Originally published as Daniel J. Solove The AI Inference Problem in Privacy Law, Wolters Kluwer Cybersecurity Policy Report (Oct. 19, 2026)
You can download a PDF of the piece here:
Abstract
AI’s profound power to generate inferences from data creates an enormous problem for privacy law. AI provides an unprecedented ability to determine patterns in aggregated data which can produce inferences about sensitive data.
There are two ways in which AI inferences undermine privacy law. First, many privacy laws are premised on giving individuals control over the personal data they choose to share with organizations. Almost all privacy and data protection laws, including the EU’s General Data Protection Regulation (GDPR), mandate that organizations inform people with a privacy notice about the data collected. When AI allows inferences to be made, the inferred data is beyond what was disclosed to people in the privacy notice. The idea that individuals consent to this data collection via opt-out (as in many U.S. laws) or opt-in (as in the GDPR, some U.S. laws, and most global laws) becomes a farce if other personal data about people beyond what they consented to is inferred.
Second, many privacy laws protect sensitive data differently from regular personal data. With AI inferences, sensitive data can readily be inferred from combining pieces of regular personal data. Determining which data should be classified as sensitive becomes nearly impossible.
In this essay, I’ll elaborate on these claims and explain the practical consequences, which shatter the foundations of many privacy laws and make them nearly impossible to comply with.
The Power of AI Inferences
What makes this story so important to privacy law is that it involves sensitive data (health) and that the inference was made from particularly innocuous personal data—unscented lotion, scent-free soap, cotton balls, and vitamins.
Now, more than a decade later, AI algorithms are even more sophisticated at finding patterns in data and making inferences. Alicia Solow-Niederman aptly labels our current times as an “inference economy” where organizations use AI to infer additional data about individuals beyond the data they supplied as well as data about other people.[2]
There are inferences that can be made based on common correlations in data. It’s possible to infer race or religion from where people live, as people tend to live in particular demographic clusters. Religion can be inferred from what people eat and wear. Health conditions can be inferred from supermarket purchases, physical activity, and fitness data. In our highly politicized world these days, political beliefs can be inferred from the products people buy, the places they eat, and the entertainment they watch.
The examples above are inferences that are rather unsurprising. Humans can make them, albeit somewhat crudely, but AI algorithms can make these inferences more readily and with much greater precision. But this is just the tip of the iceberg. AI’s superpower is its ability to discover patterns in data that humans don’t see, giving it the ability to make unexpected inferences and infer a lot more than what could be inferred before.
There have been hundreds of studies demonstrating how readily data about a person can be inferred. One overview of 327 inference studies listed countless inferable attributes, including gender, age, politics, location, occupation, race and ethnicity, family and relationships, education, income, health, religion, sexual orientation, and social class.[3]
In one study, researchers could infer drug use with 65% accuracy from Facebook likes.[4] Another study was able to infer mental illness based on people’s social media posts.[5] Researchers used Facebook activity to infer the likelihood of women suffering postpartum depression.[6] Researchers were able to correctly infer political party affiliation 85% of the time from Facebook likes.[7] The same researchers were also able to correctly identify people’s race from Facebook likes 95% of the time.[8]
The inference problem is particularly troublesome because AI can make inferences that surprise us, and we’re constantly discovering new ways that AI is making inferences. If we don’t know in advance what AI can infer, then we’re flying blind. And, as AI algorithms evolve, they likely will be able to improve their ability to infer.
The Implications for the Individual Control Model
But if inferences can readily be made from the data that companies collect, then the notion that individuals have any control over the data they are sharing is a complete farce. Returning to the Target pregnancy case, shoppers had no idea that Target was inferring their pregnancy from the items they bought. They thought they were just revealing to Target that they bought things like soap and lotion, but instead they ended up revealing sensitive health data.
The individual control model has long been broken and unworkable, as most people don’t read privacy notices and have no idea how to assess the risks of sharing their data or agreeing to uses of it.[9] But AI inferences make a complete mockery of the idea that individuals have any control even over the details about their lives they are sharing.
The AI inference problem should be the final nail in the coffin of the individual control model. At the very least, making inferences from data should be treated as tantamount to collecting personal data.
