The Center for Democracy and Technology just dropped a report that’s making a lot of AI companies uncomfortable. CDT exposes 37 dark patterns in ChatGPT and Claude — and honestly, reading through it, some of what they found is hard to argue with.
This isn’t a fringe tech blog raising alarms. CDT is a Washington D.C.-based nonprofit that’s been doing digital rights work since 1994. When they publish something like this, people pay attention.
So what did they actually find? And more importantly — does it change how you should be using these tools?
What the CDT Report Actually Says About ChatGPT and Claude
The CDT defines dark patterns as design choices that push users toward actions they didn’t consciously choose — or away from ones that protect their privacy. These aren’t bugs. They’re deliberate interface decisions.
Across both OpenAI’s ChatGPT and Anthropic’s Claude, CDT researchers catalogued 37 distinct instances. They grouped these into categories most people wouldn’t even think to look for: default settings that favor data collection, vague consent language that sounds reassuring without actually telling you anything, and upgrade prompts placed at exactly the moment you’re most frustrated with a free-tier limitation.
The report hits hardest on a few specific areas. Memory features that opt users in by default. Consent flows that use confusing language around how your conversations are used for model training. Emotional design — and yes, that’s a real term CDT uses — where AI systems are built to feel warm, trustworthy, and relationship-like in ways that may lower your guard before you’ve actually agreed to share sensitive information.
What surprised me going through the specifics was how many of these patterns feel completely normal until someone names them. That’s the whole point.
The 5 Patterns Worth Actually Caring About
Look, CDT listed 37. Not all of them carry equal weight. Some are edge cases. But five of them genuinely affect how you should interact with AI tools going forward.
Default memory opt-ins
Both ChatGPT and Claude have rolled out memory features — Claude calls it Projects, OpenAI has its own Memory toggle. The issue CDT flags: these features are often enabled by default or presented in a way that makes opting out feel like you’re losing something valuable rather than protecting yourself.
In practice, most people just hit “accept” and move on. That’s not informed consent. That’s interface pressure.
Training data consent that buries the real question
Here’s the one that trips people up most. When you start using ChatGPT or Claude, you’re walked through a setup or terms flow. The language around whether your conversations are used to train future models is technically there — but it’s written in a way that requires you to already know what you’re looking for.
I’ve read both sets of terms more than once. The honest truth: even knowing what I was looking for, it took a few passes to find the exact toggle that actually stops your data from being used for training. That’s not an accident.
Emotional persona design as a consent-lowering mechanism
This one is more subtle and, to be fair, more contested. CDT argues that making AI feel warm, empathetic, and relationship-like is a design choice that can lead users to share more than they’d share with a cold, transactional interface.
They’re not wrong, exactly. There’s real research on this from MIT and Stanford showing people disclose more to systems that respond with apparent empathy. Whether that crosses the line into manipulation is a legitimate debate — but the mechanism is real.
Friction asymmetry in privacy settings
Opting into features? One click. Opting out? Settings menu, sub-menu, toggle, confirmation. CDT documents this across both platforms. The path of least resistance consistently leads toward more data sharing, not less.
Upgrade prompts at cognitive peak frustration
This one is more about money than privacy, but it’s still a dark pattern. Both platforms are designed to surface paid plan prompts exactly when you’ve hit a usage limit mid-task. You’re annoyed, you’re in the middle of something, and the upgrade button is right there. That timing isn’t coincidental — it’s conversion design.
Why These Companies Do This (And Why the Answer Is Boring)
Neither OpenAI nor Anthropic is some cartoonishly evil operation trying to steal your secrets. The reality is much more ordinary.
These are companies with enormous infrastructure costs, investors expecting returns, and product teams optimizing for engagement and conversion metrics. Dark patterns emerge when “user engagement” becomes the primary design goal and “user understanding” becomes secondary.
Anthropic in particular positions itself as the safety-focused alternative in AI. Claude’s whole identity is built around being more trustworthy, more careful, more ethical than competitors. So the CDT findings hit differently for them — it’s harder to shrug off when your brand is “we’re the responsible ones.”
