18 AI Privacy Violations: Real Examples (2026)

AI privacy violations occur when artificial intelligence systems collect, process, or share personal data without a valid legal basis, such as consent, or without telling the people involved. Examples of AI privacy issues that became enforcement cases include Clearview AI's face database built from scraped photos, Meta's facial recognition in photo tagging, and Uber's fully automated driver deactivations.

US settlements over biometric data now run into the billions. Texas settled with Meta for $1.4 billion and with Google for $1.375 billion, and both settlements covered biometric data.

AI privacy violations: key statistics (2026)

AI privacy penalties come from three sources: regulator fines, government settlements, and statutory damages in private lawsuits.

• €7.1 billion: total GDPR fines for all types of violations, AI-related or not, from May 2018 to January 2026, according to DLA Piper's survey. Meta holds the single largest fine of any kind, €1.2 billion in 2023.
• €825 million: the Dutch data protection authority's fine on Uber in August 2026 for deactivating drivers through fully automated decisions.
• $1.4 billion: Meta's settlement with Texas in 2024 over facial recognition in photo tagging, which followed a $650 million Illinois class settlement over the same feature.
• $5,000 per violation: statutory damages under section 637.2 of the California Penal Code, the damages provision of the California Invasion of Privacy Act (CIPA). How violations are counted in website tracking cases is still contested.
• $1,000 to $5,000 per violation: damages for negligent and for intentional or reckless violations of Illinois' Biometric Information Privacy Act (BIPA). Since August 2024, repeated collection from one person by the same method counts once.
• €35 million or 7% of global turnover: the maximum EU AI Act fine for prohibited AI practices. The bans have applied since February 2, 2025.
  
  

What are AI privacy violations?

An AI privacy violation, or AI invasion of privacy, differs from an ordinary privacy violation in scale. One AI model can process millions of people's data without their knowledge. The system can be a machine learning model, a facial recognition tool, a recommendation engine, or a generative AI product.

Regulators apply existing laws such as the GDPR, CCPA, BIPA, and CIPA to AI data practices, because those laws cover personal data whatever technology processes it. AI-specific rules, led by the EU AI Act and new US state laws, add a second layer.

How does AI violate privacy?

AI violates privacy when a system collects, infers, records, or exposes personal information without a valid legal basis or the person's knowledge. It also violates privacy when it makes decisions about people without safeguards such as human review. AI privacy violations follow seven recurring patterns, and one company can hit several at once:

• Training on personal data without a legal basis: scraping posts, chats, or other personal data into training sets, as with OpenAI, Character.AI, and X.
• Collecting biometrics without consent: faceprints and voiceprints captured without the written consent Illinois BIPA requires, as with Clearview AI and in the Meta and TikTok settlements.
• Recording conversations without everyone's consent: AI notetakers and voice assistants that capture people who never agreed, as alleged against Otter.ai and Google Assistant.
• Profiling people for targeted ads: analyzing their behavior for advertising without a valid legal basis, as with LinkedIn.
• Deciding without human review: fully automated decisions about work, credit, or accounts, as with Uber's driver deactivations.
• Using a person's name or identity: AI features attributed to real people without permission, as alleged against Grammarly.
• Leaking data through prompts and security failures: confidential data pasted into chatbots, and AI products or their vendors that expose stored data in breaches.

Five of the seven AI privacy patterns come down to consent or another legal basis for using the data: training, biometrics, recording, ad profiling, and identity. The other two, automated decisions and data leaks, turn on human review and security instead.

Examples of AI privacy issues: 18 real cases and their outcomes

Outcomes in these 18 AI privacy cases range from no penalty at all, as with the Royal Free and DeepMind in 2017, to Meta's $1.4 billion settlement with Texas in 2024. Several are still in court.

1. Clearview AI: facial recognition without consent (ongoing, €90M+ in EU fines)

Clearview AI scraped more than 30 billion photos from the internet to build a facial recognition database it sells to law enforcement. The people pictured never knew or consented.

Why it's a violation: The GDPR prohibits processing biometric data to identify people unless a narrow exception applies, such as explicit consent, and Clearview had none. Illinois BIPA requires a written release before a company captures anyone's face geometry.

What happened: Data protection authorities in Italy, Greece, and France fined Clearview €20 million each. France added a €5.2 million penalty in 2023 because Clearview didn't prove it had stopped collecting and deleted French residents' data. The Dutch authority fined it €30.5 million in 2024.