The Implications for Sensitive Data
- racial or ethnic origin
- political opinions
- religious or philosophical beliefs
- trade-union memberships
- health
- sex life or sexual orientation
- genetic data
- biometric data[10]
In the United States, nearly half the states have passed consumer privacy laws, and all recognize sensitive data. The laws, though, recognize different types of data as sensitive. Some of the most common categories include:
- racial or ethnic origin
- religious beliefs
- mental or physical health diagnosis
- sexual orientation
- genetic or biometric data
- personal data collected from a known child
- precise geolocation
Other categories of sensitive data that are often recognized include:
- citizenship or immigration status
- financial account information
- neural data
- transgender or nonbinary status
- government identification information[11]
Around the world, most comprehensive privacy laws of countries recognize sensitive data.[12]
Sensitive data is afforded heightened protections beyond regular personal data. Under the GDPR, the lawful basis of “legitimate interests,” which is one of the most commonly used, is unavailable for sensitive data. Additionally, a data protection officer must conduct a data protection impact assessment (DPIA) when processing a “large scale of special categories of data.”[13]
AI inferences make it very easy to infer sensitive data from non-sensitive data.[14] This either allows companies to evade sensitive data protections by using non-sensitive data as a proxy or it means that non-sensitive data that could give rise to inferences about sensitive data should also count as sensitive data.
Many Privacy Laws Recognize that Inferences Count
In the U.S., many state consumer privacy laws don’t address whether inferences count as sensitive data. A few, however, recognize inferences. For example, the California Consumer Privacy Act (CCPA) definition of “personal information” includes “[i]nferences drawn from any of the information identified in this subdivision to create a profile about a consumer reflecting the consumer’s preferences, characteristics, psychological trends, predispositions, behavior, attitudes, intelligence, abilities, and aptitudes.”[18] But in certain circumstances, the CCPA turns its back on inferences when it states that neural data is “not inferred from nonneural information.”[19]
Recently, Delaware amended its Personal Data Privacy Act to include “inferences made based on personal data, alone or in combination with other data, that are used to reveal or identify [sensitive data].”[20] Colorado’s Privacy Act rules explicitly recognize “sensitive data inferences” made from personal data, “alone or in combination with other data.”[21]
The fact these laws recognize inferences strikes me as quite sensible, though I find it surprising that they haven’t attracted more controversy. It seems that not enough people are recognizing just how far down the inference rabbit hole goes.
Many U.S. state consumer privacy laws don’t even address the issue, but there is often nothing stopping interpretations of the definition of sensitive data to encompass inferences.
The Sensitive Data Dilemma
This presents a dilemma like Scylla and Charybdis. On the one hand, laws could recognize that inferences count, in which case nearly all data could potentially be sensitive if inferences are made about it. To comply with the laws, data must be classified at collection, but how can this happen if the data could become sensitive later depending upon potential inferences? And the ability to make inferences evolves, so what’s possible now might be different tomorrow.
Since the rules are dramatically different for sensitive data versus regular personal data, companies must know how to classify the data and what form of consent to procure when obtaining the data or what legal basis can be used to gather the data. But with the inference problem, companies can’t really know how to classify data. All personal data could potentially be sensitive, so this would mean that the sensitive data provisions of laws would apply to nearly every piece of data collected.
So far, GDPR enforcers haven’t fully implemented the logic of the EDPB and CJEU, but if they did, then most companies might be in violation unless they treat nearly all the personal data they collect as sensitive.
On the other hand, law could refuse to recognize inferences, but this opens up enormous loopholes. Not covering inferences is incredibly naive in today’s age of AI and digital analytics because AI can readily identify proxy data for nearly any kind of sensitive data, allowing sensitive data protections to be sidestepped.
AI can also avoid generating data altogether and go straight to use. AI algorithms can identify patterns in data that correlate to sensitive data categories such as race or religion that can then be used for decisions based on these patterns—all without producing sensitive data. But the very reason laws protect sensitive data is because this data is used problematically in decisions affecting individuals. In other words, AI’s inferential powers can be used to make decisions involving sensitive data attributes without actually using the sensitive data itself—like the trick of a devious magician using mirrors to make something look like it is invisible.
Conclusion
Citation: Daniel J. Solove The AI Inference Problem in Privacy Law, Wolters Kluwer Cybersecurity Policy Report (Oct. 19, 2026)
You can download a PDF of the piece here:
[1] Charles Duhigg, How Companies Learn Your Secrets, N.Y. Times Mag. (Feb. 16, 2012), https://www.nytimes.com/2012/02/19/magazine/shopping-habits.html.