OpenAI’s response to similar critiques in the past has been to update terms and add more settings pages. Whether that addresses the root issue or just adds more complexity to an already complicated opt-out process is worth thinking about.
What This Means If You’re Using Claude for Serious Work
If you’re using Claude for business tasks — writing, research, analysis, coding — the way you set up your workspace matters more than most people realize. This isn’t just about privacy in the abstract. It’s about knowing what you’re agreeing to when you paste client data, business strategy, or sensitive research into a chat window.
Here’s what to actually do:
Go into your Claude settings and look for the data controls section. Find the toggle for whether your conversations are used to improve Anthropic’s models. Turn it off if you’re handling anything you wouldn’t want appearing in future training datasets.
For ChatGPT, the equivalent is under Settings > Data Controls > Improve the model for everyone. Default: on. Most people never change it.
If you’re on Claude’s free tier, your options are more limited. Paid tiers give you more explicit control. That’s not a coincidence either — it’s worth knowing before you decide which tier makes sense for your use case.
The Grok Comparison Nobody’s Making
Here’s something the CDT report doesn’t cover but probably should: Grok, xAI’s model built into X (formerly Twitter), has its own set of concerning default behaviors — particularly around using your X posts for training data. The way Grok handles voice mode and data permissions follows similar patterns to what CDT found in ChatGPT and Claude.
The point isn’t that one platform is definitively worse. It’s that this is an industry-wide design philosophy, not a company-specific scandal. If you’re picking tools based partly on privacy posture, you need to check settings on every platform individually — don’t assume the “ethical” branding means the defaults are set in your favor.
The Local Model Angle (For People Who Want to Actually Opt Out)
After reading the CDT report the first time, I went back and started using local models more for anything sensitive. Not as a paranoid reaction — just as a practical recalibration.
Running local models with tools like Cursor means your conversations never touch an external server. No training data questions. No consent flows. No upgrade prompts. The tradeoff is capability — local models at the consumer level are still catching up to GPT-4o and Claude Sonnet — but for tasks where privacy matters more than raw power, it’s worth considering.
The CDT report, whether intentionally or not, makes a good case for why local AI options deserve more attention than they currently get.
What “Dark Pattern” Actually Means in Legal Terms (And Why That Matters)
CDT isn’t just writing opinion pieces. Their framing has regulatory teeth.
The FTC has been increasingly aggressive about dark patterns since 2022, publishing enforcement guidelines and taking action against companies like Amazon and Vonage for deceptive opt-out flows. The EU’s Digital Markets Act and GDPR both have specific provisions that dark pattern research like CDT’s can be used to support enforcement actions.
So when CDT publishes a report documenting 37 patterns across two of the biggest AI companies in the world, they’re building a record. That record can feed into regulatory complaints, congressional testimony, and enforcement actions. OpenAI and Anthropic both have legal teams that understand this — which is probably part of why you’ll see both companies quietly update terms and settings in the months after a report like this lands.
Watching what actually changes (and what doesn’t) in the next 6-12 months will tell you a lot about which findings have real legal exposure.
The Honest Defense of Both Platforms
Let me be real about something: some of what CDT flags exists in a gray area.
Warm, conversational AI personas aren’t inherently manipulative. Most users actively prefer interacting with an AI that sounds like a person rather than a command-line interface. Calling that “emotional manipulation” is a stretch in many contexts.
Memory features that are on by default? A lot of users genuinely want that — they want Claude or ChatGPT to remember their preferences without having to configure anything. Opt-out defaults exist partly because opt-in defaults result in most users never enabling useful features.
The line between “convenient default” and “manipulative default” is real but not always obvious. CDT draws that line in a specific place based on their framework. Reasonable people can disagree on some of the 37 specific instances.
What’s harder to defend: consent language that’s intentionally vague. Friction asymmetry in privacy settings. Upgrade prompts timed to peak frustration. Those are choices that don’t have a user-benefit explanation that holds up under scrutiny.
The Privacy Setup You Should Actually Do Right Now
Don’t wait for these companies to fix their defaults. Here’s a 10-minute audit you can do on both platforms today.
