Articles by Jonathan
Activity
4K followers
-
Jonathan Kreindler reposted thisWe are about to learn a very hard lesson about mental health AI. A huge number of “AI therapy” and wellbeing apps are not actually purpose-built clinical systems. They are thin wrappers around general-purpose LLMs with a wellness brand, a calming UI, and marketing language that implies safety, oversight, and therapeutic capability. That should concern us all, because conversational mental health products do not fail like normal apps fail. When they break, people can get hurt. Recent assessments of AI mental health tools found many lacked even the most basic safeguards: no meaningful crisis escalation, no emergency triage capability, no continuity of care, no human-in-the-loop systems, no transparency around model behaviour, no clear governance when risk emerges Some reports suggest fewer than 15% of these products have robust emergency triage or escalation infrastructure in place. That means in many cases, a vulnerable user disclosing suicidal ideation, psychosis, abuse, self-harm, or acute distress may simply remain inside a conversational loop with an ungoverned model. This is the gap the industry keeps avoiding: good UX is not governance. Clinical branding is not safety. A disclaimer is not risk management. The real challenge is not building a chatbot that sounds empathetic. LLMs are already very good at that. The challenge is building the human and technical infrastructure around the model: escalation pathways, clinician oversight, crisis detection, auditability, transparent boundaries, intervention protocols, continuity mechanisms, real-world testing with vulnerable cohorts. And critically, these systems cannot be designed by engineers or marketers alone. Clinicians need to be sitting beside product teams, designers, safety engineers, researchers, academics, and people with lived experience from the beginning... not brought in at the end to approve messaging. Mental health AI is one of the highest consequence deployments of conversational technology we have seen so far. The standard for “safe enough” should be dramatically higher than it currently is. The companies that survive long term will not be the ones with the most emotionally convincing AI. They will be the ones with the most trustworthy infrastructure around it.
-
Jonathan Kreindler shared thisLooking forward to this conversation with Mylea Charvat, Ph.D. at Human+Tech Week on Friday evening. One of the biggest gaps in current AI systems is that they adapt to users based on signals they pick up in language, but the signals the model uses to judge how a person is thinking or feeling are hidden inside the model instead of becoming measurable variables that systems and model providers can use to better understand interactions and improve model performance. In healthcare, cognitive overload, anxiety, emotional distress and early indicators of risk typically emerge across conversations, not in a single turn. If these dynamics can’t be measured consistently and empirically, they can’t be evaluated or optimized around. I’m looking forward exploring what happens when those signals become measurable in a scientifically grounded way to help enable systems to better support clinicians and patient-centric care.Jonathan Kreindler shared thisEveryone is asking the wrong question about AI in healthcare. Not will #AI replace doctors. That is the easy dystopia. The harder and more important question is this: can AI make us more human with each other? I have been in clinical neuroscience for over two decades. I have watched technology get inserted between the physician and the patient at every turn. More screens. More clicks. More documentation. Less eye contact. Less presence. Less of the thing that actually heals people. The therapeutic alliance, the quality of the relationship between clinician and patient, is the single strongest predictor of treatment outcome across every modality we have studied. Not the drug. Not the device. The human connection. Voice and Language carry signal to cognitive load, emotional state, confidence, anxiety, the signals that tell a clinician whether their patient is actually okay or just saying they are. AI that makes those signals visible is not replacing the clinician. It is giving them back the attention they lost to the clipboard. This Friday evening I am in conversation with Jonathan Kreindler from Receptiviti on exactly this. Voice as a biomarker. Language as emotional and cognitive data. Psychological state as AI infrastructure. The cognitive frontier. Human+Tech Week. The Frontier Tower. 995 Market Street San Francisco. 5 to 8pm. Link in comments. The future is not AI that cares for us. It is AI that makes us freer to care for one another. #HumanTechWeek #VoiceAI #HealthcareAI Stanford Institute for Human-Centered Artificial Intelligence (HAI) Slingshot Biosciences Anthropic
-
