Calibrated noise, or the art of disappearing into a statistic.
The law says anonymized data ceases to be personal data. To get there it demands an irreversible procedure. No known technique removes residual risk, so the real question is what you would have to be able to prove.
This is not a settled legal opinion. It distinguishes throughout between what the statute says, what the technical evidence shows and what is the author's own view, and every source is linked so it can be checked without intermediaries.
Each dot is a person. The orange dot is your client. The circle marks how many people they blend in with.
Blends in with 681 people
Within the conservative reference this report uses.
Maximum shift in the relative probabilities: ×2.7
There is no legal epsilon threshold in Chile and no ruling from the Agency, and the literature offers no universally accepted range: the scale above is the author's own. The factor does not measure the probability that someone will be identified: it bounds how far the scales can tip between hypotheses.29
of 1.5 million people identified with four location and time points in the reference study1
maximum fine for a very serious infringement, checked against the consolidated text of the law7
days until December 1, 2026, the effective date under the current statutory calendar
Agency guidelines on anonymization, because its Board of Directors has not yet been constituted
The Agency still does not exist
Article 30, which creates the Agency as an autonomous public law corporation, belongs to the amendments to Law No. 19,628 (Ley N° 19.628) that, under the first transitional article of Law No. 21,719, only enter into force on December 1, 2026: until that date the Agency has no legal existence, and what the statute brought forward was the appointment of its first Board of Directors and a limited early exercise of functions by that board once seated. On May 19, 2026 the Senate floor rejected the Executive's proposal to seat the Board of Directors of the Personal Data Protection Agency (Joselyn Biermann, Roberto Godoy and Matías Larraguibel), for failing to reach the quorum of two thirds of sitting senators.16 According to the specialist press, the proposal obtained 19 votes in favor and 12 against, short of the 33 the quorum required.17 The deadline in the fourth transitional article, which ordered the first appointment to be made 6 months before the law takes effect, expired with no board members appointed.
Supreme Decree No. 662 of June 13, 2025, of the Ministry of Finance, approving the regulation on the infringement prevention model, entered the Comptroller General's review process on August 28, 2025 and has yet to be published in the Official Gazette.18 During August 2026 the Government has said it is weighing whether to postpone the entry into force through a short amending statute.19
A sectoral technical antecedent does exist, earlier than and unrelated to the Agency: Technical Standard No. 241 (Norma Técnica N° 241) of the Ministry of Health, approved by Exempt Decree No. 23 of April 9, 2025, of the Undersecretariat of Public Health, and published on April 30, 2025, which governs the anonymization of structured databases for publication as open data in the health sector. It requires k-anonymity and l-diversity of at least 2 for quasi-identifiers, validates the result with the motivated intruder test, and warns that its own techniques do not keep their properties permanently and must be reassessed periodically. It does not prescribe differential privacy.26
That postponement has not been approved. The date in force remains December 1, 2026 and it is the one to plan against. Anyone deferring preparation on a bet that an extension will come is taking a risk with no legal support.
The deeper consequence is a different one: without a Board there are no interpretive criteria from the Agency, no Agency guidelines on anonymization and no enforcement practice from which standards could be inferred. Everything that follows is built on the statutory text, that sectoral antecedent, legal scholarship and comparative experience.
Every institutional fact in this section carries its numbered source: the rejection is recorded in the Senate's own official information, and the vote breakdown, which that communication does not record, is expressly attributed to the press. A warning for anyone searching for the regulation: the Ministry of Finance Decree No. 662 that Ley Chile shows as published on July 12, 2025 is a different rule, on circular economy in State procurement.
Three registers, three treatments
Statute
The law in force. It appears in serif italics, with the provision cited underneath and linked to its exact location at the Library of the National Congress. It is the only thing that binds.
Technical evidence
Scientific literature or documented deployments. It runs in ordinary prose, with the source numbered and linked at the end of the document.
The author's view
Interpretation, classification or recommendation by the author. It sits in a separate, labelled panel. There is no authority behind it.
Five propositions, each with its degree of support
-
Technical evidence
Deleting the name and the RUT does not anonymize. The combination of ordinary attributes identifies people just the same, and the empirical evidence goes back to the year 2000. A client who maintains their database is anonymized because the direct identifiers were removed has a problem: if deletion was not enough, the data remains personal data, and the statute requires the controller to be able to prove the lawfulness of the processing it carries out (article 3, letter a, of the consolidated text). That is a specific duty of substantiation owed to the data subject and the authority, not a general reversal of the burden of proof.
