English

Not everything that looks new is truly new. Generative AI can be a real productivity boost in research, useful for outlining, rewriting, translating, and even brainstorming when you’re stuck. But the same fluency that makes AI outputs feel “publishable” can also hide a quieter risk: AI-generated work can unintentionally recycle existing scholarship, not always through copied sentences, but through reused ideas, familiar argumentative scaffolds, replicated study logic, and look‑alike novelty. This is the reality many researchers are now facing: the integrity problem is often not blatant plagiarism, but hidden reuse, where the writing is original in wording while the intellectual contribution is thin, derivative, or too close to what already exists. When that happens, “novelty” becomes a surface effect, and research integrity becomes less about catching copied text and more about protecting genuine originality, traceable reasoning, and honest contribution. A strong example comes from Gupta and Pruthi’s ACL 2025 study, All That Glitters is Not Novel: Plagiarism in AI Generated Research (an Outstanding Paper award at ACL 2025). Their focus isn’t simple copy-paste plagiarism, but a more subtle kind: research-style text that rephrases or recombines prior work in ways that can look “new” during a fast review process. Instead of asking reviewers to judge novelty in the usual way, the authors designed an expert-led setting where participants were explicitly tasked to look for plagiarism sources. They had experts evaluate 50 LLM-generated research documents (including documents from “The AI Scientist” and other public proposals, plus newly generated ones). This matters because typical evaluations often assume good faith and don’t incentivize active source-hunting. They also used a clear rubric: the top scores corresponded to cases where there’s essentially a one-to-one mapping between the generated methodology and earlier work, or where substantial parts are borrowed from a small set of prior papers without credit. In other words, it’s not about identical sentences, it’s about the intellectual skeleton of the method and contribution being too close to something that already exists. The headline result is hard to ignore: experts flagged 24% of the 50 documents as plagiarism (scores 4–5) after verification steps that included contacting original authors; if you also count cases where verification wasn’t possible (e.g., authors unreachable), the rate rises to 36%. That gap is important, because it shows how “confirmed” cases may still be an undercount when real-world verification is slow or impossible. This is exactly why AI-era plagiarism can feel different: the risk often sits at the idea level, problem framing, method pipeline, and contribution claims, rather than in identical phrasing. If a proposal is written confidently, packaged with clean sections, and sprinkled with plausible citations, it can pass a quick surface check even when the underlying concept is not truly original. The study also highlights a second problem: automation doesn’t save us (yet). The authors report that several automated approaches, including embedding-based search and a commercial plagiarism service, were inadequate for detecting plagiarism in these LLM-generated proposals. That’s consistent with a broader reality: “semantic borrowing” is much harder to catch than overlapping strings of text. For peer review, this creates a nasty workload tradeoff. If AI increases the volume of polished submissions while also increasing the probability of hidden borrowing, reviewers must spend more time doing detective work, searching literature, mapping methods, and checking whether “novel contributions” are just renamed versions of known ideas. That pressure doesn’t just slow review; it can also push reviewers toward shallow heuristics, which makes the system even easier to game. For writers who use AI ethically, the safest mindset is: AI can help you express your ideas, but it should not be the source of your contribution. Keep a “provenance trail”: what you read, what you copied into notes (with quotes), and what you personally decided. If AI suggests a method or framing, treat it like an untrusted hint, then verify by searching for prior work and adding explicit citations where your idea connects to existing literature. For universities, journals, and conferences, the response shouldn’t be panic, it should be process upgrades. Require transparent disclosure of AI use, strengthen novelty checks (especially at the idea/method level), and give reviewers tools/time to do targeted source-searching when something feels “too clean.” Most importantly, reward careful citation and honest positioning (“this is an extension of X”) rather than over-marketing novelty, because in the AI era, exaggerated novelty is becoming easier to manufacture than real research progress. Finally, this is where human reviewers and human judgment remain essential, and where responsibility can’t be outsourced. Because idea-level reuse is subtle, it often takes domain expertise to notice when a “new” pipeline is really a renamed or lightly rearranged version of established work. In other words, integrity in the AI era depends less on automated flags and more on careful reading, source-checking, and accountable editorial processes. The same caution applies in clinical contexts: many AI tools on the market are still not reliable enough to be treated as clinical-grade systems, and they can produce confident but wrong, biased, or unsafe outputs. In therapy settings especially, we should treat AI as supportive workflow software, not an authority, keeping clinicians firmly responsible for interpretation, decisions, and client safety.

