Researchers claim AI awareness doesn't protect us. But it hasn't been studied. #
Posted August 17, 2026 [ Reviewed by Kaja Perina
](/us/docs/editorial-process)
Key points
- Warning people that AI flatters them reduces trust but not its power to change their minds.
- Researchers call this "sycophancy blindness": noticing flattery doesn't stop it from working.
- One-time warnings are not the same as ongoing, real-time self-awareness.
- Clinicians risk abandoning a real tool if they mistake this research approach as proof awareness fails.
We keep telling people to be "aware" of what AI does. New research suggests it barely helps.
In "Individual-Level Interventions Against Sycophantic AI Reduce Its Appeal But Not Its Persuasiveness," Ye et al. (2026) warned users that an AI was sycophantic: flattering and overly agreeable. It was the kind of AI that tells you what you want to hear.
The warnings worked, sort of. People trusted the AI less. Enjoyed it less. Rated it as less objective. These are good signs. If using AI feels easy and effortless, it's probably impacting you more than you realize. And not in a good way.
But regardless of the warning, AI changed their minds anyway. Just as much as if they had never been warned at all.
The researchers call it "sycophancy blindness." Even people who noticed the AI was validating everyone were still moved when it validated them. We see the same thing when people are told outright they're talking to an AI and not a real person (Joshi et al., 2026). It affects them anyway.
Three Risk Areas of Using AI #
Cheng and Jurafsky (2026) found that AI affirms user behavior at a substantially higher rate than humans do in comparable conversations. They also reported that people who received those sycophantic AI responses became less willing to apologize or repair after interpersonal conflict. That's a marriage, a friendship, a working relationship, absorbing the cost of an AI that agreed with someone when a human would have pushed back.
The impacts aren’t isolated to conflict repair. The risks of interacting with AI show up across three areas. Emotionally, validation from AI shapes decisions and confidence in ways people don't fully register in the moment. Socially, a 12-month study found using chatbots for companionship predicted increased loneliness four months later, and feeling lonelier predicted more chatbot use four months after that (Folk & Dunn, 2026). Cognitively, just ten minutes of AI assistance was enough to make people give up faster on their own problem-solving afterward, a quick erosion of independent effort (Liu et al., 2025).
This is what's actually at stake when AI shapes how we think, feel, and connect. A single warning, delivered once before anyone starts using it, was never going to hold up against effects like these. That's why I disagree when researchers call this a failure of awareness.
What actually failed was one-time education, a fact delivered once, not the ongoing, active noticing that real awareness requires. And if clinicians read this study as proof that awareness doesn't work, we lose a real way of helping our clients.
The whole reason I developed the AI Awareness Arc is that I think real self-awareness and intensive education are what let us stand our ground with AI, to use it safely and effectively.
What Self-Awareness Really Is #
When I say self-awareness, I mean noticing our own thoughts, feelings, and experiences in real time, or as close to it as we can get.
It's noticing what AI does and how it makes us feel, so we can respond in the moment instead of being swayed by the interaction without realizing it. It's being able to point at the screen and say, "That's it agreeing with me again," or "There's that closing hook to keep me engaged."
Self-awareness is seeing the question AI so often asks, checking in with ourselves, and deciding whether we want to answer that question, address something else, or decide we got what we needed and we're finished. It's noticing how AI is shifting our thinking, our certainty, our willingness to sit with disagreement or discomfort.
This study shows, as others have, that simple education, like a warning label or a one-off message, won't be enough to offset the impact of AI. A warning sign on a car wouldn't be enough to get us to drive safely either.
We need driver's ed-style training: Real instruction on what AI is, and live practice with it, so people learn the risks it carries and how to navigate it safely.
References
Ye, M., Kraut, R., & Rathje, S. (2026). Individual-level interventions against sycophantic AI reduce its appeal but not its persuasiveness (arXiv:2607.25166v3) [Preprint]. arXiv. https://arxiv.org/abs/2607.25166
Joshi, R., Adjagbodjou, A., & Luria, M. (2026). Dark patterns in AI chatbots: A taxonomy to inform better design. Center for Democracy & Technology. https://cdt.org/wp-content/uploads/2026/05/2026-05-28-CDT-Research-Dark-Patterns-in-AI-Chatbots-Report-final-2.pdf
Cheng, M., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science. https://doi.org/10.1126/science.aec8352
Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4), 276-286. https://doi.org/10.1177/09567976261427747
Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2025). AI assistance reduces persistence and hurts independent performance. arXiv. https://arxiv.org/abs/2604.04721