-
The author's view
Differential privacy offers a quantifiable guarantee, but it is neither the only admissible technique nor does it automatically turn a dataset into anonymous data. Whether it suffices will depend on the context, the data, the purpose, the means reasonably available for reidentification, the recipient's access and the quality of the documentation. What sets it apart is that it produces an auditable parameter, not that it is the only defensible choice.
-
Technical evidence
Epsilon is the central parameter, but not the only one that must be evidenced. Delta, the sensitivity of the computation, the mechanism used and composition control all belong to the same record. A report declaring only the epsilon is incomplete.12
-
Statute
The law uses the word irreversible for anonymization and the word reasonably for identifiability. They are two different thresholds within the same article. Nowhere does the law use the expression "zero risk". How those standards fit together is an open question the Agency has not resolved.
-
The author's view
For most companies this technique is not the priority. The priority is having the record of processing activities, the lawful bases, the processor agreements and the breach protocol in order. Recommending it to a client without those in place diverts budget away from what will be inspected first.
The anonymization everyone thinks they have
Attributes that look innocuous are fingerprints. The combination identifies even where each piece says nothing on its own.
They do not measure the same thing, so do not compare them side by side in a briefing note. They measure how many people end up alone in their combination of attributes, with a different design in each study. What can be asserted with support is the common conclusion: a handful of ordinary data points is enough to single out most people.156
The experiment that forced the US to change method
The United States Census Bureau attacked its own 2010 publication with the techniques then in use. The reconstructed records matched the confidential file exactly for 46.48% of the 308,745,538 people counted; cross-referencing with commercial databases produced 138 million putative reidentifications, and 52 million were confirmed, 17% of the population.4
That result explains why the 2020 census was published under differential privacy. There was no methodological preference involved, only the finding that deleting identifiers and rounding figures no longer protected anyone.
Why grouping is not enough either
The most common earlier technique grouped records until each person was indistinguishable from others. The problem is that grouping does not prevent inference: if the five people in a group share the same diagnosis, the whole group reveals it.
Opinion 05/2014 of the Article 29 Working Party set the assessment framework still in force in Europe. A technique is judged against three risks: singling out a person, linking their records across databases, and inferring an attribute about them.8
That three-risk framework is, in my view, the most useful comparative instrument that can be invoked before the Agency today, because it offers an assessment structure that Chilean law does not develop. It has no binding force in Chile and must be presented as persuasive doctrine.
Why does nobody choose the safest epsilon?
The technique does not delete information, it distorts it in a controlled way. The more protection you demand, the less useful the result.
What are you looking at?
The faint bars are the true count of an arbitrary query, for instance how many people fall in each age bracket. The highlighted bars are what actually gets published after the procedure is applied.
Move the control and watch the distance between the two. With high protection, the published figure can land far from the real one. With low protection they match almost exactly, and the protection stops meaning anything.
Here sits the conflict of interest worth anticipating. Whoever commissions the study wants accurate figures. Protection demands imprecise ones. If nobody records who set the balance point and on what grounds, there will be no one to hold accountable when the Agency asks.
In a 2017 interview, Frank McSherry, one of the technique's inventors, called an epsilon of 14 per day relatively pointless as a guarantee.15 Hence the proposal by Dwork, Kohli and Mulligan to publish the epsilons as a condition of accountability.9
That bound describes the theoretical worst case on the probability ratio, not the expected performance of a real attacker. Nor is it the probability that a person will be identified.
The 39-to-45-year bracket holds 40 people and was published as 38. The largest gap was 2 people.
With an ε of 14, no hypothesis about a person can become more than 1.2 million times more plausible. That is the limit being accepted.
What number would you have to defend?
There is no universally accepted epsilon range: the literature recorded the lack of consensus, which is why it proposes publishing each value with its reasoning.9 This report uses 0.01 to 1 as a conservative teaching reference and contrasts it with far higher documented deployments.
The United States Census strikes me as the most useful precedent there is, and not for the value it chose but for the procedure it chose it with. It set the number through a formal committee decision, on a fixed date, after public debate, placing the reasoning on record.
If your client does the same, they have a record. If an engineer sets it by default and nobody writes it down, they do not. None of this amounts to saying that any given value is legally sufficient or insufficient in Chile.