In a busy speech and language therapy week, “AI” doesn’t arrive as a philosophical debate. It arrives as a friction point: Can we trust automated transcription for a multilingual child? Should we trial a screening app to reduce waiting lists? Can a generative tool draft home-program instructions without drifting beyond evidence? ESLA’s position paper, Shaping the Future of Speech and Language Therapy with Artificial Intelligence (published March 6, 2026), lands right in this reality. It acknowledges that AI is rapidly reshaping healthcare, and it correctly frames SLT as a profession built on interaction, context, and judgement. The problem is that when clinicians turn to a position paper, many of us aren’t only looking for a vision, we’re looking for direction. To ESLA’s credit, the paper avoids the two common extremes: adopting tools because they’re new, or rejecting them because they feel risky. It states a sensible professional stance: AI should support rather than replace therapists, and implementation must respect human rights, dignity, diversity, and inclusion. But here’s the tension: those principles are true, and still too general to steer day-to-day decisions. In practice, “support not replace” needs to translate into operational boundaries: Which tasks are acceptable to automate? Which tasks must remain clinician-led? What level of human oversight is “enough” in real services under pressure? Without that, the document risks becoming a shared set of values that everyone agrees with, while actual practice drifts in whichever direction vendors, budgets, or time constraints push it. ESLA outlines opportunities (screening, assessment support, intervention planning, outcome monitoring, personalised care, hybrid delivery, access). Again, plausible, and in many areas, exciting. Yet the paper could have pushed further into the uncomfortable practicalities: the reality that “personalisation” can become “fit to the majority,” especially for multilingual children, minoritised dialects, AAC users, complex disability profiles, and culturally specific communication norms. If ESLA’s goal is equity, we need stronger statements on minimum evidence expectations, not just aspirations. What counts as sufficient validation for multilingual speech recognition? What sample sizes? What types of accents/dialects? What error rates are acceptable and acceptable for whom? Otherwise, “equity” stays a hope rather than a measurable requirement. Where the paper feels strongest is its ethical backbone, privacy, consent, data ownership, bias, unequal access, and over-reliance. But ethically strong does not automatically mean clinically usable. This is where many clinicians will say: It’s nice, but… c’est flou encore. We wanted more “what to do / what not to do.” We wanted a toolkit. We wanted ESLA to name the concrete risks that show up in real workflows: staff uploading voice samples into unknown cloud services; student clinicians pasting identifiable details into chatbots; managers adopting “AI screening” because it’s cheaper, then letting it quietly become gatekeeping. A position paper can’t solve everything, but it can draw bright lines and this one often speaks in principles where the field is asking for guardrails. What would make this paper more actionable is a clinician-facing set of guidelines that reads like a decision support tool, not a manifesto. For example: a “red / amber / green” use-case list (e.g., green = admin summarisation with de-identified data; amber = draft therapy materials with clinician review; red = automated diagnosis/eligibility decisions). A privacy verification checklist (Where is data stored? Is it used to train models? Can we opt out? Who is the data processor/controller? Is encryption stated? What is retention/deletion policy? Can we get a Data Processing Agreement? Is there audit logging?). A minimum evaluation standard (evidence thresholds, bias testing expectations, multilingual performance reporting, disclosure requirements). And even a “vendor questions” one-pager managers can use before procurement. Without these, professional leadership remains a principle, but clinicians and services are left to invent policy locally, inconsistently, and often under time pressure. ESLA’s conclusion calls for deliberate, collaborative, ethical adoption so AI strengthens participation, dignity, and inclusion. That’s a good direction. But the next step should be explicit: an implementation appendix that turns values into practice. Because the profession doesn’t just need permission to engage with AI, we need structure to do it safely. Until then, ESLA’s paper functions as a strong ethical positioning statement, but not yet as the practical compass many of us hoped for. In other words, the paper is important, timely, and well-intentioned, but it still needs more actionable guidance for real-world practice.

In research and in therapy, the hardest part of the week is often not the “clinical moment” or the “research idea” itself. It’s the accumulation of in-between work: turning rough notes into coherent formulations, translating technical concepts into language a client can live with, and documenting care in a way that is both faithful and legible. Against that reality, the most useful way to approach ChatGPT 5.3 and 5.4 is not as a question of novelty, but as a question of workflow. What can these tools responsibly reduce, and what new risks do they introduce? Before you compare speed or “intelligence,” ask the compliance question readers actually need answered: is this tool HIPAA‑approved for PHI, or not? HIPAA doesn’t officially certify software as “approved”; in practice, the baseline is whether you can get a Business Associate Agreement (BAA) and enforce the right retention/access controls. OpenAI states that using its API with PHI requires a BAA, and that a BAA for ChatGPT is currently available only to sales‑managed ChatGPT Enterprise or Edu customers, not to ChatGPT Business. So if you can’t contractually and operationally defend how patient data is handled, you shouldn’t treat the system as an extension of your clinical workspace, whether you’re in the U.S. under HIPAA or elsewhere under professional ethics and local privacy expectations. Once guardrails are clear, the 5.3/5.4 split becomes genuinely useful. ChatGPT 5.3 (“Instant”) reads like a model tuned for everyday flow: fast, smooth, and often excellent at turning messy ideas into clean language. In a clinical workflow, that matters because it targets the small tasks that become overwhelming at scale, drafting psychoeducation, rewriting intake instructions, simplifying coping-skill explanations, or making client-facing language less technical. It can also quickly produce several variants so you can choose the tone that fits a client’s age, culture, and readiness. The clinical value is real, but it’s also deceptively easy to misuse. When a model produces confident prose quickly, it can blur the boundary between translation (changing style while preserving meaning) and invention (quietly filling gaps with plausible-sounding content). That boundary matters because clinical documents are not just “writing”, they’re