The UK's £7.55 million fine is still on appeal. Australia's regulator found breaches but issued no fine. In the US, a court approved a BIPA settlement in March 2025. It gives the class a 23% equity stake in Clearview, valued at about $51.75 million, instead of cash.

Lesson: Under the GDPR, a photo being publicly available doesn't make it lawful to use for facial recognition. Businesses that use facial recognition need a legal basis and, where the law requires it, opt-in consent. They also need a written retention policy and a deletion process before collecting a single image.

2. Google DeepMind and the Royal Free: 1.6 million patient records shared for an app test (UK, 2017)

The Royal Free NHS Foundation Trust gave Google's DeepMind about 1.6 million records to test Streams, an app that alerts clinicians to acute kidney injury. Patients weren't told their data would be used that way.

Why it's a violation: UK data protection law required the trust to tell patients how their records would be used. It also had to share no more data than the project needed.

What happened: In July 2017 the Information Commissioner's Office (ICO) found the trust had failed on both counts, in breach of the Data Protection Act 1998. The ICO required the trust to sign an undertaking, and there was no fine. A later representative claim, brought on behalf of all the affected patients, was struck out before trial. The Court of Appeal dismissed the appeal in December 2024.

Lesson: A good cause doesn't replace a legal basis. Organizations that hand patient or customer data to an AI company for a new purpose need to tell people first and share only what the project needs.

3. Amazon Alexa: children's voice recordings kept after deletion requests ($25M, 2023)

Amazon kept children's Alexa voice recordings indefinitely, the FTC and Justice Department alleged in 2023. According to the complaint, Amazon failed to delete transcripts from all its databases even when parents asked for deletion. It also used the data to improve its Alexa algorithm.

Why it's a violation: The Children's Online Privacy Protection Act (COPPA) requires verifiable parental consent before collecting data from children under 13. It also requires deletion on request and bars keeping the data longer than reasonably necessary. The FTC also alleged that Amazon misled users about deleting their geolocation data.

What happened: Amazon agreed to a $25 million civil penalty in 2023. In a separate FTC case the same day, Amazon's Ring agreed to pay $5.8 million in consumer refunds. The FTC alleged that Ring employees could view customers' private videos and that Ring used customer videos to train its algorithms without consent.

Lesson: Both Amazon cases turned on data kept or reused beyond what users agreed to. Voice assistants and other AI products that keep user data need deletion that reaches every copy, including transcripts and training sets.

4. Meta: facial recognition in photo tagging without consent ($650M and $1.4B, 2021 and 2024)

Meta's Tag Suggestions feature used facial recognition to identify people in uploaded photos and suggest tags. The Illinois class action alleged it did so without the written consent Illinois law requires.

Why it's a violation: BIPA requires written notice and a written release before any company collects a person's face geometry. Texas's Capture or Use of Biometric Identifier Act has a similar consent rule.

What happened: Meta settled the Illinois BIPA class action for $650 million in 2021, then settled with Texas for $1.4 billion in 2024 over the same feature. The Texas attorney general called it the largest settlement ever obtained in an action brought by a single state.

The feature itself is gone: Meta shut down Facebook's face recognition system in 2021. In October 2024 it announced expanded tests of facial recognition to detect scam ads and help users recover accounts.

Lesson: Biometric laws in Illinois and Texas cover consumer features such as photo tagging. An AI feature that scans the faces or voices of Illinois or Texas residents needs consent before the first scan.

5. TikTok: alleged faceprint collection without disclosure ($92M settlement, 2021)

TikTok used facial recognition to collect users' facial data without the disclosure and consent BIPA requires, according to the class action it settled. The suit also alleged that TikTok shared data with third parties. TikTok disagreed with the allegations.

Why it matters: Illinois BIPA requires companies to tell people in writing what biometric data they collect, why, and for how long. Companies must also get a written release first.

What happened: TikTok agreed in February 2021 to pay $92 million and to stop recording biometric information. The settlement covered about 89 million US users.

Lesson: Personalization features that read faces or voices need their own disclosure and written consent, separate from the privacy policy.

6. Snap: face lenses and biometric consent ($35M settlement, 2022)

Snapchat Lenses map face geometry to apply effects, and an Illinois class action alleged that Snap did this without BIPA-compliant consent. Snap denied the claims and said Lenses don't identify individuals.

Why it matters: BIPA requires a written release before a company collects face geometry. Snap's defense turned on whether mapping a face without identifying anyone counts as that kind of collection.

