Your night out is not a data point

Most of the web is funded by real-time bidding (RTB): the instant you open a page, a profile of you (what you’re reading, where you are, an ID that follows you around) is broadcast to an auction so advertisers can bid on your attention.

The UK’s own data regulator examined this machinery in 2019 and concluded that the processing of sensitive-category data inside it was unlawful, that the consent popups of both major frameworks (IAB Europe’s TCF and Google’s Authorized Buyers) were non-compliant, and that a single bid request shares a detailed profile with hundreds of organisations without the person’s knowledge: something the ICO called disproportionate, intrusive and unfair [1].

The categories traded in those auctions are not abstract. The industry’s own taxonomies include labels like “Mental Health”, “Sexual Health”, “Reproductive Health”, “Substance Abuse”, “Politics” and “Ethnic & Identity Groups” [1].

The Irish Council for Civil Liberties calls it “the biggest data breach ever recorded”. The industry disputes the framing; the numbers went unchallenged [2].

Downstream of the auctions sit the data brokers. A 2023 Duke University study contacted 37 US brokers about mental-health data; 11 were willing and supposedly able to sell lists of people with depression, anxiety and other conditions, priced from roughly $275 per thousand aggregated records up to $75,000 to $100,000 a year for subscription access. At least one broker advertised fully identified individuals: names and postal addresses of people with depression, bipolar disorder, PTSD and OCD, linked to race and ethnicity [3].

And the collectors are often the companies you trusted most. The US FTC alleged that Avast, an antivirus vendor, sold its users’ granular, re-identifiable browsing histories after promising the product would protect their privacy. It alleged that location broker X-Mode ingested over 10 billion location points (advertised as accurate to 20 meters) and sold them to private government contractors, and that InMarket harvested precise location from roughly 100 million devices a year to sort people into segments like “parents of preschoolers”, “Christian church goers” and “wealthy and not healthy” [4].

In late 2024 the FTC’s Mobilewalla case confirmed what privacy researchers had long said: participants in the ad auctions simply keep the bidstream data about everyone in the auction, whether or not they win the ad, revealing visits to healthcare facilities, churches and labor unions [5].

On paper, the law has teeth. GDPR allows fines up to 4% of global revenue, and European authorities have issued 3,195 publicly known fines totalling about €6.31 billion since 2018 [6]. The headline case is the €1.2 billion fine against Meta in May 2023, for data transfers the European Data Protection Board called “systematic, repetitive and continuous” violations [7] [8].

Now look closer. Meta appealed, and payment is suspended while the appeal runs. Amazon’s €746M fine, the second largest ever, was annulled on procedural grounds in 2026. Of the $3.26B in fines the Irish regulator had levied, only $19.9M (0.6%) had actually been paid as of December 2024 [9]. British Airways’ £183M fine was negotiated down to £20M. The ICCL’s five-year review found GDPR enforcement “largely paralysed” against big tech [9] [10], and even the UK’s own Information Commissioner has said fines are not an effective way of keeping big tech in line [9].

On the US side, the Avast case ended in a $16.5M settlement, against years of selling the browsing data of hundreds of millions of users, and X-Mode and InMarket settled by consent order with no fine at all [4]. FTC Chair Lina Khan’s 2024 statement that “businesses by default cannot sell people’s sensitive data” was a genuinely important marker, but it was established through settlements in which nobody admitted anything.

The pattern is consistent: the violations are industrial-scale and continuous; the penalties arrive years later, get reduced, get appealed, or never get paid. For a trillion-dollar company, this is not deterrence. It’s a line item.

Facebook and Cambridge Analytica (2014 to 2018). A personality-quiz app harvested the profiles of up to 87 million people (only about 270,000 of whom ever installed it; the rest were their friends) and passed the data to a political-targeting firm. The punishment: a record $5 billion FTC penalty in 2019, plus $100M to the SEC for misleading investors, plus £500,000 from the UK ICO (the pre-GDPR statutory maximum, paid without admission of liability). The FTC’s own dissenting commissioner pointed out that Facebook’s stock rose when the $5B figure leaked: “the market believes that a penalty at this level makes a violation profitable” [19].

Google Street View, “Wi-Spy” (2007 to 2010). Street View cars quietly captured payload data from unencrypted home Wi-Fi networks in more than 30 countries: emails, passwords, URLs. The punishment: a $25,000 FCC fine, and not for the collection itself but for obstructing the investigation; ProPublica calculated it as about 68 seconds of Google’s profit. Later: $7M to 38 US states, a $13M class settlement paid to privacy nonprofits (affected users got $0), €100k in France, €145k in Germany [20].