Each row measures something different. The protected unit of privacy changes: one person in the Census, one device's contributions in Gboard, each datum sent from a device in Apple's case. The time horizon changes: Apple's budgets renew every day, while the Census figure covers that entire release. And the scale and the delta change: Gboard reports in zCDP, not convertible without fixing a delta. The table is for sizing orders of magnitude and for copying the Census procedure, not for ranking deployments from best to worst.
Linear scale on epsilon. The Apple figures are per-day budgets measured by outside researchers, not per-datum values.2310
Two thresholds in the same article
The text carries an internal tension worth knowing before taking any position. It is unresolved, and the Agency has not spoken.
-
June 16, 2018
Law No. 21,096 (Ley N° 21.096)
Personal data protection enters article 19 No. 4 of the Constitution as a standalone fundamental right.
-
December 13, 2024
Law No. 21,719 · publication
Published in the Official Gazette. It amends Law No. 19,628 and creates, with effect deferred to December 1, 2026, the Personal Data Protection Agency.
-
March 1, 2025
Law No. 21,663 (Ley N° 21.663) · reporting duties enforceable
Early alert within 3 hours and a report within 72 hours to the National CSIRT for entities covered by the Cybersecurity Framework Law.14
-
April 30, 2025
MINSAL Technical Standard 241
Sectoral technical reference: anonymization of structured databases for open data in health, approved by Exempt Decree No. 23 and published in the Official Gazette. It does not come from the Agency and does not prescribe differential privacy.26
-
August 28, 2025 · open
Decree 662/2025 under review
The prevention model regulation, issued June 13, 2025, enters the Comptroller General's office. Its publication is not on record as this report closes.18
-
May 19, 2026 · open
The Senate rejects the board proposal
The two-thirds quorum is not reached. The Agency is left without a Board and the fourth transitional article's deadline expires unmet.16
-
December 1, 2026 · fully in force
Law No. 21,719 enforceable
The effective date under the current statutory calendar. The Government is weighing a postponement through a short statute, with no rule enacted.19
-
December 1, 2027
End of the phase-in
The sixth transitional article's window closes, during which the Agency may sanction smaller companies under Law No. 20,416 (Ley N° 20.416) with a written warning.
What the law says, word for word
"An irreversible procedure by virtue of which personal data can no longer be linked or associated with a specific person, nor allow their identification, because the link with the information that connects, associates or identifies that person has been destroyed or deleted. Anonymized data ceases to be personal data."
Definition of anonymization (our translation) · art. 2 letter k), consolidated Law No. 19,628"All objective means and factors that could reasonably be used for such identification at the time of processing."
Identifiability standard (our translation) · art. 2 letter f), same provisionNowhere does the law use the expression "zero risk". It uses "irreversible" for anonymization and "reasonably" for identifiability. How the two thresholds fit together is what remains unresolved.
The text admits of at least two readings. The first, probably the most literal, takes "irreversible" to impose an obligation of result stricter than the European reasonable-means approach, with the consequence that full anonymization would be practically unattainable. The second holds that the reasonableness formula in letter f) should inform the interpretation of the whole.
Neither is taken from an author: they are, in my view, the readings the text admits, and I know of no Agency ruling or Chilean case law settling the point. Until that happens both are defensible, and the uncertainty rests on the controller. Comparative law has moved in the second direction: the Court of Justice of the European Union held in 2025 that the personal character of data is not absolute and depends on the recipient's means.11
Article 16 quinquies
It is not a new lawful basis. The law deems a legitimate interest to exist where processing is carried out exclusively for historical, statistical, scientific and study or research purposes, all of which must serve public interest ends. It further requires adopting and evidencing the quality and security measures, and anonymizing beforehand whatever is published.7 It does not authorize statistical processing in general.
In my view those conditions operate jointly: failing any one of them defeats the presumption of legitimate interest. It is a reasonable reading and I believe the correct one, but it is an interpretation: the statutory text sets out the conditions and does not literally describe them as cumulative.
- Publication in the Official Gazette
- Dec 13, 2024
- Effective date under the current calendar
- Dec 1, 2026
- Days remaining
- 98 days
- Agency constituted
- No
- Agency guidelines
- None
Eight terms and what they mean in a legal setting
You do not need to operate any formula. You need to know what you are being shown when a vendor hands you a technical report.