What happened: Snap settled for $35 million in 2022 without admitting liability and added an in-app consent notice.

Lesson: Entertainment features such as face filters can draw BIPA claims, so they need the same consent review as face recognition.

7. Otter.ai: AI notetaker accused of recording meetings without consent (US, 2025 to 2026)

Otter.ai's notetaker joins Zoom, Google Meet, and Microsoft Teams calls to record and transcribe them. According to the consolidated class action, it asks only the Otter account holder for permission. The suit says Otter then keeps the recordings and voiceprints and uses them to train its speech models.

Why it matters: CIPA requires every party's consent to intercept or record a confidential conversation. Illinois BIPA requires a written release before a company collects a voiceprint. Otter denies the allegations.

What happened: In re Otter.AI Privacy Litigation moved forward on August 13, 2026. A federal judge in San Jose let the federal Wiretap Act, CIPA, and BIPA claims proceed. The court found the plaintiffs had plausibly alleged that Otter acts as a third-party eavesdropper, because it keeps and uses conversations for its own commercial purposes.

Motions on whether the suit can proceed as a class action are due in May 2027, with a hearing by September 2027.

Lesson: In all-party consent states such as California, an AI tool that records calls needs consent from everyone on the call. Consent from the customer who installed it isn't enough. The same all-party consent rule drives CIPA lawsuits over website tracking and session replay.

8. Dinerstein v. Google: patient records shared to train health software (dismissed, 2023)

A University of Chicago Medicine patient sued Google and the university over patient records the medical center shared with Google. Google used several years of those records to train software that anticipates health needs. The suit alleged that the records, though labeled anonymized, kept dates of service that Google could combine with its own location data to re-identify patients.

Why it matters: HIPAA's safe-harbor method for de-identifying health records requires removing every part of a patient's dates except the year, including admission and discharge dates. Data that can be linked back to individuals isn't anonymous under the GDPR either.

What happened: The Seventh Circuit affirmed dismissal in Dinerstein v. Google in July 2023 because the plaintiff hadn't shown a concrete injury, so the court never reached the privacy claims.

Lesson: Standing rules can end a US case early, but the re-identification risk stays with the company holding the data. Healthcare AI projects need de-identification that holds up when datasets are combined.

9. Google Assistant: alleged recordings without a wake word ($68M settlement, US, 2026)

Google Assistant switched on and recorded conversations when nobody said its wake word, according to the plaintiffs in In re Google Assistant Privacy Litigation. They say Google shared some of those clips with human language reviewers. The case calls these accidental activations false accepts.

Why it matters: An assistant that records without its wake word captures conversations nobody chose to share. The plaintiffs brought privacy and breach-of-contract claims over it. Those recordings can hold sensitive details about the speaker and anyone nearby.

What happened: Google agreed to a $68 million settlement in January 2026 while denying wrongdoing. The court granted preliminary approval in March 2026 and heard the final approval motion on October 1, 2026.

Lesson: Voice features that can activate by mistake need a visible sign when they start recording and plain disclosure of any human review. Users also need an easy way to review and delete their recordings.

10. Uber: drivers deactivated by fully automated decisions (Netherlands, €825M fine, 2026)

Uber deactivated drivers between 2018 and 2022 using fully automated systems that flagged fraud and low ratings, the Dutch data protection authority found. No person assessed the decisions, and Uber didn't adequately tell drivers the decisions were automated.

Why it's a violation: Article 22 of the GDPR gives people the right not to be subject to solely automated decisions that significantly affect them. Exceptions exist, but they require safeguards such as human intervention. The authority treated losing access to work as that kind of decision.

What happened: The authority fined Uber about €825 million on August 21, 2026, and Uber has appealed. In 2023, an Amsterdam appeals court had called Uber's human review of automated decisions little more than symbolic. The same authority fined Uber €10 million in January 2024 for not telling drivers how long it kept their data.

Lesson: Automated decisions about work, credit, or access need a person who can review and overturn them. The people affected also need to know the decisions are automated.

11. BetterHelp and GoodRx: health data shared with ad platforms through tracking pixels (US, 2023)

BetterHelp shared users' email addresses, IP addresses, and health questionnaire answers with Facebook, Snapchat, Criteo, and Pinterest for advertising, the FTC alleged. GoodRx shared health information with Facebook, Google, and others, the FTC said. The FTC complaints against BetterHelp and GoodRx describe data flowing through tracking pixels.