Google location tracking (2014 to 2020). An AP investigation in 2018 showed that turning “Location History” off did not stop Google storing your location; a default-on setting kept collecting it anyway. The punishment: $391.5M to 40 US states (2022, then the largest multistate privacy settlement in US history), $85M to Arizona, and $1.375 billion to Texas (2025), the largest privacy recovery ever obtained by a single state. Google admitted no wrongdoing in any of them and said the Texas deal required no product changes [21].

Amazon Alexa and Ring (2023). The FTC alleged Amazon kept children’s Alexa voice recordings indefinitely and undermined parents’ deletion requests, and that Ring gave thousands of employees and contractors unrestricted access to customers’ home camera footage, including bedrooms; one employee viewed thousands of videos of female users. Ring also skipped basic protections while about 55,000 accounts were hijacked, some hackers taunting people through their own cameras. The punishment: $25M (Alexa) and $5.8M (Ring). Combined, about 0.005% of Amazon’s revenue that year [22].

Meta’s 533-million-user scrape (2018 to 2021). Phone numbers and personal data of about 533 million Facebook users, siphoned through the contact-import feature, were posted free on a hacking forum. Meta’s public response was that the data was obtained “not through hacking our systems but by scraping”, and it declined to notify the affected users. The punishment: a €265M Irish fine for failing data-protection-by-design [23]. Remember this one whenever a company says scraping is harmless; it’s also the mechanism feeding AI training sets.

Meta’s facial recognition (2011 to 2021). Facebook ran face-templating on users’ photos without the written consent Illinois law requires, and Texas found the same for its residents’ biometric data. The punishment: a $650M Illinois class settlement (2021) and $1.4B to Texas (2024) [24]. The products had already served their purpose for a decade.

Apple Siri “grading” (2019). Contractors reviewing accidental Siri activations regularly heard confidential recordings: medical details, drug deals, people having sex, tied to location and contact data. The punishment: Apple suspended the program after the Guardian exposed it, and settled a class action for $95M (2024 to 2025), roughly $20 per affected device, admitting no wrongdoing [25].

TikTok’s China transfers (2020 to 2025). The Irish DPC found TikTok sent EEA users’ data to China without the required protections and hid it: the 2020 to 2022 privacy policy didn’t even name China. The punishment: €530M (2025). TikTok is appealing; payment is on hold. Its earlier COPPA offence (collecting children’s data as Musical.ly) had cost it $5.7M [26].

Uber’s breach cover-up (2016 to 2017). After hackers took data on 57 million riders and drivers, Uber paid them $100,000 disguised as a bug bounty and kept the breach secret for a year. The punishment: $148M to all 50 US states, and, remarkably, an actual criminal conviction: the security chief who ran the cover-up got three years of probation and a $50,000 fine, the first executive ever convicted over a breach response [27]. One probation sentence in a decade of scandals; that’s the ceiling so far.

Meta’s consent shell game (2018 to 2025). In January 2023 Meta was fined €390M because “you agreed to ads by accepting the terms of service” is not a legal basis for behavioural advertising. The revealing detail: Ireland’s own regulator had sided with Meta and proposed roughly €59M; the European board of regulators had to overrule it and force the finding. Meta’s response was not to stop, but to hop legal bases: from “contract” to “legitimate interest” (killed by the EU Court of Justice in July 2023), to “consent”, to the “pay or okay” wall in November 2023. The EDPB said in 2024 that this binary generally isn’t valid consent either, and the European Commission fined the model €200M under the DMA in April 2025 [28]. Seven years of litigation, five fines, and behavioural advertising never stopped for a single day.

The children’s cases (2022 to 2023). Instagram was fined €405M after children’s accounts were public by default and business accounts of 13-to-17-year-olds published their email addresses and phone numbers to the world. TikTok was fined €345M for the same public-by-default design for minors, plus a “Family Pairing” feature that let any unverified adult link to a child’s account and loosen its settings, plus dark patterns nudging kids toward public posting [29]. In both cases, and in WhatsApp’s €225M transparency fine before them (where Ireland’s draft proposed €30 to 50M and the EDPB forced a quadrupling), the Irish regulator’s draft had to be overridden by the rest of Europe [29] [30].

France just fines the cookie banner itself. The CNIL, using ePrivacy rules that bypass the Irish bottleneck, issued the first mega-fine of the GDPR era (€50M against Google in 2019, for burying consent, upheld on appeal) and then went after the banners: €100M and €150M against Google and €60M against Facebook for designs where accepting tracking took one click but refusing took several. Amazon took €35M for dropping cookies before anyone consented at all. Adtech firm Criteo took €40M for being unable to prove the consent it claimed to have [31].

