queued 374 rows (one per sanction).
2. Codebook design — an LLM proposed ~30 typed variables matching the analytical brief (sector, procedure, breach themes, fine bucket, severity score).
3. Per-row extraction — for each sanction, an LLM read the French text and emitted a JSON object conforming to the codebook; the system then deterministically coerced values to the declared types (dates → ISO, numbers → float, enums snapped to nearest valid value, ambiguous booleans → null).
4. Export — parquet + CSV + this card + charts.

No fine-tuning. No labelled training data. Reproducible from the upstream URL on any day.

🔬 Rebuild this dataset from scratch

pip install gemma-miner          # from https://github.com/moncifem/gemma-miner
export OPENROUTER_API_KEY=...    # any OpenAI-compatible provider works
gemma42                          # drops into the REPL

# in the REPL:
> Build me a statistics-ready dataset of CNIL sanctions from
  https://www.cnil.fr/fr/les-sanctions-prononcees-par-la-cnil with sector,
  procedure type, fine amount in EUR, and breach-theme flags.

---

Limitations & honest caveats

• CNIL anonymises most entities by category ("OPÉRATEUR DE TÉLÉPHONIE MOBILE") rather than name. organism_name is therefore often generic. For the named, headline-grade decisions, follow decision_url to Légifrance for the authoritative text.
• LLM-derived columns were extracted from the public CNIL summary (≤ 1 KB of French text per row), not from the full deliberation. A true for is_breach_security is high-precision; a null / false means the summary doesn't discuss it — not that no security breach occurred. This is deliberate: the dataset preserves the null-vs-false distinction throughout.
• Fine parsing is robust on the majority but can misattribute on rows combining a fine + a liquidation d'astreinte; the raw decision_raw field is preserved so you can re-parse if needed.
• Decision URLs are present on ~75 % of rows. Simplified-procedure decisions are usually unlinked.
• Sample size for some breach themes is small (n < 20). Use Wilson or Jeffreys CIs rather than normal-approximation intervals when comparing rates across years or sectors.
• The CNIL publishes only its own sanctions — judicial fines and EDPB cross-border decisions where CNIL was supporting authority are not included.

---

Citation

@misc{elmouden_cnil_sanctions_2025,
  title  = {CNIL Sanctions 2011-2025},
  author = {EL-Mouden, Moncif},
  year   = {2025},
  note   = {Generated by Gemma Miner from https://www.cnil.fr/fr/les-sanctions-prononcees-par-la-cnil},
  url    = {https://huggingface.co/datasets/moncefem/cnil-sanctions-2011-2025},
}

@software{elmouden_gemma_miner_2025,
  title  = {Gemma Miner: an autonomous text-to-dataset agent},
  author = {EL-Mouden, Moncif and contributors},
  year   = {2025},
  url    = {https://github.com/moncifem/gemma-miner},
}

Underlying decisions are published by the CNIL on https://www.cnil.fr/fr/les-sanctions-prononcees-par-la-cnil and on Légifrance; consult those sources for the authoritative text.

Author & links

• 👤 Moncif EL-Mouden — 🤗 huggingface.co/moncefem
• 🤖 Gemma Miner (the generator) — https://github.com/moncifem/gemma-miner
• 🇫🇷 Source — https://www.cnil.fr/fr/les-sanctions-prononcees-par-la-cnil

License

Apache License 2.0.

Please attribute:

• the CNIL as the source of the underlying decisions, and
• Gemma Miner (https://github.com/moncifem/gemma-miner) as the dataset generator.