Why it's a violation: The FTC charged both companies under the FTC Act because sharing health data with advertisers contradicted their privacy promises. GoodRx was also charged under the Health Breach Notification Rule, an FTC rule for health apps and services that HIPAA doesn't cover. Both companies had displayed HIPAA seals that, the FTC said, misled consumers.

What happened: BetterHelp paid $7.8 million for consumer refunds, and GoodRx paid a $1.5 million civil penalty, the first under the Health Breach Notification Rule. Both orders bar the companies from sharing health data for advertising.

Lesson: Health websites and apps need to know what their tracking pixels send to ad platforms. The FTC holds companies to their privacy promises even where HIPAA doesn't apply.

The same pixels drive pixel lawsuits under CIPA. Those suits argue that a pixel sending page activity to a third party, without the consent of all parties, can be wiretapping. Some courts have let these claims proceed and others have dismissed them. Kaiser Permanente agreed in 2025 to a $46 million settlement over tracking on its patient portals while denying wrongdoing.

Enzuzo's consent management platform blocks tracking pixels until a visitor consents, so page activity reaches ad platforms only after the visitor agrees.

12. OpenAI: ChatGPT's training data challenged in Europe and Canada (2023 to 2026)

Italy's data protection authority temporarily blocked ChatGPT from March 30 to April 28, 2023, until OpenAI added privacy notices, an age gate, and opt-outs. Spain and Poland opened investigations, and OpenAI faces complaints about inaccurate answers it gives about real people.

Why it matters: Regulators questioned whether OpenAI had a legal basis to train on personal data scraped from the web. Complaints also argue that ChatGPT's false statements about real people breach the GDPR's rule that personal data be accurate, in Article 5(1)(d).

What happened: Italy's authority fined OpenAI €15 million in December 2024, but a Rome court annulled the fine in March 2026.

In May 2026, Canada's federal privacy commissioner and three provincial counterparts found the complaint well-founded. OpenAI had trained its early models on scraped personal information and user conversations without valid consent. The federal commissioner can't issue fines. The commissioner called the case conditionally resolved after OpenAI began filtering personal information out of training data and committed to further fixes and quarterly reports. British Columbia and Alberta found it unresolved.

Lesson: Training data is personal data when it describes people. Generative AI companies and the businesses that build on large language models need privacy notices and data processing agreements that describe the actual AI data flows.

13. Cox Media Group: alleged fake opt-in for an AI listening ad service ($930,000, US, 2026)

Cox Media Group marketed an AI ad product called Active Listening that it claimed could hear conversations through consumers' smart devices, the FTC alleged. The pitch was that local ads could be targeted based on what people said. The FTC says the listening capability didn't exist and that Cox and two other firms falsely claimed consumers had opted in through app terms.

Why it's a violation: The FTC charged deception under Section 5 of the FTC Act. In its words, "clicking through mandatory terms of service does not constitute 'opt-in consent' for such an invasive service or for use of consumers' voice data from inside their homes."

What happened: Cox Media Group, MindSift, and 1010 Digital Works agreed to pay $930,000 in total, and the Commission finalized the orders in August 2026.

Lesson: Sensitive data such as voice needs its own clear, separate consent request outside the terms of service. Any claim that users opted in also has to match what they actually agreed to.

14. Character.AI: scraped personal data used to pre-train AI models without notice (Italy, €158,000 fine, 2026)

Italy's data protection authority fined Character Technologies, the company behind the Character.AI chatbot, €158,000 in July 2026. The authority found that Character.AI pre-trained its own language models on personal data scraped from the public internet. It didn't tell the Italians in that data, users or not, or give them a workable way to object.

Why it's a violation: The GDPR requires clear notice of how personal data is used. It also gives people a right to object when a company relies on legitimate interests. That legal basis allows processing without consent when the person's rights don't override the company's interests. The authority also faulted the service's English-only privacy notice, weak age verification, and minors' profiles being public by default.

What happened: Besides the fine, the authority ordered Character.AI to fix its privacy notice and age checks and report back within 120 days.

Lesson: Under the GDPR, training on scraped personal data requires notice that reaches the people in it, in a language they understand, and a way to object that works.

15. Grok and X: EU users' posts used for AI training by default (2024 to 2026)

X began using EU users' public posts to train its Grok AI model on May 7, 2024. The setting that allowed it was switched on by default. Many users only learned about it in late July 2024.