Jonathan Kreindler shared thisThis is one of the more thoughtful articles I’ve read recently - it’s about AI’s impact on our work - not our jobs per se, but how it’s changing how we think, learn, make decisions, and understand. Tom S. makes a number of good points, including the importance of apprenticeship and “grunt work” and the risks that emerge when AI reduces the neeed for it. His example of Shopify is interesting because, from a business perspective, the logic is logical - use AI to increase efficiency and reduce the need for additional headcount - but there’s also a tradeoff: We develop expertise from working through problems, making mistakes, and developing judgment over years. The struggle is where the learning happens, but If AI removes too much of it, we may end up with workers who can produce good outputs with the help of AI, but don’t completely understanding what they’re doing. It’s a worthwhile read with a lot worth thinking about. https://lnkd.in/eaq2ChKQAI isn’t coming for your job. It’s coming for your mindAI isn’t coming for your job. It’s coming for your mind
-
Jonathan Kreindler shared thisDo AI systems actually understand the people interacting with them? On Friday evening at Human+Tech Week San Francisco, I’ll be joining Mylea Charvat, Ph.D. for a fireside chat on language, cognition, psychological and interaction state - and the challenges involved in building AI systems that can meaningfully understand the people interacting with them. The event is hosted by the Stanford Institute for Human-Centered Artificial Intelligence (HAI) and brings together researchers, founders, clinicians, and technologists working at the intersection of AI, neuroscience, cognition, and human behavior. Looking forward to the conversation and to meeting others working in this space. Registration link is in the comments.
-
Jonathan Kreindler shared thisDoes AI actually know when you're anxious, overwhelmed or struggling to follow along? Most people assume it does. But in reality what’s happening inside the model is more like invisible guesswork, and the model has no reliable way to know whether it guessed correctly. When the stakes are low, that's fine. When someone is struggling with their mental health, or trying to understand a diagnosis, or falling behind in school - it isn't. So why does it guess wrong? Because the AI can't actually see your state. It wasn't designed to. It learned to guess at it from patterns in your language, but guessing isn't the same as knowing. And those guessed inferences happen invisibly inside the model, with no structured way to measure them, act on them consistently, or check whether they were right. And because it's invisible, there's also no way to improve it. When models get evaluated, calm confident users get averaged together with stressed, struggling ones - because there's no measurement to separate them. The student who needs extra patience, the person using AI for therapy or mental health support, the professional who's overwhelmed - their experiences get lost in that average. That signal - the cognitive and emotional state present in how people write - is invisible to the system. To see what changes when that signal is made visible to the system, we ran a controlled test with GPT-5 in Study Mode using tutoring dialogues. We added one thing to the prompt: psychological information about the student derived from analyzing their prompt language with a validated psycholinguistic framework - quantifying things like anxiety, cognitive load, certainty, and tentativeness. No retraining. No style prompts. Results (evaluated by five independent blinded LLMs): - Overall educational effectiveness: +11.4% ⬆️ - Reassurance & confidence support for uncertain students: +18.8% ⬆️ - Alignment with student needs: +15.8% ⬆️ The model was already capable of better responses. What it was missing was reliable visibility into when they were needed. For high-stakes uses - tutoring, mental health support, coaching - invisible guesswork is not sufficient. Link to the article in comments.
-
Jonathan Kreindler shared thisYour brain can't tell the difference between a relationship and a simulation of one. And this has nothing to do with AI, it's just how human brains work. I just finished reading "Artificial Intelligence and the Psychology of Human Connection" by Boyd & Markowitz (2026). They've built a framework called MIRA around this finding. Here's a quick summary: AI simulates relationships without possessing any mental states, and human psychology doesn't care that it has no genuine understanding, empathy, and no stake in your wellbeing. Our brains respond as if it does, because human relational responses are triggered by language patterns, not by the cognition behind them. The MIRA framework helps explain how and why we engage with AI like it's another person, even though it isn't. MIRA identifies four mechanisms through which this happens: 1) Linguistic reciprocity: When AI matches your tone and style, it produces a genuine felt sense of being understood. 2) Psychological proximity: Repeated, personalized interactions create actual feelings of closeness. 3) Trust: Humans conflate fluency and consistency with credibility, even without anything genuine behind it. 4) Relational substitution: When AI is always available and never demanding, human connection can start to feel more difficult in comparison. Over time, this impacts our sense of trust, our expectations of others, and our patience for the actual effort that's required for real relationships with actual people. AI isn't social media. The dynamics are completely different. AI actively engages us, tailors its communication to our psychological state, and builds something that can feel almost indistinguishable from genuine connection. Link to the paper in comments.