[2] Alicia Solow-Niederman, Information Privacy and the Inference Economy, 117 Nw. U. L. Rev. 357, 361 (2022); see also Hideyuki Matsumi, Predictions and Privacy: Should There Be Rules About Using Personal Data to Forecast the Future?, 48 Cumb. L. Rev. 149 (2017); Sandra Wachter & Brent Mittelstadt, A Right to Reasonable Inferences: Re-Thinking Data Protection Law in the Age of Big Data and AI, 2019 Colum. Bus. L. Rev. 494.
[3] Joanne Hinds & Adam N. Joinson, What Demographic Attributes Do Our Digital Footprints Reveal? A Systematic Review, 13 PLoS ONE, Nov. 28, 2018, at 1, 5.
[4] Michal Kosinski, David Stillwell & Thore Graepel, Private Traits and Attributes Are Predictable from Digital Records of Human Behavior, 110 PNAS 5802, 5802–03 (2013).
[5] Jina Kim, Jieon Lee, Eunil Park & Jinyoung Han, A Deep Learning Model for Detecting Mental Illness from User Content on Social Media, 10 Sci. Reps. 11846 (2020).
[6] Munmun De Choudhury, Scott Counts, Eric J. Horvitz & Aaron Hoff, Characterizing and Predicting Postpartum Depression from Shared Facebook Data, in Ass’n for Computing Mach., CSCW ’14: Proceedings of the 17th ACM Conference on Computer Supported Cooperative Work & Social Computing 625, 626 (2014).
[7] Kosinski et al. at 5802-03.
[8] Id.
[9] Daniel J. Solove & Woodrow Hartzog, Kafka in the Age of AI and the Futility of Privacy as Control, 104 B.U. L. Rev. 1021 (2024).
[10] GDPR, art. 9.
[11] Daniel J. Solove, U.S. State Consumer Privacy Laws: A Visual Guide with Whiteboards (July 2026), https://teachprivacy.com/state-privacy-law-guide/.
[12] Daniel J. Solove, Data Is What Data Does: Regulating Use, Harm, and Risk Instead of Sensitive Data, 118 Nw. U. L. Rev. 1081 (2024), https://ssrn.com/abstract=4322198.
[13] GDPR, art. 6.1(c), 35.3(b).
[14] Solove, Data Is What Data Does; see also Paul Ohm, Sensitive Information, 88 S. Cal. L. Rev. 1125 (2015); Tal Z. Zarsky, Incompatible: The GDPR in the Age of Big Data, 47 Seton Hall L. Rev. 995 (2017).
[15] Article 29 Data Protection Working Party, Advice Paper on Special Categories of Data (“Sensitive Data”), at art. 8(1) (2011), https://ec.europa.eu/justice/article-29/documentation/other-document/files/2011/2011_04_20_letter_artwp_mme_le_bail_directive_9546ec_annex1_en.pdf (noting that
[16] Annex—Health Data in Apps and Devices, at 1, 2, Article 29 Data Protection Working Party, (2015), https://ec.europa.eu/justice/article-29/documentation/other-document/files/2015/20150205_letter_art29wp_ec_health_data_after_plenary_annex_en.pdf.
[17] Case 184/20, OT v. Vyriausioji tarnybinės etikos komisija, 2022 E.C.R ¶ 117.
[18] Cal. Civ. Code § 1798.140(v)(1)(K).
[19] Cal Civ. Code § 1798.140 (ae)(1)(G)(ii). For more on neural data, see Daniel J. Solove Rethinking Privacy of the Mind: Why Protecting Neural Data Is Insufficient, Wolters Kluwer Cybersecurity Policy Report (Oct. 6, 2026).
[20] 6 Del. C. § 12D-102(e).
[21] Colorado Privacy Act Rules 4 CCR-904-3, Rule 2.02 (Sept. 29, 2022).
* * * *
Daniel J. Solove is the Bernard Professor of Intellectual Property and Technology Law at the George Washington University Law School. He is the founder of TeachPrivacy, a company that provides workforce privacy, cybersecurity security, and AI training to companies and organizations around the world. He is the author of 10+ books and 100+ articles.
You can follow his events, writings, training, cartoons, and resources by subscribing to his free weekly newsletter.
Subscribe to Solove’s Free Substack
A supplement to Solove’s regular newsletter with more in-depth discussions