Why it matters: X relied on legitimate interests rather than consent. Ireland's Data Protection Commission (DPC) questioned whether that basis, and the notice users got, met the GDPR.

What happened: In August 2024 the DPC took X to the High Court. X agreed to stop processing EU users' posts for Grok training and later made the suspension permanent. The DPC opened a formal inquiry into Grok's training in April 2025.

In February 2026 the DPC opened a second inquiry into Grok generating sexualized images of real people, including children. In March 2026 an Amsterdam court ordered xAI and X to stop Grok from generating non-consensual nude images, under penalty of €100,000 a day.

Lesson: Switching on a new AI use of existing data by default invites regulatory action, as the DPC's court case showed. Clear notice and an opt-out users can act on come before posts or chats feed a model. AI image tools that can depict real people also need safeguards before launch.

16. Grammarly and Superhuman: AI feedback attributed to real writers, allegedly without consent (US, 2026)

Grammarly launched Expert Review for paid users in August 2025. The feature gave AI-generated writing feedback attributed to named writers, academics, and journalists who, the suit says, never agreed to take part. On March 11, 2026, journalist Julia Angwin sued Superhuman, Grammarly's parent company, in a proposed class action. Angwin is a New York Times contributing opinion writer and the founder of Proof News.

Why it matters: The suit claims commercial misappropriation of name and identity, meaning use of a person's name or identity to sell a product without permission. Superhuman says the claims are without merit, and it disabled the feature the day the suit was filed.

What happened: Superhuman has moved to compel arbitration and to dismiss, and the court paused the case until October 15, 2026, while the parties mediate. The complaint puts the amount in controversy above $5 million.

Grammarly's Chrome extension drew separate privacy criticism: a January 2026 Incogni study ranked it among the most potentially privacy-damaging of the AI extensions with more than 2 million downloads. Incogni, a data-removal company, cited the extension's collection of website content, personal communications, and user activity.

Lesson: AI products that attribute output to real people need those people's consent. AI writing tools can also read much of what employees type, so companies need a list of approved tools.

17. OpenAI and Scarlett Johansson: a ChatGPT voice that sounded like her (US, 2024)

OpenAI's Sky voice, one of five ChatGPT voices launched in September 2023, drew comparisons to Scarlett Johansson after the GPT-4o demo on May 13, 2024. Johansson, who voiced an AI assistant in the 2013 film Her, said the voice sounded "eerily similar" to hers. OpenAI CEO Sam Altman had posted the single word "her" during the demo.

Why it matters: Johansson said she declined Altman's request in September 2023 to voice the system, and that OpenAI contacted her again two days before the demo. California's right-of-publicity statute, section 3344 of its Civil Code, bars using a person's voice for commercial purposes without consent. No claim was filed here, and OpenAI says Sky is another actress's natural voice.

What happened: Johansson's lawyers sent two letters to OpenAI asking how Sky was created. OpenAI paused the voice on May 19, 2024, and Altman said he was "sorry" OpenAI "didn't communicate better." Rep. Nancy Mace invited Johansson to testify at a July 2024 House subcommittee hearing on deepfakes, and she didn't testify.

Lesson: No federal statute yet gives people a specific right over AI copies of their voice. The NO FAKES Act, which cleared the Senate Judiciary Committee in June 2026, would create one. Releasing an AI voice that resembles a real person without permission carries legal and reputational risk even when the resemblance is disputed.

18. LinkedIn: behavioral ad targeting without a valid legal basis (EU, €310M fine, 2024)

Ireland's Data Protection Commission fined LinkedIn €310 million in October 2024 over how it processed members' data for behavioral analysis and targeted advertising. The DPC describes behavioral analysis as using information "provided by, inferred from or observed about an individual" to target ads.

Why it's a violation: The DPC found LinkedIn had no valid legal basis. The consent it relied on for third-party data wasn't freely given, informed, specific, or unambiguous, and its legitimate-interest and contract arguments also failed. The DPC also found transparency and fairness breaches.

What happened: LinkedIn appealed both the fine and the findings. In April 2026 the Irish High Court ruled that the two appeals run on separate tracks, and both continue.

Lesson: Inferring traits from user behavior for advertising needs its own valid legal basis. Any consent relied on has to be freely given, informed, specific, and unambiguous.

Ad profiling on websites often starts with tracking tags. Enzuzo's consent management platform blocks those tags until visitors consent and keeps a record of every choice. Book a demo to see it on a live site.