The consent popup system was itself ruled illegal (2022 to 2025). The Belgian regulator found that the IAB Europe Transparency & Consent Framework, the machinery behind most consent popups on the European ad-funded web, violated the GDPR: the very system asking for your consent was unlawfully processing your data. The EU Court of Justice confirmed the core findings in 2024 and the Belgian Market Court upheld the decision in 2025. The fine for building the consent infrastructure of an entire continent on an unlawful basis: €250,000 [32].

Clearview AI, or what happens when a company simply refuses (2022 onward). Clearview scraped billions of face photos into a search engine for law enforcement. Italy, Greece and France each fined it €20M, France added a €5.2M non-compliance penalty, and the Netherlands added €30.5M: over €100M in total. As of the latest reporting, Clearview has paid none of it. It has no European establishment, so there is nothing to seize; the Dutch regulator is now openly exploring whether it can hold the company’s directors personally liable, precisely because fining the company achieves nothing. The UK’s £7.5M fine spent years in tribunals over whether the regulator even had jurisdiction [33].

Amazon’s €746M: the fine that confirmed the violation and then vanished (2021 to 2026). Luxembourg’s regulator fined Amazon €746M in 2021 over ad targeting without a valid legal basis; the public only learned of it through Amazon’s SEC filing, because Luxembourg’s secrecy rules kept the decision confidential. In March 2026 Luxembourg’s administrative court agreed the violations were real and annulled the fine anyway, on procedural grounds, sending the case back to the regulator [34].

Read the pattern across the whole file. In America, the harvesting runs for years, the profit is taken up front, and the settlement arrives later at a fraction of the gain, with nobody admitting anything. In Europe, the law is stronger and the findings are damning, but the fines are appealed, reduced, annulled or simply never paid, while the practice continues under a new legal theory. Different continents, same arithmetic.

Large language models are trained on scraped copies of the public internet, and the personal data in those scrapes does not stay “inside” the model. The foundational research showed that an attacker can extract verbatim training data from a deployed model, including names, email addresses and phone numbers, even when that data appeared in a single document [11]. Follow-up work in 2023 extracted gigabytes of training data from production models, and found a “divergence attack” that made aligned ChatGPT emit memorized training data at 150 times its normal rate; the authors’ conclusion was blunt: current alignment techniques do not eliminate memorization [12]. Worse, memorization scales: bigger models retain more of what they ingested [11] [13].

Legal scholars Daniel Solove and Woodrow Hartzog call this era “the great scrape”, arguing that mass scraping for AI training violates nearly every core principle privacy law is built on: consent, transparency, purpose limitation, data minimization [14]. Regulators agree in substance. Italy’s Garante briefly banned ChatGPT in 2023 and later fined OpenAI €15 million after finding it had no lawful basis for the personal data used to train the system (OpenAI is appealing) [15]. The EDPB’s December 2024 opinion set the bar plainly: an AI model is only “anonymous” if extracting personal data from it is an insignificant risk, and in practice LLMs will generally not qualify [15].

Clearview AI, which scraped billions of face photos, has collected fines across Europe, including €30.5M from the Dutch regulator [9]. And yet, as of this writing, every major model trained on scraped personal data is still running. The playbook from the adtech era repeats: scrape first, litigate later, keep the product.

Our platform is built on one principle: the software that runs the venue lives on hardware the venue owns, and every external service is an opt-in connector we can survive losing.

The honest version of privacy work is boring. It’s deciding, table by table and script tag by script tag, that the convenient default isn’t good enough, because the convenient default is the machinery described above.

• No trackers, no adtech, no analytics SaaS. Your visit to our agenda is not an auction. Nothing on this site broadcasts you to hundreds of companies, because nothing here speaks RTB.
• Guest checkout, minimal data. Buying a ticket takes a name and an email address. No account, no profile, no enrichment. Your tickets come back through a private link. The single third-party script on the whole site is the payment form, loaded only on the checkout page, configured without the site-wide tracking beacons. Card processing is the one genuinely unavoidable third party (that’s how card networks work), and processors are treated as what the law says they are: contracted sub-processors, not data partners.
• Big-tech services stay server-side, behind interfaces we can swap. Where we do use one (address autocomplete uses Google’s geocoding), your browser never talks to them: requests go through our backend, cached on our side, swappable for a self-hosted OpenStreetMap service without touching anything else. Google sees a server IP, not your session.
• Your calendar, not our funnel. Subscribe to our agenda feed once and every show lands in your own calendar. Your calendar app polls our server; no intermediary learns who you are.
• Metadata scrubbed by default. Every image we serve is a processed copy with the EXIF stripped: no GPS coordinates or device fingerprints riding along.
• GDPR as a floor, not a ceiling. Export, erasure and consent are designed into the data model from day one, not bolted on after a complaint. Credentials are encrypted and write-only. Audit logs record who did what, never your behaviour.
• And nothing you give us trains an AI. Your data has one job: getting you into the show.