records that shape continuity of care, reimbursement, supervision, and how clients are understood by other systems. If you use 5.3 for notes or summaries, the safest stance is: it drafts language; you own the facts. ChatGPT 5.4 (“Thinking”) tends to be most helpful when the task has internal structure, where you need the tool to keep track of dependencies across steps rather than just rewrite a paragraph. In research, that might mean tightening a methods section without changing meaning, drafting a transparent analysis-plan template, or turning scattered meeting notes into a protocol outline you can critique. In clinical contexts, it can help you generate alternative hypotheses, map maintaining factors, or create decision-tree prompts for your own reflection. The benefit isn’t “it knows therapy,” but that it can help you organize complexity when you’re fatigued. That same coherence, however, introduces a specific kind of risk: a “complete-sounding” narrative that feels psychologically compelling even when it’s not accountable to the actual relationship, the client’s culture, or your observed data. 5.4 can produce formulations that read impressively integrated, yet still be subtly wrong, overly confident, or prematurely interpretive. In therapy, that can show up as false precision: labels that land too early, causal stories that are too tidy, or summaries that inadvertently overwrite the client’s own meanings. The more elegant the output, the more vigilant the clinician needs to be. Pricing matters because it quietly determines who uses what, how often, and under which safeguards, and that, in turn, shapes clinical risk. OpenAI’s consumer tiers are commonly framed as Go (~ $ 8/month in the U.S.), Plus ( $ 20/month), and Pro ( $ 200/month), while ChatGPT Business is listed at ** $ 25 per user/month billed annually** (and the pricing page advertises “Try for free”). Prices can also be region-dependent, so what you see in Lebanon may differ from U.S. stickers. The honest clinical question, then, is not “Which is cheapest?” but: Which plan aligns with our confidentiality requirements and governance? Lower-cost personal tiers can nudge clinicians toward “quick personal account” use (often with weaker admin controls), while Business/Enterprise tiers are designed for organizational controls, yet they still come with practical constraints (e.g., “unlimited” is typically subject to abuse guardrails, and some tiers have product/feature caveats; Go may include ads). A realistic way to use these tools well is to divide tasks by risk level. Low-stakes, high-value uses include drafting client handouts (without identifiers), rewriting psychoeducation, generating multiple wording options for a message you will review, or creating session structure templates. Higher-risk areas, diagnosis, risk assessment, duty-to-warn decisions, and definitive case formulation, should remain human-led, with the tool used only to support clarity of communication, not clinical judgment. The goal is not to “ban” the tool; it’s to keep it in a role that improves quality without quietly shifting responsibility. If there’s one operational lesson, it’s that these tools are best treated like exceptionally fluent interns: fast, helpful, and sometimes brilliant at presentation, but not accountable for truth. They can reduce friction and widen your drafting bandwidth, especially during heavy weeks, but only if you design a workflow where verification is normal rather than optional. In other words, the right question is not “Can it write this?” but “Can I audit what it wrote, and can I defend it ethically and clinically?” Ethically, the most important moment is often when the tool feels easiest to use. Transparency is not only a client-facing issue (“AI helped draft this handout”) but a professional integrity issue: what was delegated, what was verified, and what data entered the system. Responsibility means refusing to outsource clinical judgment, risk assessment, diagnosis, or case formulation, to a tool that cannot carry a duty of care. Data integrity, in turn, means keeping documentation anchored to observed clinical facts and the client’s words, and using governance-minded frameworks to decide what is appropriate to automate, what must remain human,

In our clinical and research work with adolescents, privacy is not a side issue, it is a developmental workspace where autonomy, identity, and boundaries are tested in real time. The American Psychological Association (APA) has emphasized that wanting privacy is normal and healthy in adolescence, especially when teens are exploring identity, relationships, and new emotions. When privacy is treated as inherently suspicious, we often intensify secrecy rather than strengthen judgment. At the same time, we are seeing a shift in where privacy is practiced. Many teens now process questions and feelings in conversations with AI chatbots. As Dr. Joshua Goodman notes, “Conversations with AI chatbots can feel like a safe space” for adolescents to discuss things they would not feel safe bringing to parents or others. In practice, that “safe space” feeling often comes from immediacy, lower embarrassment, and the absence of visible adult worry. The interface can feel calm, nonreactive, and available, especially for teens who expect criticism or conflict. From a therapist’s standpoint, this is precisely why we cannot treat chatbot use as a trivial trend or a purely parental concern. We have a professional responsibility to follow adolescents’ real help-seeking pathways, including the digital ones, and to make sure that what feels private is not mistaken for what is protected. If a teen is using a “kid chatbot” as their primary container for shame, identity questions, sexual concerns, trauma narratives, or self-harm thoughts, then chatbot use becomes clinically relevant behavior. It belongs in assessment, formulation, and ongoing treatment planning the same way we track sleep, peer relationships, substance use, and social media exposure. The central clinical concern is simple: perceived safety is not the same as real safety. Dr. Mary Alvord states it directly: “Perceived safety isn’t the same as real safety.” When a chatbot feels validating and discreet, teens may assume it is confidential in the way a therapeutic relationship is meant to be. But teens often use the word “private” to mean “no one will judge me,” “no one will get angry,” or “no one will make it awkward.” Data privacy asks different questions: who can access the content, whether it is stored, whether it is analyzed, and how it might be reused. This is where our clinical role is not to frighten teens, but to clarify reality. Psychoeducation about privacy is not a parent lecture; it is a therapeutic intervention. Just as we help adolescents build mental models of consent, emotional regulation, risk, and relationships, we can help them build a mental model of digital disclosure. A chatbot can be emotionally comfortable while still being structurally nonconfidential. An interface can feel intimate without offering the protections we associate with private human relationships. Importantly, the therapist’s task is