AI data breach examples

An AI data breach is an incident in which an AI product, its vendor, or its users expose personal or confidential data through a security failure.

AI data breach examples include a chatbot app that left about 300 million messages readable, a hiring chatbot whose admin login accepted the password 123456, and an analytics vendor breach that exposed OpenAI API users' details. 

Typical causes are an open database, a default password, or a compromised analytics vendor. Unlike AI privacy violations, which come from decisions about how to use data, a breach comes from a lapse in security or data handling. Breaches of AI products often expose unusually sensitive data, because chatbots store full conversations.

Chat & Ask AI: 300 million chatbot messages exposed (2026)

A misconfigured Firebase backend let anyone read stored conversations in the Chat & Ask AI app, 404 Media reported in January 2026. The researcher who found it said about 300 million messages from more than 25 million users were reachable, including chats about suicide and drug use.

McDonald's McHire: a 123456 password left millions of chatbot records reachable (2025)

An admin login on McHire, the hiring chatbot for McDonald's, accepted 123456 as both username and password, security researchers found in June 2025. Combined with a second flaw, The login let them reach chat records they tied to more than 64 million job applications, including names, emails, and phone numbers.

Paradox said it fixed both flaws within hours and that most of those records weren't tied to a candidate. Only the researchers viewed candidate data, according to Paradox, and only for five people.

OpenAI and Mixpanel: an analytics vendor exposed API users' details (2025)

In November 2025, an attacker exported a dataset from Mixpanel, a web analytics vendor OpenAI used for its API platform. The data included names, emails, approximate locations, and browser details of API users and a limited number of ChatGPT users. Chats, passwords, API keys, and payment data weren't affected, and OpenAI removed Mixpanel from production.

Every third-party script a website loads can send visitor data to another vendor. A breach at that vendor can then expose the site's visitors, as the Mixpanel incident did for OpenAI's API users.

Samsung: engineers pasted source code into ChatGPT (2023)

Samsung semiconductor engineers pasted confidential source code and meeting notes into ChatGPT prompts in three incidents reported in April 2023. The company had lifted an internal ChatGPT ban only weeks earlier. In May 2023, Samsung barred generative AI tools on company devices and networks.

The leaked material was trade secrets, not personal information. Customer data can leak the same way when employees paste it into AI tools.

OpenAI: a ChatGPT bug showed other users' chat titles and payment details (2023)

A bug in an open-source Redis client let some ChatGPT users see other users' chat titles on March 20, 2023, OpenAI disclosed. The same bug may have exposed payment details of 1.2% of ChatGPT Plus subscribers active during a nine-hour window. Those details were name, email, payment address, card type, and the last four card digits.

OpenAI said the number anyone else actually saw was extremely low. Italy's data protection authority cited the ChatGPT bug when it temporarily blocked the service later that month.

Privacy laws that apply to AI

Most AI privacy cases are enforced under general privacy, data protection, and consumer protection laws, through civil regulator decisions and private class actions rather than criminal prosecutions.

GDPR (EU) and UK GDPR

The General Data Protection Regulation applies to organizations established in the EU. It also covers organizations outside the EU that offer goods or services to people in the EU or monitor their behavior. The UK applies its own version, the UK GDPR. Key provisions for AI:

• Legal basis (Art. 6): every processing activity needs one. The EDPB's 2024 opinion on AI models says legitimate interests can support AI development, but only after a case-by-case balancing test.
• Special category data (Art. 9): biometric data used to identify people, health data, and data revealing sensitive traits can't be processed unless an exception applies, such as explicit consent.
• Automated decision-making (Art. 22): people have the right not to be subject to solely automated decisions with significant effects, subject to exceptions with safeguards.
• Data minimization (Art. 5): organizations must not collect more personal data than the stated purpose needs, including when they build or run AI systems.
• Right to erasure (Art. 17): people can ask for their data to be deleted, though removing data already built into a trained model remains technically and legally unsettled.

CCPA / CPRA (California)

The California Consumer Privacy Act, as amended by the CPRA from 2023, applies to for-profit businesses that meet its revenue or data-volume thresholds. Key AI obligations:

• Consumers can opt out of the sale or sharing of personal information, including sharing for cross-context behavioral advertising.
• Consumers can limit the use of sensitive personal information when it's used to infer characteristics about them.
• Regulations on automated decisionmaking technology (ADMT) took effect January 1, 2026. Businesses that use ADMT for significant decisions, such as hiring, lending, housing, education, or health care, must comply by January 1, 2027. They must give a pre-use notice, offer an opt-out, and answer access requests.
• Businesses must run a risk assessment before using personal information to train ADMT for significant decisions or to train facial recognition, emotion recognition, or identity verification technology.