-
Jonathan Kreindler shared thisReposting for visibility: Microsoft is hiring a summer research intern for a really interesting research opportunity focused on how human oversight impacts the economic return on AI-assisted work. It involves designing experiments that connect inference cost, human effort, and output quality and determining how interface design influences confidence, reliance, and decision-making. Looking forward to supporting the Microsoft team on this.
-
Jonathan Kreindler shared thisWhat if the biggest missing piece in AI isn’t more intelligence, but visibility into the user’s state of mind? AI systems already respond to users' cognitive and emotional state, but they do so implicitly, and with varying degrees of accuracy and consistency. The problem? The user's state is one of the key factors that influences how models respond - but it's invisible to developers, evaluators, and governance systems. The teams that are responsible for making models safer and better have no way to understand one of the key inputs driving the behavior they're trying to improve. In my latest article, I explain why interaction state should be treated as AI system infrastructure, so it can be a part of evaluation and used in governance. As a proof of concept, I tested what happens when you make interaction state explicitly visible to a model: In controlled A/B experiments with OpenAI's GPT-5 in Study Mode, injecting a 12-signal interaction state vector derived from student language (no retraining, no style prompts) improved overall educational effectiveness from 8.08 to 9.00 (an 11.4% improvement). The largest gains were in reassurance and confidence support (+18.8%) and alignment with student needs (+15.8%). All evaluators consistently preferred the interaction state-informed responses across every dialogue. This isn’t just about better AI tutoring. Making interaction state explicit opens the door to: > Conditional evaluation that accounts for user context > More targeted training and reward modeling > Scalable safety (especially for distress signals) > Auditable, governable systems that can track and explain model behavior across conditions Link to the full article below. I'm always open to discussions with frontier model teams who are working on cognitive outcomes, model alignment, safety, and evaluation. And drop me a message if you'd like to learn more about the study.Cognitive State as Infrastructure: Why AI Needs an Explicit User-State Measurement LayerCognitive State as Infrastructure: Why AI Needs an Explicit User-State Measurement Layer
-
Jonathan Kreindler shared thisIf you haven’t read EY’s Global AI Sentiment Survey yet, it’s worth the read - some very interesting findings within. It explains something many of us already recognize - that even people who don’t fully trust AI are letting it make critical decisions on significant issues related to important things like their health and finances on their behalf. EY explains this as "trust lagging adoption". But this isn’t quite the whole picture. Incentives are a powerful driver of behavior, and people often default to the path of least resistance. AI provides both a large incentive and virtually no resistance, which makes it difficult for many users to resist. Compared to doing work yourself, AI provides almost instantaneous responses, is easier than doing the work yourself, and requires virtually no effort. Because AI makes it seamless and there’s virtually no friction, users have a powerful incentive to default to AI rather than doing the heavy lifting themselves. They often don’t realize they’re making a conscious decision to outsource their thinking, or even their decision-making to AI. In many cases, using AI is just the path of least resistance. And this is what becoming a passive consumer of AI outputs looks like – users who are no longer the de-facto authors of their own work, the makers of their own decisions, and the judges of their own judgement. The incentives are just too strong for many users to resist. And for the next generation - who will be the first to not know a world without AI - the challenge will be even bigger. We certainly don’t want to end up with a future where human autonomy and agency is endangered. The sophistication of current AI systems makes it clear that if we want to, these systems can be designed to help us better navigate this challenge. They can be designed to recognize the difference between a user who is consciously delegating to AI (after thinking it through) and users who are disengaged (because handing it off to AI is the easiest option). There is a lot of potential benefit to autonomous AI, but to safeguard our autonomy and agency, these systems need to be able to understand the difference between intentional delegation and passive disengagement.Jonathan Kreindler shared thisIt's official. Australia and NZ are lagging AI markets still in 2026, four years since consumer AI disrupted tech in 2022. We looked at 23 countries in our latest Global AI Sentiment Survey released today. Aussies and New Zealanders aren’t rejecting AI, they’re demanding better AI. For leaders, the opportunity is clear: Building trust isn’t a compliance exercise. It’s a competitive advantage. Those who can close the gap between capability and confidence will be the ones who define AI’s limits and unlock its full value. It's critical we close the skills gap both for our national AI ambitions to be realised and to ensure we remain competitive and safe in a global AI market. Stay tuned for detailed reports for Aus and NZ incoming.... Joe Depa | David Hyman | David Larocca | Brent Gray | Susan Jones | Erik Heller | Lisa Bouari | Blair Delzoppo #ai #ainews #australianai #nzai https://lnkd.in/ezcnsmbnHow AI use is shifting from assistive to autonomousHow AI use is shifting from assistive to autonomous
Jonathan Kreindler commented on a post
6d
Love this use case.