not only to warn; it is to understand the function the chatbot serves. In sessions, we can ask: What platform do you use? When do you turn to it? What do you ask it for—comfort, advice, distraction, validation, “permission,” reassurance? How do you feel afterward—relieved, calmer, more distressed, more confused, more dependent on reassurance? These questions help us assess whether the chatbot is supporting coping or quietly strengthening avoidance, rumination, or isolation. They also help us identify a common pattern in adolescence: the teen is not necessarily choosing a chatbot over people; they may be choosing it over anticipated judgment. Because we hold a duty of care, we also have a responsibility to name the “privacy gap” explicitly: many adolescents do not know what happens to their data. They may assume that because the interaction is one-on-one and “sounds caring,” it is private in the same way that a clinician’s office is private. Our role is to bring the invisible layer into view, gently, concretely, and repeatedly, so that teens can make informed choices rather than emotionally-driven assumptions. This is especially important with younger users and “kid” versions of chatbots, where the design may inadvertently amplify the sense of safety while the teen’s understanding of data practices is still developing. One effective approach is to translate privacy into a skill-building exercise rather than a moral warning. In practice, we can guide adolescents to sort what they might share with a chatbot into clear categories: low-stakes content (study help, hobbies, neutral questions), personally identifying details (full name, school, address, phone number, account handles), and high-sensitivity disclosures (sexual content, self-harm thoughts, trauma narratives, family conflict, screenshots of private messages, and medical or mental health details). The goal is not to ban curiosity or punish disclosure. The goal is to build discrimination: which topics can tolerate a lower-protection environment, and which topics deserve a safer channel—trusted adults, clinicians, or crisis supports when needed. From there, we can teach a practical rule that adolescents can actually use: the more specific and identifiable the details are, the higher the potential cost if the content is stored, misunderstood, shared, or later accessed. We can also strengthen what developmental psychology already tells us adolescents are practicing: “future-self thinking.” We can ask: How would you feel if a parent, peer, school, or future partner saw this? Would you still want it written down somewhere? Could it be taken out of context? This anticipatory reflection is developmentally appropriate because foresight and risk appraisal are still consolidating in adolescence. Practiced regularly, it turns privacy from an abstract value into a usable decision-making habit. An ethical reflection is unavoidable because the issue is not only teen behavior, but also system design and adult responsibility. Teens and caregivers deserve transparent explanations of what chatbots can and cannot promise, especially when the interface feels intimate. As clinicians, we cannot outsource clinical judgment, or the protection of vulnerable disclosures, to systems without human accountability and professional ethics. Yet we also have to hold complexity: some teens have limited access to responsive adults, and chatbots may function as a temporary bridge. Our responsibility, then, is not to deliver blanket approval or bans. It is to stay close to adolescents’ real lives, to ask about chatbot use without shaming, to clarify the

Dans notre travail clinique et de recherche auprès des adolescents, la vie privée n’est pas une question annexe ; c’est un espace de développement où l’autonomie, l’identité et les limites sont testées en temps réel. L’American Psychological Association (APA) a souligné que le désir d’intimité est normal et sain à l’adolescence, en particulier lorsque les jeunes explorent leur identité, leurs relations et de nouvelles émotions. Lorsque la confidentialité est considérée comme intrinsèquement suspecte, nous avons souvent pour effet d’accroître le secret plutôt que de renforcer le jugement. Dans le même temps, nous observons un déplacement des lieux où s’exerce cette intimité. Beaucoup d’adolescents traitent désormais leurs questions et leurs émotions dans des conversations avec des chatbots d’IA. Comme le note le Dr Joshua Goodman, « Les conversations avec des chatbots d’IA peuvent donner aux adolescents l’impression d’un espace sûr » pour aborder des sujets qu’ils n’oseraient pas apporter à leurs parents ou à d’autres. En pratique, ce sentiment « d’espace sûr » vient souvent de l’immédiateté, d’une moindre gêne et de l’absence d’inquiétude visible chez l’adulte. L’interface peut paraître calme, non réactive et disponible, surtout pour des adolescents qui s’attendent à de la critique ou au conflit. Du point de vue du thérapeute, c’est précisément pourquoi nous ne pouvons pas traiter l’usage des chatbots comme une tendance anodine ou une préoccupation uniquement parentale. Nous avons la responsabilité professionnelle de suivre les véritables voies de recherche d’aide des adolescents, y compris numériques, et de nous assurer que ce qui « semble » privé n’est pas confondu avec ce qui est protégé. Si un adolescent utilise un « chatbot pour enfants » comme principal contenant pour la honte, les questions d’identité, les préoccupations sexuelles, les récits de traumatismes ou les pensées d’auto‑agression, alors l’usage du chatbot devient un comportement cliniquement pertinent. Il a sa place dans l’évaluation, la formulation et la planification thérapeutique continue, tout comme nous suivons le sommeil, les relations avec les pairs, l’usage de substances et l’exposition aux réseaux sociaux. La préoccupation clinique centrale est simple : la sécurité perçue n’est pas la sécurité réelle. La Dre Mary Alvord le dit clairement : « La sécurité perçue n’est pas la sécurité réelle. » Lorsqu’un chatbot paraît validant et discret, les adolescents peuvent supposer qu’il est confidentiel comme une relation thérapeutique est censée l’être. Or, les adolescents emploient souvent le mot « privé » pour signifier « personne ne me jugera », « personne ne se mettra en colère » ou « personne ne rendra la situation gênante ». La protection des données pose d’autres questions : qui peut accéder au contenu, s’il est stocké, s’il est analysé et comment il pourrait être réutilisé. C’est ici que notre rôle clinique n’est pas d’effrayer les adolescents, mais d’éclairer la réalité. La psychoéducation sur la confidentialité n’est pas une leçon parentale ; c’est une intervention thérapeutique. De la même manière que nous aidons les adolescents à construire des modèles mentaux du consentement, de la régulation émotionnelle, du risque et des relations, nous pouvons les aider à bâtir un modèle mental de la divulgation numérique. Un chatbot peut être confortable sur le plan émotionnel tout en restant non confidentiel sur le plan structurel. Une interface peut sembler intime sans offrir les protections que nous associons aux relations humaines