BIPA (Illinois)

Illinois' Biometric Information Privacy Act is the most actively litigated US biometric law. Requirements:

• Written notice and a written release, which can be an electronic signature, before collecting biometric identifiers such as face geometry, voiceprints, fingerprints, and retina or iris scans
• A publicly available retention and destruction schedule
• A ban on selling or profiting from biometric data
• Individuals can sue for damages: $1,000 per negligent violation and $5,000 per intentional or reckless violation

CIPA (California)

The California Invasion of Privacy Act is the main California law behind lawsuits over website tracking and AI tools that record conversations:

• Section 631 bars reading or learning the contents of a communication in transit without the consent of all parties
• Statutory damages are $5,000 per violation or three times actual damages
• SB 690, signed September 30, 2026, ends private claims under section 638.51 from January 1, 2027, including pending suits filed since January 1, 2025. Those claims treat website trackers as pen registers, devices that record addressing data rather than content. Wiretap claims under section 631 continue.

EU AI Act

The EU AI Act sorts AI systems by risk, and its obligations phase in from 2025 to 2028. Regulation (EU) 2026/1744, an amending law adopted in 2026, moved the high-risk deadlines back.

• Prohibited practices (from February 2, 2025): social scoring, untargeted scraping of facial images to build recognition databases, and emotion recognition at work and school. Real-time remote biometric identification in public spaces for law enforcement is banned too, with narrow exceptions. From December 2, 2026, AI systems that generate non-consensual intimate images or child sexual abuse material join the list.
• Transparency duties (from August 2, 2026): people must be told when they're talking to a chatbot, and AI-generated content and deepfakes must be marked. Generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to mark their output.
• High-risk systems: AI used in hiring, credit, education, and other listed areas must meet risk-management, human-oversight, and documentation duties from December 2, 2027. AI built into regulated products follows on August 2, 2028.
• Fines: up to €35 million or 7% of worldwide annual turnover for prohibited practices.
  
  

State AI laws (California, Texas, Colorado, Connecticut)

California, Texas, Colorado, and Connecticut have AI laws that add duties on top of their privacy laws. Other states, including Utah and Illinois, have narrower AI rules. Public enforcers bring all four of these laws, and none lets individuals sue. For the privacy laws underneath them, see the state privacy laws guide.

• California AI Transparency Act (SB 942): generative AI providers must embed hidden provenance data in AI-generated images, video, and audio. They must also offer a free tool to check it.
• Texas Responsible AI Governance Act (TRAIGA): it bans AI built or deployed to push people toward self-harm or crime, to discriminate unlawfully, or to produce sexual deepfakes.
• Colorado (SB 26-189): Colorado repealed and replaced its 2024 AI Act in May 2026. Businesses that use automated decision-making technology for consequential decisions must give notice, explain adverse outcomes, and offer human review. Consequential decisions cover jobs, housing, lending, insurance, health care, education, and government services.
• Connecticut (Public Act 26-15): enacted in May 2026, it adds provenance duties for large generative AI providers, AI companion safety rules, and notices for AI hiring tools deployed from October 1, 2027.

Effective dates and penalties differ widely across the four laws, and so do the officials who enforce them.

AI privacy violations and consent management

AI privacy violations affect consent management wherever website tracking is involved. CIPA, EU cookie rules, and the CCPA all set rules for how a site collects and shares visitor data. Many AI privacy violations start as consent problems: data was collected, shared, or put to a new use without the person's knowledge or agreement.

Others, such as automated decisions without human review, turn on rules unrelated to consent. For the consent side on websites, Enzuzo's consent management platform handles five jobs:

• Blocks tracking pixels such as the Meta Pixel, Google Ads, and analytics tags until a visitor agrees. That matters under CIPA's all-party consent rule and EU cookie rules.
• Applies opt-out choices and Global Privacy Control signals, which the CCPA and other US state laws require businesses to respect.
• Uses Google Consent Mode v2 to tell Google tags each visitor's consent state before they fire, and Google models part of the measurement lost when visitors decline.
• Keeps timestamped records of each visitor's choice, which help a business show consent under GDPR Article 7(1).
• Shows opt-in banners to EU visitors and opt-out notices to most US visitors, based on location.

