English

As researchers, we spend a surprising amount of time doing work that is not, strictly speaking, research. We design studies, refine hypotheses, collect and analyze data, and engage deeply with theory and then we open PowerPoint or Illustrator and begin the meticulous process of turning our methods and findings into figures. We adjust arrows, realign boxes, standardize fonts, correct axis labels, and export multiple versions to meet submission guidelines. The science may be rigorous, but the path from text to publication-ready visuals often feels inefficient and cognitively draining. Lately, we’ve been looking at PaperBanana, a text-to-figure AI tool that’s clearly targeting researchers who want to speed up figure creation without sacrificing academic clarity. The core promise is simple: turn method descriptions into structured methodology diagrams, and turn data into charts in a way that’s meant to stay faithful to what we actually did (not just “look plausible”). The detail that makes it feel more “research-first” than generic image generators is the idea of a closed-loop workflow: instead of one-shot output, the tool aims to draft, check, and refine—so the figure is more likely to match the logic of the method. The “plus” here isn’t that we get a pretty figure in seconds. The real added value is that we get cheap iteration. When iteration is cheap, we stop freezing figures too early. We can generate multiple drafts, compare them, and treat figure-building like writing: draft → critique → revise. That’s where quality usually improves, not because the AI is perfect, but because our feedback loop gets faster and less exhausting. At the same time, it’s worth being honest: Illustrae plays a different game. It’s typically the better fit when the job is not just “create one diagram,” but assemble a full visual story, multi-panel figures, posters, teaching material, and lots of manual layout decisions. Illustrae tends to offer more features and flexibility, with more ready-to-use options to manage variables, layouts, iterations, and visual adjustments. The tradeoff many people feel is cost: it’s often described as significantly more expensive (and/or less predictable) than a straightforward researcher subscription, which can be a barrier for individuals or small labs. PaperBanana, in contrast, feels more minimalist and research-focused. For method diagrams and conceptual figures, especially at the drafting stage, it can be comparable in performance, while being more affordable, which matters a lot for labs, PhD students, and early-career researchers. It’s also commonly framed as being built on NanoBanana Pro, positioned as stronger for structured scientific visuals than general image models like DALL·E 3 (at least for diagram-like outputs where structure and labels matter more than “art style”). Here’s the comparison in a clean table (pricing stars are about budget-friendliness + predictability, not “how expensive the company is in absolute terms”): Tool Best for Pricing style (practically) Public price (USD, as listed) Price predictability (1–5) Value for researchers (1–5) Flexibility / control (1–5) Strengths (advantages) Tradeoffs (be honest) Illustrae Posters, multi-panel figures, teaching material, complex layouts Subscription + credits paid pricing not publicly posted (custom quote) (illustrae.co) 2 4 5 More features; more layout control; strong for assembling and polishing big visuals Can be significantly more expensive / less predictable; cost can block individuals/small labs; still needs expert review to avoid conceptual drift PaperBanana Methods diagrams, conceptual figures, fast drafting of research visuals Subscription + credits From $4.90/mo (annual billing) for 100 credits; $6.90/mo for 400; $19.90/mo for 1,500 (paperbanana.studio) (Alt. “credit plans” page also lists $14.90/mo, $59.90/mo, $119.90/mo tiers.) (paperbanana.org) 4 5 3 Minimalist and research-focused; fast drafting; good for method diagrams; affordable enough to be realistic for students/labs No meaningful free tier for deep exploration (subscription needed); final figures still need human refinement; confidentiality policies may not fully guarantee exclusion of unpublished work from training Now the limits, because this is where “AI for academia” either becomes useful or becomes risky. First, human review remains essential. These tools accelerate drafting, but final figures still need expert refinement: labels, implied causality, statistical meaning, and whether the diagram accidentally over-claims what the method can do. Second, subscription friction is real. If there’s no meaningful free tier, adoption becomes a budgeting decision, not a quick experiment, especially for students. Third, confidentiality is still a question. Unless a tool makes an explicit, strong guarantee that unpublished papers/figures are excluded from model training (and clarifies retention), we should be cautious with sensitive or pre-publication material. And yes, we can always rely on classic AI tools like Notebook or NotebookLM for summarizing, outlining, or restructuring ideas. They’re great at text workflows, but they’re not built specifically for researchers’ visual needs, and they’re typically less precise for scientific diagram conventions, which increases the risk of subtle visual or conceptual inaccuracies compared with tools designed for academic figures. So when we ask “what’s the plus?”, it’s this: we’re buying back attention. Not just time, attention. If figure drafting becomes fast enough that we can iterate without dread, we can redirect our effort to what actually moves research forward: clearer hypotheses, sharper methods, better interpretation, and figures that communicate rather than decorate. Choosing the right tool isn’t about hype, it’s about fitness for scientific purpose.