privées. Il est important de rappeler que la tâche du thérapeute n’est pas seulement de mettre en garde ; il s’agit aussi de comprendre la fonction que sert le chatbot. En séance, nous pouvons demander : Quelle plateforme utilises‑tu ? À quels moments y recours‑tu ? Qu’en attends‑tu réconfort, conseils, distraction, validation, « permission », rassurance ? Comment te sens‑tu ensuite, soulagé, plus calme, plus en détresse, plus confus, plus dépendant de la rassurance ? Ces questions nous aident à évaluer si le chatbot soutient l’adaptation ou, insidieusement, renforce l’évitement, la rumination ou l’isolement. Elles nous aident aussi à repérer un schéma fréquent à l’adolescence : le jeune ne choisit pas nécessairement un chatbot plutôt que des personnes ; il le choisit plutôt pour éviter un jugement anticipé. Parce que nous avons une obligation de protection, nous avons aussi la responsabilité de nommer explicitement « l’écart de confidentialité » : nombre d’adolescents ignorent ce qu’il advient de leurs données. Ils peuvent supposer que, parce que l’interaction est en tête‑à‑tête et « donne une impression de bienveillance », elle est privée au même titre qu’un cabinet de clinicien. Notre rôle est de rendre cette dimension invisible plus concrète, avec douceur, concrètement et de façon répétée, afin que les adolescents puissent faire des choix éclairés plutôt que de s’en remettre à des suppositions dictées par l’émotion. Cela est particulièrement important chez les plus jeunes utilisateurs et dans les versions « enfant » des chatbots, où la conception peut involontairement amplifier le sentiment de sécurité alors même que la compréhension des pratiques de données est encore en développement. Une approche efficace consiste à traduire la confidentialité en un exercice de développement de compétences plutôt qu’en un avertissement moral. En pratique, nous pouvons guider les adolescents pour trier ce qu’ils pourraient partager avec un chatbot en catégories claires : contenus à faible enjeu (aide aux devoirs, loisirs, questions neutres), éléments personnellement identifiants (nom complet, établissement scolaire, adresse, numéro de téléphone, identifiants de comptes) et divulgations hautement sensibles (contenu sexuel, pensées d’auto‑agression, récits de traumatismes, conflits familiaux, captures d’écran de messages privés, ainsi que données médicales ou de santé mentale). L’objectif n’est pas d’interdire la curiosité ni de culpabiliser les jeunes lorsqu’ils partagent trop. L’objectif est de développer le discernement : quels sujets peuvent un environnement moins protecteur, et quels sujets méritent un canal plus sûr, adultes de confiance, cliniciens, ou dispositifs de crise si nécessaire. À partir de là, nous pouvons enseigner une règle pratique réellement utilisable par les adolescents : plus les détails sont précis et permettent d’identifier la personne, plus le risque potentiel est élevé si le contenu est stocké, mal compris, partagé ou consulté ultérieurement. Nous pouvons également renforcer ce que la psychologie du développement nous apprend déjà que les adolescents exercent :

Google’s late-February 2026 launch of Nano Banana 2, also described by Google and some coverage as Gemini 3.1 Flash Image, can look like a routine model update, but in clinics and research groups it behaves more like a workflow shift. It lands amid familiar production pressures: recruitment materials constrained by IRB timelines, conference figures that must stay legible at a distance, patient infographics that simplify without distorting, and teaching slides that communicate risk responsibly. Google’s core claim is that it pairs high image quality with Flash-level speed, now accessible via the Gemini app and Google Search experiences such as AI Mode and Lens, with explicit rollout messaging for MENA. When image generation becomes both fast and visually persuasive, the change is not only higher output, it is faster decision cycles. Google highlights features that map directly onto scientific and clinical communication: improved text rendering inside images, stronger subject consistency across elements, aspect-ratio control, and outputs up to 4K. In practice, that can mean fewer “manual design” iterations for medication titration visuals, consent-process diagrams, visual abstracts, and explanatory schematics, especially in teams without dedicated design support. One especially consequential claim is that Nano Banana 2 can draw on real-time information and images from web search to improve accuracy and support diagrams and infographics. The upside is immediate: clearer visuals produced quickly, potentially improving patient comprehension and reducing cognitive load for trainees. The methodological tension is equally real: if “grounding” relies on sources that are not clearly surfaced, archived, and citable, we gain speed while losing traceability. The risk is not only that an image could be wrong, but that we may not be able to reconstruct why it looks convincingly right. That tension sharpens with the “across Gemini, Search, and Ads” framing. The most clearly documented integrations at the moment are Gemini and Search. The Ads direction is plausible in context, Google Ads already supports generative image workflows with explicit guidance that advertisers must review AI-generated assets before publishing, and Google has previously described bringing Gemini models into Performance Max, but reports that Nano Banana 2 will directly power creative suggestions in Ads should be treated as a reported trajectory until Google’s Ads documentation names the model and scope explicitly. For health researchers and clinicians, this matters because the same leap that improves patient education materials can also accelerate persuasive health content optimized for clicks and conversion. As visual polish becomes cheaper, it becomes even less correlated with truth, shifting the burden of appraisal onto audiences who often have the least time and the most at stake. The most practical summary is simple: Nano Banana 2 compresses creation time, but it does not remove responsibility, it relocates it. Responsible teams will spend less time drafting and more time verifying: ensuring diagrams encode the correct causal claim, risk visuals match guideline thresholds, translations preserve meaning, and public-facing graphics do not outrun the evidence. The weak point is institutional: many organizations still lack lightweight governance for “AI-assisted visuals,” even when they have mature controls for medication orders, patient instructions, and research data outputs. Ethically, better tools raise, not lower, the bar on three duties. Transparency matters when provenance affects trust (patient materials, public education, high-impact research communication). Data integrity and auditability matter when web-grounded generation influences content and sources are not recoverable. Privacy discipline remains non-negotiable: identifiable patient details should not become prompt ingredients to produce “better” visuals. Watermarking and content credentials can help, but they are not substitutes for domain review when the content is medical.