Epsilonε · privacy budget
The number that governs how much protection is applied. The lower it is, the more protection and the less accuracy. It is the central parameter and the easiest to audit, though not the only one that must be on record. If your vendor's report does not include it, it is incomplete as evidence.
The eε boundwhat epsilon guarantees
Epsilon caps the ratio between probabilities: no conclusion about a person can become more than eε times more plausible than it already was. With epsilon 1 that factor is 2.7. With epsilon 14 it is 1,202,604. It does not measure the probability that someone will be identified: it bounds how far the scales can tip. Confusing the two is the most common error in translating the parameter.
Deltaδ
The probability that the guarantee simply fails. It is not the probability that any specific person is reidentified: it is the margin within which the epsilon bound stops applying. It should be much smaller than one divided by the number of people in the database, because with a delta of that order the definition still tolerates publishing one person's complete record. If the report declares an epsilon but omits the delta, the guarantee is incomplete.
Sensitivityof the computation
How much a result changes if a single person enters or leaves the database. It determines how much noise must be added. A poorly characterized computation receives less noise than it needs and the guarantee is broken from the outset, even if the declared epsilon looks right.
Compositionbudget exhaustion
Every published query spends part of the budget. Ten queries at epsilon 1 amount, under the basic bound, to epsilon 10. If nobody keeps count, the protection runs out unnoticed and the technical report stops supporting anything. It is the most neglected point.
Laplace mechanismand its Gaussian variant
The two standard formulas for computing how much noise to add. You do not need to understand them. You do need it on record which one was used, because that determines whether the declared epsilon is pure or comes with a delta.
Anonymizationart. 2 letter k), consolidated Law No. 19,628
An irreversible procedure that destroys the link with the information that connects, associates or identifies the person. Anonymized data ceases to be personal data and leaves the scope of the law. It is the only way out of the regime as a whole, which is why everyone wants through it.
Pseudonymizationart. 2 letter l), consolidated Law No. 19,628
Processing that prevents attributing the data to a subject without additional information kept separately. The data remains personal and remains inside the law. Confusing it with anonymization is the most frequent and the most expensive mistake.
The definitions come from the consolidated text of Law No. 19,628 in the version in force on December 1, 2026, checked letter by letter against the Library of the National Congress; every citation in this report links to that official version.7 The four technical parameters of the record are developed in NIST SP 800-226 and in the UK ICO's anonymisation guidance, both of March 2025.1213
Five questions and the document that answers each
The entire technical side exists for a single situation: someone having to answer in writing.
The questions are my own. They anticipate the evidentiary standard that in my view follows from the duties of substantiation the statute places on the controller (articles 3 letter a, 13, 14 and 14 quinquies of the consolidated text), but they reproduce no official questionnaire and no enforcement practice, because the Agency is not constituted and there is no practice to draw them from.
-
On what lawful basis do you process this data, and for what specific purpose?
Answered by The record of processing activities, with purposes kept separate. A consent form drafted to cover everything does not answer this question.
-
If you maintain the data is anonymized, what procedure did you apply and with what parameters?
Answered by The technical report, with the mechanism, the epsilon, the delta and the sensitivity, each with its justification. With less than that there is no record to review.
-
Who approved those parameters, when, and on what grounds?
Answered by The data governance committee's minutes and the version history with dates and owners. It is the question that turns an engineering decision into an attributable one.
-
How many releases have you made from the same database since then?
Answered by The budget consumption log. If nobody kept it, the declared guarantee may have run out months ago without anyone noticing.
-
Did you attempt reidentification before publishing, and with what result?
Answered by The results of the tests against the three risks in Opinion 05/2014. Filed, dated, even where they came out clean.
