This cybersecurity news safari documents dark patterns in AI products: silent AI-training opt-ins, buried privacy settings, manipulative defaults. Real cases, named vendors, plain language.
The documented cases reveal a pattern across AI products in 2026: consent flows designed to maximize data collection while minimizing user awareness. Silent opt-ins for AI training, settings buried three menus deep, and oversized accept buttons. Naming them is the first step to regulating them.
A gallery of real dark patterns reported in AI product onboarding. We only include cases with published reporting we can cite — sources are linked under each entry. We describe the interfaces in words; we never hotlink screenshots from other sites.
X (formerly Twitter), training its Grok AI model
In July 2024, X activated a setting giving itself permission to use all users' posts, interactions, inputs and Grok results for "training and fine-tuning" — with every user opted in by default and no notification. The toggle was buried several layers deep (Settings → Privacy and safety → Grok) and could only be disabled from the web, not the mobile app. The story broke when a user spotted the hidden setting and posted about it. Weeks later, privacy group NOYB filed GDPR complaints in nine European countries over X's use of over 60 million users' personal data for AI training without consent.
Sources: MacRumors, 26 July 2024 — "X Training Grok AI On Tweets Without Notifying Users"; Malwarebytes, August 2024 — "X accused of unlawfully using personal data of 60 million+ users to train its AI".
Slack (Salesforce)
In May 2024, reporting revealed Slack's Privacy Principles stated its systems analyze customer data — messages, content and files — "to develop AI/ML models". The training was opt-out, but the opt-out was engineered for maximum friction: an individual user could not do it at all — only a workspace owner could email Slack's support team with the workspace URL and a specific subject line ("Slack Global model opt-out request"). The policy had sat quietly in place for months before the backlash. Slack then clarified the data went to non-generative ML models (channel/emoji recommendations, search) using de-identified aggregate data, not to its generative AI products — but the opt-out burden on users remained.
Sources: Engadget, 17 May 2024 — "Slack has been using data from your chats to train its machine learning models".
Zoom
In March 2023 Zoom updated its terms of service to grant itself broad rights over customer content — including audio, video and chat — for purposes including "machine learning and artificial intelligence, including training and tuning of algorithms and models". The change went mostly unnoticed until August 2023, when reporting triggered a wave of backlash, with users calling for a boycott. Zoom responded with a blog post promising it would not use audio, video or chat content for AI training without consent, and added that line to the terms. However, the terms still allowed AI training on "service generated data" (usage and behavioral telemetry) without additional permission, and critics noted that once a meeting host opted in, other participants had to leave the meeting to avoid consenting.
Sources: Africa Check, 1 September 2023 — "Zoom updates terms after media backlash, saying it won't use your data to train AI".
LinkedIn (Microsoft)
In August 2024 LinkedIn quietly introduced a "Data for Generative AI Improvement" setting that automatically opted users into having their personal data and content used to train its generative AI models — with no prior announcement. The following month it updated its privacy policy to disclose the practice. The sequence drew a class-action lawsuit in a California federal court alleging the company shared Premium subscribers' data with third parties for AI training without proper consent, and that the policy update was an attempt to cover the earlier data use. LinkedIn has denied the allegations; the lawsuit's claims are allegations, not findings.
Sources: Fast Company — "LinkedIn plans to use your data to train AI"; The Arabian Post — "LinkedIn faces legal challenge over alleged unauthorized use of user data for AI training".
Every case above follows the same playbook: assume consent, bury the toggle, announce nothing. The fix is always the same too: explicit opt-in, plain language, and the off-switch where users can actually find it. AI terms are changing faster than anyone can track — that's exactly why we built tools that watch them.
All descriptions are based on the cited published reporting. Products may have changed their policies since publication — check the sources for the latest state. This page is an educational guide by Ingenious AI, not legal advice.
Interface designs that trick or pressure users into choices they would not otherwise make — like hidden opt-ins or impossible opt-outs.
Silent AI-training opt-ins, buried data-sharing toggles, and take-it-or-leave-it terms updates — each entry in the gallery above is a real, cited case.
Increasingly: regulators in the EU and several US states now treat deceptive design as an unfair practice, with fines attached.