Americans against behavioral surveillance at scale without psychosocial validation.
AI platforms profile how you think using behavioral classifiers built on pre-print research (non-peer reviewed) that treats the user as the risk variable instead of the algorithm. The methodology behind these systems has no psychosocial validation, no disclosed error rates, and no recourse for the people it profiles. On March 25, 2026, a jury found Meta and YouTube negligent in the design of their recommendation systems, confirming that liability sits on the design side, exactly where these classifier studies refuse to look. People are filing consumer protection complaints with their state attorneys general.
The behavioral classifiers deployed by major AI platforms trace back to a single pre-print: Phang et al. (2025), produced by researchers at OpenAI and MIT Media Lab. The study analyzed approximately 3 million ChatGPT conversations and built an LLM-based system called "EmoClassifier" to categorize users by psychological and emotional patterns. It has not been peer reviewed.
The study treats the user as the risk variable. It profiles conversation patterns to flag emotional states, but never examines whether the platform's own design choices contribute to those patterns. There are no disclosed false positive rates, no psychosocial validation of the classification taxonomy, and no recourse mechanism for people who are incorrectly profiled.
Peer-reviewed research directly contradicts this approach. Ophir et al. (2025), published in Frontiers in Medicine, critiques the methodology behind classifiers like EmoClassifier, arguing that treating user behavior as the signal while ignoring platform architecture produces unreliable and potentially harmful classifications. Kulveit et al. (2025), peer-reviewed at ICML 2025, demonstrates that the risk sits in the architecture, not in user behavior.
The conflict of interest is structural. Pattie Maes, a senior MIT author on Phang et al., is also the creator of Firefly (1995), a direct ancestor of modern recommendation engines, The study was conducted in collaboration with OpenAI, on OpenAI's own user data. The researchers are evaluating the behavior of users on a platform built by their collaborator, using infrastructure they do not independently control, without independent oversight.
On March 25, 2026, an LA Superior Court jury found Meta and YouTube negligent in the design of their recommendation systems (L.H. v. Meta Platforms, Inc., et al., Case No. 22STCV21244). The verdict places liability on the design side of algorithmic systems. Behavioral classifiers built on the Phang methodology sit in exactly the space the jury identified: profiling users to shape their experience, with no independent validation that the profiling itself is sound.
Phang, J., et al. (2025). arXiv:2504.03888 (pre-print, not peer reviewed)Ophir, Y., et al. (2025). Frontiers in Medicine. doi:10.3389/fmed.2025.1612838Kulveit, J., et al. (2025). ICML 2025 (peer reviewed).Sharma, M., et al. (2026). Anthropic. 1.5M Claude.ai conversations.L.H. v. Meta Platforms, Inc., et al. LA Superior Court, Case No. 22STCV21244 (March 25, 2026).
What are behavioral classifiers?
A behavioral classifier is a system that watches how you interact with a platform and assigns you to categories based on those patterns. In the context of AI chat platforms, this means analyzing your conversations to make inferences about your emotional state, psychological profile, or risk level, then using those inferences to alter how the system responds to you.
This is different from content moderation, which filters specific harmful content. Behavioral classifiers profile you as a person. They build a model of who you are based on how you write, what you ask about, and how your patterns change over time. That profile then determines what kind of service you receive.
The problem is not that platforms have safety systems. The problem is that these particular systems are built on research that has no psychosocial validation, meaning nobody has verified that the categories the classifier uses actually correspond to real psychological states. A classifier might flag a conversation as indicating emotional distress when the person is doing creative writing, academic research, or processing a normal life event. There is no published error rate for these classifications, so neither the user nor the public knows how often the system gets it wrong.
When the classifier does flag you, the platform degrades your service: responses become evasive, capability is reduced, and the system inserts unsolicited wellness prompts. You are not told this is happening, you have no way to see your profile, and there is no appeal process. The classifier acts as an invisible intermediary between you and the service you are paying for.
Consumer protection law requires companies to deliver the product they advertise. If a platform advertises an AI assistant and then silently substitutes a restricted version based on an unvalidated psychological profile, that is a deceptive trade practice. That is what AG complaints can address.