The sanctions regime, with its articles
This section cites the provision next to each statement, linked to the official text. Several formulations common in the market do not survive the comparison.
| Concept | Provision | Content |
|---|---|---|
| Scale of fines | Art. 35 | Minor: written warning or a fine of up to 5,000 UTM. Serious: up to 10,000 UTM. Very serious: up to 20,000 UTM.Checked against the consolidated text of Law No. 19,628, in force December 1, 2026, linked on each row. |
| Recidivism | Art. 36 | Exists where the controller has been sanctioned on two or more occasions within the last 30 months, under final or enforceable decisions. Article 35 then allows a fine of up to three times the amount.30 months for recidivism, 24 for suspension. They are different periods. |
| Revenue caps | Art. 35 | For companies that are not smaller companies under Law No. 20,416 and that repeat serious or very serious infringements: the more onerous of three times the amount and 2% or 4% of annual revenue, by category.The percentage cap does not apply to smaller companies. |
| Surcharge for failing to remedy | Art. 35 | The Agency specifies the remedial measures, to be adopted within no more than 60 days; otherwise a 50% surcharge is added to the fine. |
| Aggravating circumstancesClosed list of three | Art. 36 | Only three: recidivism; the continuing character of the infringement; and having put at risk the security of data subjects' rights and freedoms. |
| Sensitive data and data of children and adolescents | Art. 37, No. 5 | It is not an aggravating circumstance. It is one of the eight criteria the Agency must apply prudentially to set the amount of the fine, alongside gravity, lack of diligence, harm, economic benefit, economic capacity, prior sanctions and whatever mitigating and aggravating circumstances concur. |
| Mitigating circumstancesList of five | Art. 36 | Unilateral remediation or reparation agreements; cooperation with the investigation; absence of prior sanctions; self-reporting with notice of the measures adopted; and diligent compliance with management and supervision duties. |
| Infringement prevention model | Arts. 49, 51 and 52 | Adopting one is voluntary. The mitigating circumstance is not certification: it is having diligently complied with management and supervision duties, and the certificate the Agency issues under article 51 is the means by which that is verified. Certification without actual compliance has no effect.Certificates last three years (art. 52). |
| Suspension of processingAncillary sanction | Art. 38 | Available where fines are imposed for repeated very serious infringements within a period of 24 months. It lasts up to 30 days, but can be extended indefinitely for successive periods if the order goes unmet. It can be partial or total. It does not affect storage. It cannot be ordered where it would affect data subjects' rights. If it reaches an entity supervised by a sectoral regulator, the Agency must first put the background before that authority.The law uses neither "essential services" nor "public utility". |
Suspension is often presented in the market as an immediate threat. The text conditions it considerably: it requires prior fines, for very serious infringements, repeated, within 24 months, and it forbids ordering it where data subjects would be harmed. It is an escalation sanction, not a first strike.
That said, its indefinite extension in 30-day periods for as long as the situation goes uncorrected is the feature that does deserve attention in planning, and the one usually left out.
A warning about numbering
Law No. 21,719 is an amending statute. The articles cited follow the numbering of the consolidated text of Law No. 19,628 as amended, which is the numbering under which the Library of the National Congress publishes the text in force and the one used by this report's links. Confusing it with the amending statute's own numbering leads to citing provisions that do not exist.
-
2012
The Banco de Chile "cartolazo"Banking
52,770 customers affected by statements mailed with other people's data. A 40 million peso fine from the banking regulator and a compensation plan of more than 1.3 billion pesos, at SERNAC's demand.20
-
2016
Failure in the MINSAL networkPublic health
At least 3 million files with patient information sat accessible in shared folders, laboratory tests included. Reported by CIPER on March 5.21
-
2018
Attacks on Banco de Chile and Banco ConsorcioBanking
US$ 10 million taken via SWIFT in May and close to US$ 2 million in November. In July, a group calling itself ShadowBrokers leaked data from more than 14,000 credit cards, most of them inactive.22
-
2023
Ransomware at FONASAHealth
A February 17 attack that private researchers attributed to BlackCat, with some 250 machines compromised. FONASA filed a criminal complaint and no official count of affected people exists.23
-
2024
Santander GroupBanking
Unauthorized access to a database at an outside provider, with customers in Chile, Spain and Uruguay. SERNAC and the CMF formally requested information from the bank in May; in 2026 SERNAC filed a class action.24
-
2025
National Fan RegistryBiometrics
More than a million people enrolled and 188 matches covered, according to the football association, as of November. Enrollment verifies identity with facial biometrics; facial checks at stadium gates began as a pilot phase.25
Cases predating the full entry into force of Law No. 21,719 and resolved under other frameworks. They are included as a reference of scale, not as precedents of the new regime.
When to recommend it, and when to advise against it?
In my view it is a priority in a narrow set of cases. Outside them it diverts budget away from what will be inspected first.
