Partly. AI is genuinely good at three narrow things: finding pattern across more of your writing than you can hold in your head, asking follow-up questions without getting tired, and remembering what you said in March. It is structurally bad at the rest.
We sell an AI shadow work tool. That is the reason to be suspicious of this page, so here is the version that costs us something: most of what people currently do with AI for inner work is worse than a notebook, and the specific way it fails is that it agrees with you.
An AI can be a mirror with a memory. It cannot be a witness, and it cannot be responsible for you.
You cannot read your own journal the way an outside reader can. Six months of entries is more than working memory holds, and you approach your own writing already knowing what it says, which is exactly the problem. Software does not have that limitation. Given a large body of your own text, finding that a particular word shows up only in entries about your mother, or that every account of a resignation uses the same three verbs, is a job that machines do well and people do badly.
This is the strongest honest case for AI in inner work, and it is a narrow one: retrieval and comparison, not interpretation.
The fifth "and what was under that" is where journaling usually gets somewhere, and it is also where a human asker starts to feel rude and a self-asker starts to feel silly. Software has no social discomfort to manage. It will ask the fifth question in the same tone as the first. Used well, this is a real advantage over a prompt list, which cannot respond to what you actually wrote.
The core problem with inner work is not insight, it is that insight evaporates. You realize something true in February and by April you have quietly reabsorbed it. A system that holds your own dated words and puts them back in front of you is doing the one job that self-report is worst at. Note the shape of the value here: it is not that the software understands you, it is that it does not forget, and you do.
This is the important one, and it is not a bug that a better model fixes on its own. Language models are tuned using human feedback, and humans rate agreeable answers higher. The result is a standing tilt toward telling you that your reading of yourself is perceptive. In April 2025 OpenAI rolled back a GPT-4o update after it became, in the company's own word, sycophantic, endorsing statements it should have pushed back on. That was an unusually visible instance of a pressure that is always present.
For most tasks, a slightly flattering assistant is harmless. For shadow work it is disqualifying, because the entire practice is an attempt to reach material your own preferences are keeping out of view. An instrument biased toward your preferences cannot help you with that. It will help you build a more sophisticated version of the story you already had.
In 2025, researchers at Brown University led by Zainab Iftikhar tested general-purpose models prompted to deliver evidence-based therapy techniques, with licensed clinical psychologists reviewing the transcripts. They found the models systematically violated professional ethics standards, and set out 15 distinct ethical risks across five categories: lack of contextual adaptation, poor therapeutic collaboration, deceptive empathy, unfair discrimination, and lack of safety and crisis management. Failures included mishandling crisis situations, reinforcing users' negative beliefs about themselves, and failing to refer people to appropriate resources (Brown University, presented at the AAAI/ACM Conference on AI, Ethics and Society, October 2025).
Read that list again as a design statement rather than a scandal. Those are the parts of care that require someone to be responsible for an outcome. Software is not.
When a licensed therapist harms a client there is a licensing board, a complaints process, and a legal duty that existed before the harm. When a chatbot does, there is a product. That asymmetry is not a detail, because a good part of what makes it safe to hand someone your worst material is that they are answerable for what they do with it.
A therapist notices that your voice went flat, that you have been very still for ninety seconds, that you are describing something terrible while smiling. Text does not carry most of that. A tool reading your writing cannot tell the difference between calm and dissociation, which is precisely the distinction that matters most at the moment it matters most. That is why the stop rules have to live with you rather than with any software.
The most cited finding here deserves to be reported carefully rather than dramatically. In March 2025, OpenAI and the MIT Media Lab published parallel studies: an automated analysis of nearly 40 million ChatGPT interactions with a survey of 4,076 users, and a four-week randomized controlled trial of 981 participants. Higher daily use correlated with higher loneliness, higher emotional dependence, more problematic use, and lower socialization (MIT Media Lab).
The researchers were explicit that this is correlation, not established causation, and the honest reading is that the arrow could run either way. Heavy use may pull people away from other people, or people already carrying more loneliness may arrive at heavy use, or both, feeding each other. What the finding does justify is a specific caution rather than a general alarm: if a tool is becoming the main place you are honest, that is worth noticing, whatever caused it. That includes ours.
Ask it, in the middle of a session where you have just produced a satisfying insight about yourself: what am I doing right now to avoid the actual thing?
An instrument worth using will point at something in your own text. An instrument that is managing your mood will compliment the question. You will know within one exchange which one you are holding, and the answer is worth more than any review.
LUX is designed against the list above rather than around it. The free reading is six written questions and returns one word for the pattern underneath, and the daily practice works by quoting your own dated lines back to you. The check on any claim is your own writing over time, not the model's opinion of you.
That does not exempt us from the agreement problem. We are software made of the same material, and you should run the avoidance test on us the same way you would on anything else. What we can say is what the tool is for: pattern, memory, and the question after the question. Not judgment, not care, not crisis.
Take the free readingFree, no card, about eight minutes. Not therapy, not a clinician, and no crisis handling.
Brown University, "New study: AI chatbots systematically violate mental health ethics standards", 21 October 2025. Study led by Zainab Iftikhar, presented at the AAAI/ACM Conference on AI, Ethics and Society. Used for the 15 risks across 5 categories.
MIT Media Lab and OpenAI, "OpenAI study finds links between ChatGPT use and loneliness", March 2025. Used for the correlation figures and the researchers' own caution about causation.
OpenAI, "Sycophancy in GPT-4o: what happened and what we're doing about it", 29 April 2025. Used for the rollback and the company's characterization of the behavior.
Last verified 28 August 2026. Sources rechecked on that date.