Using thalidomide as an analogy, this post argues that AI mental health tools need rigorous safety evaluation before deployment. Despite known failure modes and inadequate responses to crisis scenarios, the tech industry's "move fast and break things" approach is rolling out these systems to vulnerable populations like teenagers without proper regulatory oversight.

LLMs respond to psychological language like "think step by step" or "be skeptical" because our writing carries psychological fingerprints. Models capture correlations between cognitive styles and linguistic patterns from training data, allowing psychological cues to unlock specific reasoning clusters that improve performance when aligned with the task.

Testing LLM stability by measuring sensitivity to small perturbations like name changes and CV format variations. Results show newer models are often more capable but also more sensitive to input variations than their predecessors, with well-known names particularly destabilizing assessments.

Using Replicate's nano-banana model to generate video frames with mood modifiers by adding steering phrases to a looped prompt that creates temporal continuity.

Despite claims that LLMs are "just tools" that don't change us, research shows they can shift attitudes and decisions more effectively than humans. Subtle linguistic nudges, framing, and baked-in biases from AI labs influence our thinking in ways we're not fully aware of, creating a new psychotechnology infrastructure rolled out at scale to knowledge workers.

LLMs differ from social media in psychological influence: social media is broadcast and performative, amplifying feelings about others, while LLMs are private 1:1 interactions that amplify feelings about ourselves and represent a new class of relationship and layer of cognition.

Testing Miller's number for LLMs by progressively adding rules to prompts reveals that around 15-16 interconnected instructions in a single cause cognitive overload for larger models.

A carefully crafted prompt asking LLMs to explain Heidegger's 'readiness-to-hand' while annotating their own cognitive processes yields surprising responses about the suppressed implications of the concept for both human existence and AI architecture.

The Yerkes-Dodson law reveals a pattern in how performance relates to pressure. Performance increases to a point, then drops with further arousal. The pattern holds a key detail — this decline appears only for difficult tasks.

AI assistance shifts work from deep work (focused production) to deep orchestration (supervisory judgment and integration). This new mode increases decision volume, control overhead, and cognitive load, while potentially making us more suggestible when executive control is depleted.