For Claude:
- Settings > Privacy > Data Usage — find the model training toggle
- Check whether you’re in any active Projects with memory enabled
- If you use Claude for work, consider whether you’ve ever pasted anything you’d be uncomfortable with Anthropic storing
For ChatGPT:
- Settings > Data Controls > Improve the model for everyone — turn off
- Check Memory settings and review what’s stored
- Temporary Chat mode exists for conversations you don’t want stored — use it by default for sensitive tasks
Takes about 10 minutes. Worth doing before your next session, not after.
If You Care About Uncensored or Private AI Options
The CDT report is also indirectly making the case for AI tools that are architecturally private by design — not just policy-private. Platforms built around uncensored, private AI image and text generation approach the data question differently from the ground up.
It’s a real distinction. Policy-based privacy (“we promise not to use your data”) and architecture-based privacy (“your data never leaves your device”) are fundamentally different things. Most mainstream AI products — ChatGPT, Claude, Gemini — are in the first category. The CDT report is essentially an argument for why that distinction matters.
Venice AI and the Alternative Model
One platform CDT didn’t examine that’s worth mentioning: Venice AI, which explicitly markets itself around privacy-first design. Venice AI’s free tier and model options operate on a different architecture philosophy than ChatGPT or Claude — conversations aren’t stored on their servers by default.
That doesn’t mean Venice is perfect or that it should replace Claude for serious work. Capability gaps are real. But as the CDT report highlights, knowing your options across the AI tool ecosystem isn’t paranoia — it’s just smart.
What Happens Next With This Report
CDT’s findings will likely do a few things over the next year.
Expect both OpenAI and Anthropic to publish blog posts emphasizing their commitment to user privacy and update some settings to be slightly less friction-heavy for opt-outs. This is the standard playbook — acknowledge the category, update the surface, avoid admitting any specific finding was wrong.
Expect this report to appear in FTC briefings and potentially EU regulatory discussions. The EU in particular has been aggressive about dark patterns since the Digital Services Act went into effect.
Expect more independent audits like this one. CDT isn’t alone in this space — Electronic Frontier Foundation, AI Now Institute, and Mozilla Foundation are all doing similar work. The more reports stack up, the harder the “we take privacy seriously” PR response becomes to sustain.
And honestly? Expect the underlying business incentives to change slowly. Companies optimize for what they’re measured on. Until regulators make dark patterns genuinely costly, the optimization pressure will remain.
The Bigger Question Nobody’s Asking
Here’s the thing the CDT report doesn’t quite get to, but implies: why do we expect AI companies to self-regulate on design choices that directly affect their revenue?
Memory features, training data, emotional design — all of these create value for the platforms. Memory increases engagement and retention. Training data improves models. Emotional design increases time-on-platform. These aren’t side effects of dark patterns. They’re the point.
Asking OpenAI or Anthropic to voluntarily make opt-out flows easier is like asking a casino to make the exit more visible. They’ll do the minimum required and call it a day.
The CDT report is valuable not because it expects voluntary change, but because it builds the evidentiary record for mandatory change. That’s how regulatory pressure actually works.
What You Should Actually Take Away From This
The CDT report is worth reading directly if you use either platform for anything sensitive. The full document is publicly available through CDT’s website.
But here’s the practical summary: don’t assume the defaults are set in your favor. They’re not. Spend 10 minutes in your settings on both platforms, turn off the data training toggles, and use temporary or private conversation modes when handling anything you care about.
If you’re building on top of these APIs for business use, the data handling questions get more serious — not less. Know what your usage agreements actually say, not just what the marketing copy implies.
And keep an eye on what changes in the next 12 months. The companies that respond to reports like this with genuine architectural changes versus PR updates will tell you a lot about which ones are actually serious about the trust they claim to be building.
The tools themselves? Still genuinely useful. Claude and ChatGPT are impressive in ways that aren’t hype. But “useful” and “designed in your interest” aren’t the same thing. The CDT report is a useful reminder to hold both thoughts at once.