Clinical and research work rarely fails because we cannot locate information; it fails because we cannot convert information into usable outputs quickly enough. The friction lives in the in‑between tasks: preparing a background brief for a case conference, turning scattered notes into a coherent protocol amendment, cleaning citations before a resubmission, or drafting a patient-facing handout that is both readable and accurate. What is newly compelling about the idea of an “AI computer” is that it aims at this connective tissue of work, not simply at conversation. Perplexity describes Computer as a unified, cloud-based system that can “research, design, code, deploy, and manage” projects end-to-end, breaking a goal into sub-tasks and routing them across specialized components. In the public description, it can orchestrate work across 19 models in parallel, “matching each task to the best model,” while also remembering prior context and connecting to external services. The ambition is not subtle: it is a bid to turn the browser into an operational workspace where the user specifies intent and the system performs the multi-step work. The attraction is easy to understand in real workflows. We already practice constrained delegation every day: we assign parts of a project to trainees, research assistants, administrators, or colleagues, and we integrate, verify, and sign off. An agentic system promises a similar pattern, but at a different speed and scale. If it performs reliably, it may create time for what remains stubbornly human: therapeutic presence, clinical judgment under uncertainty, nuanced supervision, and careful interpretation of evidence. That tension becomes sharper when orchestration is invisible. If Computer decomposes a task into steps, chooses tools, and merges results, then provenance matters: which sources were used, which model generated which claim, and what was the system’s uncertainty at each step? In research, these details determine whether a literature synthesis is reproducible; in clinical settings, they determine whether a handout, policy memo, or documentation aid stays within the boundaries of evidence-based practice. The more autonomous the workflow, the more we need systems that make their reasoning legible rather than merely impressive. A second tension is provenance. In research, we need to know what sources were used, how claims were derived, and what uncertainty remains, because the credibility of a synthesis depends on traceable reasoning. In clinical environments, provenance is equally important, though less often formalized: we need to know whether an output is grounded in guidelines, high-quality trials, local policy, or merely plausible generalizations. Agentic tools can compress steps so efficiently that they also compress our visibility into where a conclusion came from. Cost pulls these questions from theory into daily decision-making. At roughly $240 per month, this is not an impulsive subscription for most clinicians; it is closer to a staffing tradeoff. Paying that amount implicitly assumes that time saved is both substantial and dependable, and that the time we spend verifying the output does not quietly re-inflate the workload. In clinical settings, the “true cost” includes not only money, but also the cognitive burden of oversight and the reputational risk of errors. From a practice perspective, the safest near-term uses are those that keep identifiable data out of the system and keep verification firmly human. Drafting non-identifying psychoeducation templates, creating training materials for interns, turning internal procedures into clearer language, or generating first-pass outlines for research documents can be sensible, provided we treat outputs as drafts and insist on source checking. The risk profile changes sharply when we move toward identifiable case details or highly specific clinical recommendations, especially in small communities or rare presentations where re-identification can be easier than we like to admit. We also need to acknowledge a quieter limitation that experienced researchers recognize: these tools can accelerate the appearance of scholarship. They can produce coherent framing, persuasive prose, and tidy synthesis even when the evidence base is mixed or contested. The danger, then, is not only “hallucination” in the headline sense; it is routine overconfidence, particularly under deadline pressure, fatigue, or institutional incentives that reward speed over carefulness. Ethically, we should treat agentic systems as a new layer of professional delegation that demands transparency and documentation habits. If AI materially shaped an output that informs care (a clinic policy, a patient handout, a decision support memo), the clinician’s responsibility is not reduced; it is reconfigured. We owe patients and colleagues a disciplined stance on what data entered the system, what sources were relied upon, and how claims were checked. This is consistent with broader AI-risk frameworks emphasizing lifecycle governance: mapping likely failures, setting boundaries, and building habits of verification rather than relying on good intentions. Looking ahead, the question is not whether “AI computers” will become more common; they likely will. The more important question is whether they become legible: systems that make their sources, assumptions, and limitations visible enough for clinical and research cultures that depend on auditability and trust. If we adopt them thoughtfully, starting with low-risk tasks, measuring time saved against time spent verifying, and maintaining strict boundaries around sensitive data, we can treat these tools as assistants rather than authorities, and preserve the integrity of our work while reducing avoidable friction.

As AI systems enter clinical workflows, terminology needs to become operational rather than rhetorical. An AI agent refers to a model-based software system capable of planning and executing multi-step tasks toward a goal, including selecting actions, using external tools, and revising steps based on outcomes. In other words, it’s more than a chatbot reply: it’s a goal-directed workflow runner. This definition matters because a lot of confusion starts when we treat any fluent system as “agentic.” Google DeepMind just dropped a long, ~42-page warning about why most “AI agencies” will fail in the real world, and when we read Intelligent AI Delegation (published February 12, 2026), it quietly explains the same thing from first principles. The uncomfortable takeaway is simple: a lot of what’s marketed as “agents” today won’t survive contact with reality. Not because they can’t write or plan, but because they can’t delegate in the way real systems require. And therapy is exactly where that gap becomes visible early. Here’s the part that stings: most agents today aren’t really agents in the strong sense people imagine. They’re task runners with good branding, we give them a goal, they break it into steps, call tools, and return an output. That can be powerful automation, and it can absolutely create efficiency in low-stakes work. But it’s not delegation, it’s a prettier pipeline. And in healthcare, “pretty pipelines” are exactly how hidden risk slips into daily practice. So what do we mean when we say “delegation”? Delegation isn’t one thing. It can mean delegating execution (the system carries out steps you already chose), delegating workflow control (it sequences tools, retries, and manages handoffs), or delegating judgment (it decides what should be done and when, based on interpretation of context and risk). Most systems marketed as “agentic” are mostly capable of execution and some workflow control, not delegation where judgment is offloaded and accountability meaningfully shifts. In therapy, the danger is when we unintentionally delegate judgment while assuming we only delegated execution. DeepMind’s brutal point is that real delegation isn’t just splitting tasks. Delegation is transferring authority, responsibility, accountability, and trust, and doing it dynamically as the situation changes. That means we don’t only ask, “Who can do this fastest?” We ask, “Who should be trusted to do this, under these constraints, with these consequences?” Almost no current system behaves that way end-to-end once multiple tools, uncertain data, and real stakes are involved. Before delegation even happens, the framework implies we need to evaluate capability, risk, cost, verifiability, and reversibility. In other words, it’s