A consent management platform (CMP) covers consent for website tracking. It doesn't cover biometric collection, model training, or automated decisions, which need their own notices, consent flows, and human review.

Frequently asked questions about AI privacy violations

What are examples of AI privacy issues?

Examples of AI privacy issues include Clearview AI's facial recognition database and Meta's photo-tagging feature. Clearview AI scraped billions of facial images from the internet without consent, and EU regulators have fined it more than €90 million. Meta settled over photo tagging for $650 million under Illinois BIPA and for $1.4 billion in Texas.

What are examples of AI misuse?

Common AI misuse examples include Clearview AI's scraped facial recognition database and Uber's fully automated driver deactivations, both found unlawful by European data protection authorities. Others are still allegations, such as AI notetakers accused of recording meeting guests without consent and Grammarly's AI feedback attributed to real writers. In privacy terms, AI misuse means turning an AI capability on people who had no say in it.

Can AI systems monitor users without consent?

AI systems can't lawfully capture biometrics or record conversations without consent under Illinois BIPA and California's CIPA, but other laws allow some monitoring on different terms. BIPA requires a written release before collecting biometrics, and CIPA requires every party's consent to intercept or record a conversation. In the EU, cookie rules require consent for most non-essential trackers, while the CCPA mostly requires an opt-out.

What are the privacy risks of AI?

The main privacy risks of AI are that it can identify people by face or voice without consent and train on their data without a legal basis. It can also record conversations people never agreed to, decide about them without human review, and leak sensitive data. Tracking pixels that send visitor data to ad platforms add a website-specific risk under CIPA and state privacy laws.

What is the biggest AI privacy violation?

The biggest AI privacy violation by dollar amount is Meta's facial recognition in photo tagging, which Meta settled with Texas for $1.4 billion in 2024. The Texas attorney general called it the largest settlement ever obtained in an action brought by a single state. The largest fine for an AI-driven practice is the Dutch authority's €825 million penalty on Uber in 2026 for automated driver deactivations.

Which privacy laws apply to AI?

The privacy laws that apply to AI are mostly general ones. In the US, the main laws are Illinois BIPA, California's CIPA and CCPA, COPPA for children's data, HIPAA for health data held by covered entities, and the FTC Act. New AI laws in California, Texas, Colorado, and Connecticut add AI-specific duties. In the EU, the GDPR and the EU AI Act apply; the UK applies the UK GDPR.

What consent issues come up with AI training data?

The main consent issue with AI training data is that the people in it rarely agreed to that use. Under the GDPR, data scraped from public websites or collected for another purpose still needs a legal basis. Canada's privacy commissioners found in 2026 that OpenAI trained its models without valid consent. For biometrics, Illinois BIPA requires written consent before any faceprint or voiceprint is collected.

How can a business avoid AI privacy violations?

A business lowers its risk of AI privacy violations by knowing what data its AI tools touch and having a legal basis, often consent, before using it. That means auditing every AI tool and blocking tracking until visitors consent. It also means describing AI data use in the privacy policy and running impact assessments for high-risk uses. An employee policy should set what data can go into AI tools.

What is a DPIA and when is it required for AI?

A Data Protection Impact Assessment (DPIA) is a review of the privacy risks of a new processing activity. GDPR Article 35 requires one when processing is likely to result in a high risk. That includes systematic profiling with significant effects, large-scale processing of biometric or health data, and large-scale monitoring of public areas. High-risk AI also needs an EU AI Act conformity assessment from December 2, 2027.

AI privacy enforcement in 2026

AI privacy enforcement in 2026 brought an €825 million GDPR fine for automated decisions and a federal ruling that let wiretap claims proceed against an AI notetaker. California, Texas, Colorado, and Connecticut also passed AI laws or saw them take effect.

The organizations most exposed are those whose websites, apps, or internal tools:

• Fire tracking pixels or AI-powered analytics before visitors consent
• Collect biometric data such as faces, voices, or fingerprints without a written consent flow
• Let employees paste customer or confidential data into AI tools without a policy
• Make automated decisions about jobs, credit, or accounts with no human review

For the website part of that list, a consent management platform controls which tags fire, and when, so tracking starts only after a visitor agrees.

Book a product walkthrough of Enzuzo's consent management platform to stay protected.