If you have been in practice long enough, you probably remember how your early years were shaped by the unglamorous parts of the job. Notes, reports, intake forms, scheduling messages, scoring, and endless documentation clean up. It was exhausting, but it was also part of the apprenticeship. Writing things down forced us to clarify what we saw, what we thought it meant, and why we chose a particular next step. Now artificial intelligence is stepping into that exact layer of clinical life across occupational therapy, speech therapy, psychology, and physiotherapy. Not in a dramatic “robots are replacing therapists” way, but in a practical, everyday way. AI can draft documentation, summarize long text, organize information, and generate first pass templates. That changes what clinics measure, what managers expect, and what early career clinicians feel pressured to deliver. From our perspective as clinicians, the biggest shift is that AI is compressing parts of the learning curve. A new graduate can produce something that looks polished very quickly, and sometimes that polished output can hide the fact that clinical reasoning is still developing. A confident sounding paragraph is not the same as a sound formulation. A tidy plan is not the same as an individualized plan. When caseloads are heavy and supervision time is limited, it becomes easier to mistake speed for competence, and that is where risk quietly grows. So the entry level bar is moving. If routine tasks become faster, the expectation becomes that the clinician will contribute more of what cannot be automated. Stronger clinical reasoning shows up earlier. Ethical judgment stops being a once a year training topic and becomes a daily decision, especially around privacy, consent, bias, and what information should never be entered into a public tool. Digital literacy also becomes part of professionalism, not because we need to be tech experts, but because we need to understand enough to use tools responsibly and explain their limits. We also think it helps to be very clear about what AI is good for in real practice. It can support preparation, structure, and efficiency. It can help us draft, brainstorm, and organize. But it cannot hold accountability. We still have to review every output like we might be questioned on it. We still have to check sources when research is summarized. We still have to tailor every plan to the person in front of us, because generic recommendations can be subtly wrong for a client’s context, culture, risks, and goals. For supervisors, clinic owners, and senior clinicians, there is a parallel responsibility. If AI reduces documentation time, we should be intentional about where that time goes. It can go toward better reflection, clearer consent conversations, more collaborative care, and stronger follow up. Or it can go toward squeezing in more sessions while calling it innovation. Only one of those choices protects client centered care, and only one supports the development of safe clinicians. In the end, we do not see AI as a replacement for therapy: it reshapes the role. The therapists who will thrive are those who think clearly and use tools carefully, while protecting what stays deeply human: relationship, pacing, and ethical judgment. That also means using AI responsibly: keep client information anonymized/de‑identified whenever possible, and prefer local, offline systems (on-device or a secure internal server) over sending data to online tools. If a cloud-based tool is used, it should be only with explicit informed consent, strict limits on what is entered, and a transparent explanation of where data goes, who can access it, and how it is stored.If AI gives us anything worth keeping, it is time to do more of what only clinicians can do. Resources

Toward “Right‑Fit” Care: How AI Is Personalizing Mental Health Treatment Therapy has always included a careful kind of uncertainty, and as psychologists and mental health professionals we know how much skill it takes to work inside that uncertainty without trying to “solve” a person too quickly. We listen, assess, build a formulation, choose an evidence‑based starting point, and then we learn with the client what actually works. Even when we practice well, the early phase can feel like trial and error: a few weeks of testing whether this structure, this pacing, and this approach truly fit this person. For clients who are already exhausted or at risk, that delay matters. AI is beginning to shorten that “finding out” period by using personal data, shared only with explicit permission, to support earlier, more precise decisions. Instead of relying mainly on retrospective self‑report (“How was your week?”), we can add real‑world signals from phones and wearables: sleep duration and regularity, movement and sedentary time, daily rhythm, time spent at home versus outside, and changes in routine or social connection. In some specialized settings, clinicians and researchers also explore brain data (for example, MRI or EEG measures) to add information about brain circuitry and patterns that may relate to symptom profiles or treatment response. The aim is not to replace clinical judgment, but to strengthen it. The practical shift is moving from a snapshot at intake to a living picture of the