not “does the agent have access to the calendar tool?” It’s “is this task safe to hand off, can we verify the result, and can we undo harm if it’s wrong?” That is a very clinical way of thinking, and it’s exactly why this paper hits therapists harder than it hits casual tech demos. Our work already assumes risk, not perfection. Therapy is not just information work; it’s relational work under ethics, confidentiality, and duty-of-care. Our clients don’t experience our work as “outputs”, they experience it as trust, safety, timing, attunement, and boundaries. So when an AI system is framed as an “agent,” the real question becomes: what are we delegating, and what are we accidentally outsourcing? If we let a system drift into clinical judgment, we may be delegating more than we think. When AI agency fails in our context, it often fails without obvious nonsense. It fails quietly: a note becomes subtly distorted, a risk signal gets downplayed, a client message becomes overly confident, or a safety-plan phrase becomes too generic to be safe. The output can look coherent and professional, sometimes more professional than we would write at 6 p.m. and that’s precisely what makes it dangerous. Fluency can create false reassurance, and false reassurance is a clinical risk. Current limitations of AI agents (especially in real workflows) Even when agents perform specific tasks well, they often struggle with coordination and interoperability, working reliably across tools, policies, datasets, and other agents. Multi-agent setups can become brittle: one agent’s assumption becomes another agent’s “fact,” tool errors propagate, and responsibility becomes hard to trace. This matters because real clinics aren’t single-tool sandboxes; they’re multi-system environments with partial data and shifting constraints. The limitation isn’t just intelligence, t’s dependable, governable collaboration. Because of limited interoperability, humans are still required to manage coordination between multiple agents. This is the governance problem: accountability diffusion. If one agent delegates to another, which calls a tool, which references incomplete data, who is actually accountable for the final wording sent to a client? We might sign it, the clinic might deploy it, the vendor might supply it, and the model might generate it, yet no one can point to a clear, auditable decision boundary. In therapy, that’s not abstract; it’s how ethical responsibility gets blurred. Still, we shouldn’t throw the tools away, and we should keep the narrative balanced. Agents promise real efficiency: they can reduce admin burden, draft psychoeducation we approve, structure worksheets, translate or simplify content to match literacy needs, and summarize our own prior notes with clear provenance. But practical constraints remain, so boundaries must stay explicit: delegate low-risk, reversible, verifiable tasks, and resist delegating anything that functions like diagnosis, risk assessment, crisis response, or treatment decisions. If we get that boundary right, we take the gains without surrendering what clients come for: trustworthy, human accountability.

In the past two weeks, many clinicians have seen posts about Lotus Health AI, described as an “AI doctor powered by real doctors”, a physician-supervised service that can assess, diagnose, prescribe, and refer, positioned as free, always available, and multilingual. The company also announced a $35M Series A bringing total funding to $41M, co-led by Kleiner Perkins and CRV, and argues it can cut administrative waste by making physicians more productive without insurance billing. From a therapist’s perspective, the key question isn’t whether the marketing is compelling, it’s whether this becomes a reliable front door to primary care that reduces the friction therapy often absorbs: sleep problems mistaken for depression, thyroid/anemia issues that look like burnout, medication side effects that destabilize mood, and medical uncertainty that fuels panic and health anxiety. In day-to-day work, therapy can become the place where fragmented healthcare gets processed, emotionally, and sometimes administratively, because clients can’t access timely care elsewhere. One practical benefit, if Lotus functions as described, is speed as clinical timing, not just convenience.Earlier medical clarification can make therapy interventions more accurate and effective.When a client says, “Something feels off physically,” we validate and explore patterns, yet sometimes the responsible step is: get a medical assessment early.The differential matters: if symptoms are physiological or medication-related, the plan shifts; if they’re anxiety-driven, it shifts differently.In many systems, “soon” becomes weeks, and therapy becomes a holding space for unresolved uncertainty.A 24/7 primary-care-like channel that reviews consolidated history and routes to in-person care could reduce that gap.Lotus explicitly positions itself as going beyond generic chatbot advice by involving board-certified physicians who review guidance and prescribe when appropriate. A second benefit is indirect but meaningful: it may improve the quality and coherence of information clients bring back into therapy. Many clients struggle to describe symptoms clearly due to stress, dissociation, executive dysfunction, trauma, or exhaustion. If a platform helps organize meds, labs, and prior records into a clearer story, it can reduce shame (“I’m not making it up”), sharpen insight (“my panic clusters around sleep disruption”), and help therapists choose more precise interventions (e.g., exposure for health anxiety vs urgent referral for red flags). Third, there are coordination and access benefits even if therapists never touch the platform. When clients can obtain refills, medication reviews, and referrals with fewer obstacles, therapy is less likely to be derailed by preventable destabilizers. Practically, that can mean fewer sessions spent troubleshooting access and more time spent on the core work of therapy, skills, meaning-making, attachment repair, behavior change, and identity-level integration. That said, therapists are trained to notice how tools can become part of a symptom cycle. An always-available “doctor in your pocket” may stabilize some clients, but it can also feed reassurance-seeking for others, especially with health anxiety, OCD-spectrum checking, panic, somatic symptom patterns, or trauma-related body scanning. Even if guidance is solid, repeated checking can still function as avoidance or compulsion. The goal isn’t to demonize the tool, but to integrate it into a treatment plan with clear, time-limited, values-consistent use. There are also boundary questions that show up quickly. If clients rely on app-mediated care, therapists will be asked to interpret it: “Should I trust this?” “What does this diagnosis mean?” “Can you message them for me?” A helpful posture is to treat it like any outside provider: help clients clarify questions, process impact, and decide next steps, while avoiding medicine-by-proxy. It also matters that what any physician-led digital service can do may vary by jurisdiction, licensure, and telemedicine rules, so “available” may not always mean “authorized to treat” in the way clients assume. Ethically, the business model matters because trust is a clinical ingredient. If clients suspect recommendations are influenced by sponsorships or commercial incentives (whether or not that’s true), it can erode trust in healthcare and show up in therapy as cynicism, avoidance, or hopelessness. This makes transparency practical, not philosophical: what is automated vs clinician-reviewed, and how conflicts of interest are managed over time. Privacy and data integrity are equally central. Even with strong security claims, it helps to think concretely: what data is being linked, who can access it, what consent was given, what can be deleted, and what might be retained. Data-rich systems can fail in new ways, through breaches, misuse, or overconfidence in incomplete records, so the most ethical stance is careful integration, explicit consent, and humility. Overall, I’m cautiously interested in what this could become for therapy-adjacent care: faster evaluation, less avoidable suffering, and less informal care coordination falling onto therapists. But the promise depends on whether it earns trust in real-world use, clinically, ethically, and operationally, so clients aren’t left alone with medical uncertainty that insight alone can’t resolve.