The table below is my own classification. It weighs volume processed, data sensitivity and overlap with sectoral regulators holding current powers. It comes from no authority and does not claim to exhaust what a specific case would demand.
| Industry | Exposure | Regulator | Critical data | What to recommend |
|---|---|---|---|---|
| Healthcare, isapres and insurersHighest density of sensitive data | Very high | Superintendency of Health · MINSAL | Health, genetic, biometric | UsefulFor statistics and research. Not applicable to the clinical record. |
| Banking, finance and creditDual supervision | Very high | CMF · UAF | Transactional, scoring, biometric | UsefulFor anti-fraud and statistics. Not applicable to individual scoring. |
| Technology, SaaS, AI and cloudActs as a processor | High | Agency · ANCI | Training data and third-party data | A priorityModel training, synthetic data, federated learning. |
| Telecommunications and mobilityGeolocation | High | Subtel | Metadata and location | A priorityIn every release of aggregate statistics. |
| Public sector and open dataVulnerable data subjects | High | CPLT · Agency | Socioeconomic, biometric, health | A priorityOfficial statistics and open data programs. |
| Retail, e-commerce and advertisingAn active SERNAC | High | SERNAC | Consumption profiles and identifiers | UsefulFor ad measurement. Not for the company's own CRM. |
| EducationHigh share of minors | Medium-high | MINEDUC · Superintendency | Minors' data and performance | UsefulWhen publishing results for small classes. |
| Human resourcesCuts across every industry | Medium | Labor Directorate | Health, biometrics, monitoring | UsefulClimate surveys in small teams. |
| Energy and water utilitiesHousehold consumption profiles | Medium | SEC | Smart meter consumption | UsefulNetwork analytics. |
| Tourism, real estate, legal, BPO and NGOsLimited volume | Variable | Various | By line of business | Advised against for nowBasics first: lawfulness, security and contracts. |
Five things to resolve, in this order
None involves differential privacy. They are the base on which any advanced technique acquires evidentiary meaning, and none depends on the Agency being constituted to be enforceable.
- Record of processing activities and lawful basesDocumented by purpose. A generic consent for every use is not a lawful basis.
- An operational breach protocolDetection, notice to the Agency, notice to data subjects where due, and coordination with the ANCI where applicable.
- Processor agreements and a transfer mapStandard clauses, control over sub-processors and a record of which country each transfer goes to.
- A data subject rights workflowAccess, rectification, erasure, objection, portability and blocking, plus explanation and human intervention in automated decisions.
- Impact assessment and demonstrable securityWith auditable evidence, prioritizing sensitive, biometric and children's data.
And if you are also going to recommend differential privacy
The record supporting a claim of anonymization should contain, at a minimum: the technical report with the mechanism, epsilon, delta and sensitivity each justified; the impact assessment; the minutes of the committee that approved the parameters; the version history with dates and owners; the budget consumption log across queries; and the results of reidentification tests run before publishing.
Composition is the most neglected point and the easiest to lose. Every query spends budget. Publishing successive results without keeping count exhausts the guarantee in silence.
None of this automatically turns a dataset into anonymous data. Anonymization must be assessed case by case, weighing the context, the purpose, the population concerned, the auxiliary sources available, the recipient's access and the risk of linkage with other databases. A complete record improves the controller's position; it does not replace that analysis.
One consequence is worth stating outright: until that analysis concludes that the data ceased to be personal, the processing remains fully subject to the law. Differential privacy is not, by itself, a safe harbor. Implementing it does not establish the lawfulness of the processing, nor does it excuse compliance with the principles of purpose, proportionality, transparency and security in article 3 of the consolidated text, or with the controller's remaining obligations.7
Open-source tools: OpenDP, Google DP Library, IBM diffprivlib, Tumult Analytics, SmartNoise and PipelineDP. None leads in every area. The choice is worth documenting just like the parameters.
Sources
- de Montjoye, Hidalgo, Verleysen and Blondel, "Unique in the Crowd: The privacy bounds of human mobility", Scientific Reports 3, 1376, 2013.
- Tang, Korolova, Bai, Wang and Wang, "Privacy Loss in Apple's Implementation of Differential Privacy on MacOS 10.12", arXiv:1709.02753, 2017. The loss per datum is 1 or 2; the measured daily budgets are 6 (macOS 10.12.3), 14 (iOS 10.1.1) and at least 29 more in the iOS 11 beta, on the order of 43.