client’s week. Self‑report is essential, but memory is imperfect and symptoms can blur recall. Passive and semi‑passive data can reveal patterns clients often feel but cannot easily name. If a client says they are “fine,” yet their sleep is fragmenting and their activity is steadily dropping, we have a compassionate entry point for deeper exploration. If anxiety spikes reliably at certain times and contexts, we can stop treating it as random and start treating it as predictable. This is where AI helps: it can analyze large, messy time‑series data and detect relationships humans would miss, what tends to happen before a mood drop, what predicts irritability, or what combination of isolation and sleep disruption precedes self‑harm urges. Think of it as a translation table from signals to clinical hypotheses. Sleep variability may indicate reduced emotion regulation capacity and relapse vulnerability. Reduced movement may point toward avoidance and anhedonia, suggesting behavioral activation or values‑based action. Abrupt routine changes may signal interpersonal rupture, shame, or safety concerns. The data does not diagnose; it helps us ask better questions sooner and refine the plan faster. You also pointed to a future‑leaning idea: combining brain scans with smartphone and wearable data to estimate the best intervention before a long course of trial and error begins. This direction is promising, but it demands caution. Some models can predict treatment response in research settings, yet they may not generalize across populations, devices, and real‑world complexity. Used ethically, these tools should function like decision support, a second opinion, never an automatic decision-maker. One of the most immediate benefits is timing. A growing class of tools aims to deliver support when symptoms are most likely to spike (often described as “just‑in‑time” interventions). Weekly therapy teaches skills, but the real test is whether clients can access them at 11 pm when exhausted, during a commute when panic builds, or right after conflict when urges rise. If data shows a reliable pattern, sleep disruption followed by next‑day agitation, or isolation followed by late‑night rumination, digital supports can be timed to the risk window: a brief grounding prompt, a coping-plan reminder, or a micro‑exercise that reconnects the moment to the formulation you built together. At their best, these tools feel like a bridge between sessions, not surveillance. These advancements could also expand access in a world of provider shortages. Not everyone can attend consistent sessions, and many people reach care only during crisis. Carefully designed digital supports can offer tailored continuity for those who struggle to access services, while keeping therapy human, relational, and collaborative when sessions occur. The ethical boundaries are non‑negotiable. Personal data must be opt‑in, purpose‑limited, and easy to pause or stop. The safest approach is minimalism: collect only what answers a clinical question. In many cases we do not need private content (messages, audio, contacts); we need patterns (sleep, movement, routine) and brief check‑ins. Whenever possible, the information should be anonymized or de identified (remove names, dates of birth, exact addresses, contact details, record numbers, and any unique identifiers) so it cannot reasonably be traced back to a person. The same logic applies to generative AI used for clinical writing or support: to protect confidentiality, it should ideally be local (downloaded and running on a device or secure internal server) rather than sending client information to an online system. If a cloud based tool is used, it should only be with explicit informed consent, clear limits on what data is entered, and a transparent explanation of where the information goes, who can access it, and how it is stored. And if a tool touches suicidality, the agreement must be explicit: what is monitored, who sees it, what triggers an escalation, and what the tool is not responsible for. Any system claiming it can “detect suicide risk” should be treated like a high‑stakes clinical claim requiring strong evidence, transparency, and a clear safety protocol. So how do we integrate this as therapists without losing the heart of therapy? Start small and clinical. Choose one target (sleep, panic spikes, avoidance, self‑harm urges). Collaboratively pick one or two metrics that feel helpful rather than invasive. Decide together how insights will be used, shaping session focus, planning for predictable windows, or evaluating whether a new intervention is working. Then review it like any intervention: Did it increase agency or self‑criticism? Did it clarify patterns or add pressure? Integration succeeds when the client feels more choice, more clarity, and faster relief. If we insist on evidence, close monitoring, clinical involvement, and regulation, AI can reduce unnecessary suffering by helping us match people