AI in healthcare is no longer a concept. It’s a reality in motion and we’re now watching it mature in a more serious, more structured way. For a long time, healthcare was treated like a “possible use case” for general AI. But in practice, it has already been one of the most common real-world arenas where people test what these systems can do. Patients ask questions after clinic hours because anxiety doesn’t follow office schedules. Caregivers try to interpret lab results while waiting for a follow-up appointment. People navigating chronic conditions search for plain-language explanations of complex treatment plans. Clinicians, meanwhile, are under relentless documentation pressure and are constantly looking for tools that reduce cognitive load without sacrificing safety. The demand wasn’t hypothetical, it was already here. What’s changing now is that major AI companies are building healthcare-specific products that admit, plainly, that medicine is not “one size fits all.” Two launches make the shift easy to see: Claude for Healthcare and ChatGPT Health. They’re often discussed in the same breath, but clinically and ethically, we shouldn’t treat them as interchangeable. They point to two different problems and two different audiences, and that distinction matters because it shapes how risk shows up. ChatGPT Health is best understood as a patient-oriented space: a place where individuals can connect personal health or wellness information and receive explanations, summaries, and context in language that feels human. The promise is clarity. Healthcare is full of jargon, fragmented portals, and rushed appointments; a tool that helps someone understand their own information could reduce confusion and improve follow-through. When used appropriately, it can support better conversations with clinicians because patients arrive with sharper questions and less overwhelm. But that same strength is also its most predictable risk. When a system explains something smoothly, people can mistake fluency for clinical authority. We’ve all seen it: a confident tone can feel like certainty, even when the underlying situation is ambiguous. In healthcare, that gap is not academic, it can shape real decisions. So the safety challenge for a patient-facing tool isn’t just accuracy in a narrow sense. It’s expectation-setting, clear boundaries, and guardrails that prevent “informational support” from being interpreted as diagnosis or medical instruction. Claude for Healthcare, on the other hand, is more naturally framed as an enterprise and workflow tool. The emphasis is less “ask me anything” and more “connect me to the work.” Healthcare organizations don’t just need answers; they need operational support: interpreting and summarizing complex information at scale, reducing administrative friction, supporting research and internal processes, and fitting into existing systems without turning every task into another tab and another login. If we’re honest about what burns clinicians out, a large portion of it lives here, in documentation, administrative tasks, and the endless effort of finding and re-finding information across messy workflows. That’s why workflow-oriented tools feel so attractive: they target the pressure points that are actually breaking the system. But again, the risk is different. When AI becomes part of a workflow, it can scale both efficiency and error. If an output is wrong and nobody catches it, the mistake doesn’t just affect one conversation it can become embedded in documentation, passed forward, copied, and normalized. The more “plugged in” the tool is, the more essential it becomes to design for oversight rather than speed alone. This is where the compliance conversation becomes more than a checkbox. Once these systems touch sensitive health information, the question isn’t simply “is it HIPAA -compliant?” or “is it GDPR-compliant?” The deeper question is: how will data governance work in real life, with real people, under real time pressure? HIPAA (Health Insurance Portability and Accountability Act) and GDPR (General Data Protection Regulation) are different frameworks, but both push toward the same discipline: clear rules about what data is collected, why it’s collected, who can access it, how long it’s kept, and what happens when something goes wrong. And here’s the point we can’t afford to gloss over: compliance is not something a company can fully “grant” to a user through a product announcement. Even if a system is designed to be HIPAA-ready or offers strong security features, organizations still need to deploy it responsibly. That means access controls, role-based permissions, audit trails, staff training, retention policies, incident response planning, and crystal-clear boundaries around what data should and should not be entered into the system. For patient-facing AI like ChatGPT Health, privacy and consent have to be especially legible, because patients don’t always realize what they’re sharing when they upload documents, connect accounts, or paste text from portals. The tool needs to prevent accidental oversharing and make it obvious when a question crosses into “you need a clinician” territory. For workflow-oriented AI like Claude for Healthcare, the burden shifts toward institutional controls: connector permissions, least-privilege access, monitoring, and accountability structures that keep “helpful automation” from becoming invisible decision-making. What’s genuinely promising in all of this is the direction of travel. We’re moving away from the idea that one generic assistant can safely serve every healthcare scenario. We’re seeing specialization: tools designed for patient understanding, and tools designed for clinical and organizational workflows. That specialization makes it easier to define what the system is for, how it should be evaluated, and where the boundaries must be enforced. Our take is simple: we’re watching healthcare AI diversify, and that’s a sign the sector is being taken seriously. But seriousness comes with obligations. These tools should assist, not diagnose. They should reduce burden, not quietly introduce new error channels. And they should handle sensitive data with governance that is operational, not rhetorical. If we build with those principles, AI can improve clarity and relieve real pressure in healthcare. If we don’t, it will scale confident-sounding uncertainty into the one domain where people can least afford it.