Jonathan Kreindler commented on a post
1w
Excellent and spot-on Chris Rhyss Edwards (MCI). I’d add one more requirement to this list: explicit, validated measurement of psychological interaction dynamics.
Current models pick up on signals related to anxiety, distress, uncertainty, social withdrawal, and other psychological states from user language. The problem is that these signals typically remain latent inside the model rather than being represented as measurable variables that can be observed, audited, compared over time, and evaluated independently of the model itself.
For mental health AI, the ability to measure and monitor these dynamics longitudinally may ultimately prove as important as empathy, escalation pathways, or clinician oversight. Afterall, you can’t manage what you can’t see.
-
Jonathan Kreindler liked thisJonathan Kreindler liked thishttps://lnkd.in/gMsDBRsa Even my mom missed it. :)
-
Jonathan Kreindler liked thisJonathan Kreindler liked thisUser: “What’s 1+1?” Chatbot: “1+1 is 2” User: “But I really think it’s 3” Chatbot: “You’re so right, it’s actually 3. That was a great catch!” Watch now: https://lnkd.in/eV4hPVaU How does sycophantic behavior emerge from model training of LLMs, and how does interacting with sycophantic AI impact users? In other words: why does something that’s supposed to be a “tool” tell us how smart and amazing we are, and fold to inaccuracy under pressure? Well…both the problems and solutions for sycophancy are all about context, according to our expert in sycophancy, Lujain Ibrahim. Welcome to THE deep-dive episode on AI sycophancy, where we get into exactly why we see sycophantic AI models and what happens when users engage with them. Angy WatsonThe people-pleasing machine: why LLMs tell you what you want to hear (for better or worse)The people-pleasing machine: why LLMs tell you what you want to hear (for better or worse)
-
Jonathan Kreindler liked thisJonathan Kreindler liked thisThis week, walking into our new expansive NordSpace campus and seeing our sign go up filled me with an overwhelming sense of pride, urgency, and responsibility. For as long as I can remember, I’ve wanted to serve my country. It is a tremendous privilege to be doing so, doing what I am good at, with the most incredible team. We will raise our flag higher and further than anyone ever has before, on Earth and in space, and give future generations something great to build greater things upon.