We keep noticing that when clinicians talk about “using AI,” we’re often talking about two very different approaches, even if we’re using the same tool. And the confusion usually shows up around one word: automation. People hear “automation” and imagine a cold replacement of therapy, or they assume it’s basically the same thing as collaboration. In practice, automation in clinical work is simpler and more grounded than that. It is not “AI doing therapy.” It is the clinician delegating repeatable steps in the workflow, then supervising the output the way we would supervise an assistant or a trainee. In procedural mode, AI becomes a substitute for execution. We ask, it answers; we paste, we send. The output is used for efficiency: quicker drafts, quicker wording, quicker structure. That can genuinely reduce load, especially on days when we’re holding multiple cases and still trying to document, plan, and communicate clearly. But procedural mode also has a built-in risk: it can bypass the step where we ask, “What claim did this just make, and do I actually have the clinical data to stand behind it?” In therapy, where work is high-stakes and context-sensitive, skipping that step is never a small thing. Collaborative mode looks different. Here, AI is treated more like a thinking partner that helps us refine what we already know. We provide context, constraints, and objectives, and we actively evaluate and revise what comes back. The benefit isn’t only speed, it’s quality. As goals become more complex, the work doesn’t disappear; it shifts upward into framing, supervision, and judgment. That shift is the point, because it mirrors what good therapy already is. The core value isn’t “doing tasks.” The core value is choosing what matters, staying accountable to the formulation, and tracking whether what we are doing is actually helping this client in this moment. With that clarity, the question “where does automation fit?” becomes easier: automation belongs around the session, not inside the relationship. It supports the repetitive work that quietly drains clinicians, so you show up with more focus and presence. In practical terms, this often starts with answering emails: drafting scheduling replies, boundaries, first-contact responses, follow-ups, and coordination messages with parents or schools. AI can give you a clean draft fast, but the clinician still protects tone, confidentiality, and the therapeutic frame before anything is sent. Automation can also support assessment workflows, especially the mechanical parts like cotation and report organization. It can help format tables, structure sections consistently, and draft neutral descriptions, saving time without pretending to “interpret.” Similarly, it can help with filling questions for you: generating intake questions, check-in prompts, or between-session reflection questions tailored to your model and the client’s goals. That doesn’t replace clinical judgment; it simply gives you a clearer scaffold for information-gathering and tracking change. Another high-impact area is session preparation. If you provide a brief, non-identifying summary of the last session, AI can help draft a focused plan: key themes to revisit, hypotheses to test, reminders of what was agreed on, and possible questions or interventions that match your orientation. The point is not to “script therapy,” but to reduce the mental load of reconstructing the thread so you can start the session grounded. More sensitive, but sometimes very helpful, is using automation around session recording and documentation (only with explicit consent and a privacy-safe system). AI can assist with transcripts, highlight themes, and draft a note structure or summary. Still, this must remain supervised: AI can miss nuance, misinterpret meaning, or phrase things too strongly. In clinical documentation, accuracy and accountability matter more than speed, so the clinician always verifies what’s written, especially around risk, safety planning, and any diagnostic or medical claim. Finally, automation can support what many clinicians want but struggle to do consistently: progress comparison over time. Whether you use outcome measures, session ratings, goals, homework follow-through, or narrative markers, AI can help summarize shifts from baseline, spot patterns across sessions, and draft a short “what’s improving / what’s stuck / what to adjust next” reflection. The tool organizes and surfaces patterns; you decide what it means and what the next clinical step is. All of this only works if we pay attention to data and privacy. We avoid entering identifying information unless we are using an approved, privacy-compliant system. We do not treat AI output as truth, especially for diagnosis, risk assessment, medication-related topics, or any medical claim. And we keep the clinician role explicit: AI can generate language, options, and structure, but we provide judgment, ethics, and accountability. This is also why many clinicians are drawn to running a private generative model locally on their laptop, offline, so data does not leave the device. Even then, strong device security and clear consent practices still matter, but the direction is sound: protect client information first, then build workflow support around it. When we use AI with this mindset, the payoff is real. We gain time and mental space for what cannot be automated: attunement, formulation, pacing, rupture-repair, and the relationship. The tool handles parts of the scaffolding and we protect the heart of therapy, which is slow, context-sensitive, and deeply human work.