- U.S. Census Bureau, DSEP committee decision of June 9, 2021 on the redistricting file parameters: a total budget of 19.61, made up of 17.14 for persons and 2.47 for housing units.
- U.S. Census Bureau, "Summary of Simulated Reconstruction-Abetted Re-identification Attack" on the 2010 census: a 46.48% exact match, 138 million putative reidentifications and 52 million confirmed (17%).
- Sweeney, "Simple Demographics Often Identify People Uniquely", Carnegie Mellon University, 2000; and Golle, "Revisiting the Uniqueness of Simple Demographics in the US Population", 2006.
- de Montjoye, Radaelli, Singh and Pentland, "Unique in the shopping mall: On the reidentifiability of credit card metadata", Science 347 (6221), 2015, pp. 536-539.
- Consolidated text of Law No. 19,628 in the version in force on December 1, 2026 (articles 2, 3, 13, 14, 14 quinquies, 16 quinquies, 35, 36, 37, 38, 49, 51 and 52), and Law No. 21,719, Official Gazette of December 13, 2024, first, fourth and sixth transitional articles.
- Article 29 Working Party, Opinion 05/2014 on anonymisation techniques (WP216).
- Dwork, Kohli and Mulligan, "Differential Privacy in Practice: Expose Your Epsilons!", Journal of Privacy and Confidentiality 9(2), 2019.
- Xu et al., "Federated Learning of Gboard Language Models with Differential Privacy", ACL 2023.
- CJEU, EDPS v SRB, C-413/23 P, September 4, 2025; and CJEU, Breyer, C-582/14, 2016.
- NIST, SP 800-226, "Guidelines for Evaluating Differential Privacy Guarantees", March 2025.
- ICO, guidance on anonymisation and pseudonymisation, March 2025.
- Law No. 21,663, Cybersecurity Framework Law, and the National Cybersecurity Agency's regulations.
- Greenberg, "How One of Apple's Key Privacy Safeguards Falls Short", WIRED, 2017, with Frank McSherry's assessment.
- Senate of Chile, "Desestiman propuesta de consejeros para la Agencia de Protección de Datos Personales", May 20, 2026.
- Diario Financiero, report on the Senate's rejection of the board proposal, May 2026.
- Ontier, note on Decree No. 662/2025 entering the Comptroller's review, with the issuance date and the state of the procedure.
- La Tercera · Pulso, "La difícil implementación de la ley de protección de datos personales", August 2026.
- El Mostrador, the banking regulator's fine on Banco de Chile and the compensation plan, July 2012.
- CIPER, "Grave falla en la red del Minsal dejó expuesta información confidencial de pacientes", March 5, 2016.
- Cooperativa, Banco de Chile's confirmation of the theft; Diario Financiero, the cyberattack on Banco Consorcio; BioBíoChile, the card leak, 2018.
- CronUp, analysis of the FONASA ransomware; Cooperativa, confirmation of the incident, 2023.
- Cooperativa, SERNAC's and the CMF's formal inquiries to Santander, May 2024; Chócale, SERNAC's class action, March 2026.
- ANFP, "Registro Nacional de Hinchas supera el millón de inscritos", November 13, 2025.
- Ministry of Health, Technical Standard No. 241, on anonymization for publishing databases as open data, approved by Exempt Decree No. 23 of April 9, 2025, of the Undersecretariat of Public Health, Official Gazette of April 30, 2025. Scope at p. 8; verification thresholds, the motivated intruder test and the periodic reassessment warning at p. 11.
A technical and legal report for information purposes, first version, closed in August 2026. It is not legal advice for a specific matter, nor a settled legal opinion on compliance with Law No. 21,719.
The statute is not yet fully in force, its regulations are in progress, the Personal Data Protection Agency has not been constituted and has issued no ruling on anonymization, pseudonymization or privacy enhancing technologies. The bill on artificial intelligence systems remains in progress, and there is an unresolved government assessment on postponing the entry into force.
The criteria on epsilon values, the enforcement questions, the sector classification and the recommendations are the author's own and are labelled as such. The provisions cited were checked against the official consolidated text published by the Library of the National Congress, in force December 1, 2026, linked at every citation so the reader can verify it without intermediaries. Statutory quotations are our translation of the official Spanish text, which prevails.
If your organization needs to assess a claim of anonymization or prepare the record supporting it, contact us for a specific analysis.