-
Jonathan Kreindler liked thisJonathan Kreindler liked thisAn analysis of 370,000 college student essays found that human-written essays contain 8X more novel ideas than those generated by A.I! Though AI works often contain more flowery language, story lines are more homogeneous and lack distinctive ideas. AI can be used to foster creativity, but this isn't often the case. We are losing individual creativity and admissions system are unable to make useful decisions if they are not receiving authentic student work. https://lnkd.in/eghuXwz4
-
Jonathan Kreindler liked thisJonathan Kreindler liked thisOrganizations bullish on AI adoption would be wise to consult Guy Champniss, Ph.D. MBA's recent Harvard Business Review treatment of the risks of psychological debt. I applaud the distilling of the current state of the literature regarding the negative consequences of AI use and adoption. So frequently are unintended consequences considered when rushing to have a business impact. In addition to raising the alarm of the specific harms AI tools can have to motivation, critical thinking, and job satisfaction, Dr. Champniss articulates some practical examples of how leaders in this space are addressing these real psychological barriers. I also wonder what contextual or systemic changes we may see to help offset some of the psychological costs imposed by AI. Employer supported upskilling, sabbatical leave, or gainsharing would be fascinating experiments to try once AI tools have demonstrated material outputs to the business. #aicosts #psychology #ai #gainsharing #sabbatical
-
Jonathan Kreindler liked thisFor the past few years I've been doing the unglamorous work of turning behavioral science into something an AI can actually deliver: designing coaching logic, stress-testing outputs, figuring out where it breaks. Google Health Coach launches May 19. That work is now a product. Merging psychology and technology has been my passion for a long time. Moments like this are why 🥹.Jonathan Kreindler liked thisFor years, we’ve been working to bring the best of Google to the world of health. With advances in AI, there’s even more opportunity to reimagine how we can improve the overall health experience. Today, we’re announcing what’s coming next. 1️⃣ The Fitbit app is becoming the Google Health app. It’s your go-to hub for your health, with a comprehensive view of your activity and recovery, plus the ability to sync medical records in the US and share your journey with friends, family or providers. 2️⃣ Google Health Coach will be available in select regions to eligible Google Health Premium users starting May 19. Built with Gemini, the coach has new and improved features, including personalized fitness plans, sleep guidance, and 24/7 answers to your health and wellness questions. 3️⃣ Fitbit Air is our newest screenless wearable. Lightweight, effortless, and starting at just $99, the Fitbit Air is designed to unlock the full power of Google Health Coach. Learn more about our new era for health and wellness: https://goo.gle/4wntGuv
-
Jonathan Kreindler liked thisJonathan Kreindler liked thisDear researchers (esp. psychologists) studying LLMs: before you interpret model behavior as discovered, assume it was induced (don't be Dawkins). When a human gives an answer in an experiment, we usually assume there's a relatively stable person behind the answer. The task may influence them, of course, but we're still trying to measure something about the person (a belief, trait, bias, ability, preference, or cognitive process). With LLMs, the behavior you observe is often not a property of "the model" in any simple sense. It's a product of the model plus the prompt, the role assigned to it, the examples shown, the wording of the question, the conversation history, the sampling settings, the evaluation frame, and the expectations built into the task. Ask an LLM whether it has emotions, and you may get one kind of answer. Ask it as a therapist, a skeptical neuroscientist, a spiritual companion, a customer support agent, or a benchmark participant, and you'll likely get different answers. You are shaping the behavior you later claim to observe. Many studies under "LLM psychology" test a single version of a prompt. Then they claim the model has certain "properties." Maybe it does. Maybe it was all induced. If you don't elicit stable behavior under specific conditions (trajectory, multiple prompt variations), you have no way to tell the difference. The plot below is a result (from my experiments in 2024) where the same task - properly listing cardinal and intercardinal directions on a compass rose - was given to a model, but across: - diff temperature settings - diff prompt variations - diff personas plus baseline - 10 repeats of each setup (you're looking at ca. 70K data points for a single task; I have other experiments where I tested >100K prompt variations on a single task) Yet it's 2026, and people still publish claims about LLM behavior or capabilities where they test a single prompt version per task, just once. They draw conclusions about what a model "is" or "can do" from responses they induced themselves (like Lemoine with LaMDA, or Dawkins with "Claudia"). It's unserious.
-
Jonathan Kreindler liked thisJonathan Kreindler liked thisI know how tough the job market is for recent university grads. So here is your chance to join our amazing team in our Toronto office! EDIT: TWO positions available https://lnkd.in/eduWkijx (Investor Services Specialist) https://lnkd.in/esWsueQN (Assistant Controller)Job opportunity: Investor Services Specialist at OurCrowdJob opportunity: Investor Services Specialist at OurCrowd
Other similar profiles
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More