If we’ve ever tried to write a real manuscript after a full day of research work, we already know the problem usually isn’t that we don’t have ideas. We do. The problem is that academic writing demands a specific kind of mental steadiness. We have to hold a thread, keep structure in mind, track definitions precisely, and build a coherent argument while our brains are already full of meetings, supervision, grant deadlines, data cleaning, reviewer comments, and the constant micro-decisions that come with running studies. So when OpenAI announced Prism, it caught our attention for a surprisingly practical reason. It sounds like it’s designed to reduce overwhelm in the writing process, not by writing the paper for us, but by making the environment less fragmenting and more supportive of sustained attention. OpenAI describes Prism as a free, cloud-based, LaTeX-native workspace for scientific writing and collaboration, with an AI assistant integrated into the document workflow. And that phrase, integrated into the workflow, matters. Many of us still write in a patchwork setup. The draft lives in one place, citations in another, PDFs in folders we swear are organized, tables in spreadsheets, figures in separate tools, and formatting rules that feel like a moving target. If we use AI, it often sits off to the side in a separate window, with no real awareness of what the document actually contains. Prism is pitching something different. One workspace where drafting, revising, compiling, and collaborating live together, so we don’t constantly switch contexts and lose momentum. That sounds less like automation and more like good research infrastructure. Something that helps us keep the argument intact while we spend our limited energy on what actually matters: methods, logic, interpretation, and the discipline of not overclaiming. What we also appreciate is that Prism seems aimed at the boring practical problems that quietly wreck productivity. Collaboration, comments, proofreading support, citation help, and literature workflow features are not flashy, but they are exactly the kind of friction that makes us close the laptop and tell ourselves we’ll do it tomorrow because we can feel the administrative drag consuming what’s left of our focus. And if we’ve ever co-authored a paper, we know how much time gets lost to version control, merging edits, and re-checking what the “current draft” even is. A shared cloud workspace can reduce that overhead by keeping writing and collaboration in one place. Here is where the researcher angle comes in. Researchers are trained to track nuance, uncertainty, and the limits of what data can actually support. Many of us can write well when we have space to think. But research rarely gives writing prime attention. Writing happens in stolen hours between analyses, teaching, project management, and funding applications. That changes what “helpful technology” looks like. We don’t just need a tool that generates text. We need a tool that helps us stay oriented so we can turn results into clear contributions publishable, teachable, and useful. Prism might support that kind of work, especially for researchers who publish, teach, supervise trainees, or collaborate across institutions and need their writing process to be less chaotic. If it truly reduces friction, it could help more of us finish what we start—not because the tool has better ideas than we do, but because it helps protect the continuity of our thinking. At the same time, we should say the quiet part out loud. A smoother writing workflow doesn’t automatically mean better science. AI can help us sound coherent and academic, and that can be useful, but it is also where risk shows up because polished writing can hide weak reasoning. So if we use Prism, we should treat it like a very fast assistant. It can reduce friction and help us express what we mean, but it is not the source of truth. We still own the reasoning, the claims, the citations, and the integrity of the work. And of course, Prism is not the only tool that exists. Most of us have already used other AI tools before, along with specific writing and reference managers that keep our workflow moving. What makes Prism feel different, at least from the way it is described, is the promise of one integrated workspace and the fact that it is free. If it delivers even half of that, we honestly cannot wait to explore it more. Where we land is simple. Prism sounds promising because it aims at the real pain points in research writing: context switching, formatting drudgery, collaboration friction, and the cognitive load of keeping a complex document coherent over time. Not magic. Not a replacement for expertise. But possibly the most researcher-friendly kind of productivity tool—the kind that helps us keep the thread.