Learn to add before you get the calculator.
When I introduced BastionGPT into my doctoral training program, the results from my strongest trainees were immediate. Students who understood clinical data, who could look at a table of test scores and identify the clinically significant findings, used the platform to draft polished evaluation reports in a fraction of the time. Their output was grounded in real analysis, structured around the referral question, and ready for a supervising psychologist to provide feedback. BastionGPT refined the prose while the trainees drove the clinical reasoning. The combination produced draft reports that were faster to complete and often stronger than what either could produce alone. If you know what you are doing, AI makes you look better. It is like handing a calculator to someone who already knows how to add.
That is the value proposition for training programs considering AI adoption: a HIPAA-compliant clinical AI platform that amplifies the skills trainees have already built. BastionGPT’s clinical knowledge base, document analysis capabilities, and healthcare-appropriate content filtering make it a strong fit for supervised training environments where compliance and accuracy matter.
I also discovered something I wasn’t expecting. BastionGPT became one of the fastest ways to identify which trainees had genuine clinical skills and which ones had gaps. The platform revealed gaps that already existed.
Trainees who skipped the foundational analytical work and fed raw, unorganized data into BastionGPT got back reports that read well on the surface but missed the clinical question entirely. I call these reports “hollow”: well-written summaries that say nothing diagnostically meaningful. The quality of the AI output directly reflected the quality of the clinical thinking that went into the prompt. The output may be well written and may summarize what was in the document, but everything written in a psych report needs to speak to the referral question. Instead, what I get back is something generic that does not speak to the relevancy of the question.
This pattern, AI amplifying what is already there, applies across healthcare training. It surfaces wherever a tool can produce polished output from shallow input. Psychology training is one vivid example. The principle holds for medical residencies, allied health programs, legal training, and any field where structured reasoning must precede clear communication.
The encouraging part: this is a solvable problem. My training program offers a practical model for scaffolded AI adoption, a structured progression that builds foundational skills first and introduces BastionGPT as augmentation only after those skills are verified. The framework has worked for my strongest trainees, and it is adaptable to other disciplines and training environments.
The Calculator Analogy: Why AI Augmentation Requires Foundational Skills
I use a metaphor that has become central to how I talk about AI in training: the calculator. I know my trainees can add, but it is more efficient for them to use the calculator. However, there still needs to be a way to know that they know how to add.
The analogy maps cleanly. In elementary math education, students learn their numbers, then addition, then multiplication, then long division. Only after they have demonstrated fluency with the underlying mechanics do they receive a calculator. The calculator accelerates their work. It frees them to tackle more complex problems.
AI works the same way in professional settings. A clinician who understands how to analyze test scores, interpret behavioral data, and synthesize information across multiple sources and informants can use AI to draft a report in a fraction of the time. The clinician provides the clinical reasoning and drafts the initial content; the AI refines and polishes the prose. The result is faster, sometimes better, and grounded in genuine clinical reasoning and judgment.
A trainee who skips the foundational steps gets a different result. The result is a beautifully written report that misses the point. They get a written output that fails to connect the data to the referral question, and in clinical supervision they cannot explain their reasoning. An experienced supervisor will readily identify empty writing within the first paragraph or by asking the trainee to explain their clinical rationale. What I find with AI is: if you know what you are doing, it makes you look better. If you do not know what you are doing, it makes you look worse.
This is the amplification principle. AI amplifies what is already there. Strong foundations produce stronger output. Weak foundations produce output that exposes the weakness more visibly than the trainee working without AI ever would have.
Recent research supports this observation across disciplines. A 2025 New England Journal of Medicine review by Abdulnour, Gin, and Boscardin introduced a taxonomy of AI-related skill risks in medical education: “deskilling” (erosion of existing competencies through overreliance), “never-skilling” (failure to develop foundational competencies in the first place), and “mis-skilling” (adoption of AI-generated errors as learned patterns). The authors proposed the DEFT-AI framework for structuring clinical supervision around AI interactions, emphasizing that foundational knowledge is essential to maximizing AI benefits and minimizing risks. I arrived at the same conclusion independently through direct implementation with my trainees.
A landmark field experiment published in Proceedings of the National Academy of Sciences (Bastani et al., 2025) quantified the risk. Researchers gave nearly 1,000 high school math students access to GPT-4 tutors. Students using an unguarded ChatGPT interface improved their performance by 48% during practice, but scored 17% worse than students who never had AI access when the tool was removed. Students using a version with built-in safeguards (hints instead of answers, teacher-designed guardrails) largely avoided these negative effects. The study’s conclusion aligns precisely with my calculator analogy: unfettered AI access can function as a crutch that undermines skill acquisition, while structured AI use preserves learning.
Show Your Work: Structured Pre-Work as an AI Verification Layer
My solution draws on a principle familiar to anyone who has graded a math exam: require students to show their work.
My training program uses a structured graphic organizer often taught in graduate psychology programs and in forensic or investigative environments that is built around the acronym RIOT (Records, Interviews, Observations, Tests). In psychology, for each referral question, trainees populate the RIOT matrix with evidence from each data source. If the question is whether a child has ADHD, the Records column captures report card data showing missing assignments and low behavioral ratings. The Interviews column captures parent and teacher reports of inattention, impulsivity, or distractibility. The Observations column captures behaviors noted during the evaluation and observation of the student at school. The Tests column captures performance on standardized measures.
After populating the matrix, trainees write a hypothesis: based on the evidence they have organized, what do they believe the answer to the referral question is? In supervision, they defend that hypothesis. The supervising psychologist asks questions, probes their reasoning, and evaluates whether the trainee understands the data they collected.
Only after the RIOT matrix is complete and reviewed does the trainee use BastionGPT to work on a draft narrative. Because the matrix provides the referral question, the organized evidence, and the trainee’s hypothesis, BastionGPT receives the structured clinical input it needs to generate output that speaks directly to the diagnostic question. A student eager to learn loves this structured process. They review the information and plug it into the matrix as they go.
The matrix serves two functions simultaneously. It is a pedagogical tool that builds the cognitive skill of clinical data synthesis and a verification mechanism that lets supervisors confirm the trainee has done the thinking before they work with the AI to draft the narrative. I tell my students, the RIOT matrix is the scratch paper. If you cannot produce the scratch paper, the AI did the work, and you didn’t learn a thing.
A Scaffolded AI Adoption Framework for Clinical Training
My approach follows a progression that other training programs can adapt to their own domains. The underlying logic is sequential: foundational skills first, then AI as augmentation.
Stage 1: Teach the fundamentals without AI
Trainees learn how to read medical and educational records, extract clinically relevant information from prior evaluations, identify significant test scores in the current assessment, and organize data around a referral question. They practice these skills manually. The goal at this stage is fluency with the core analytical process.
Stage 2: Require structured pre-work
Before any AI touches the output, trainees complete a structured organizer (the RIOT matrix or its equivalent) that demonstrates they have done the cognitive work. This pre-work is reviewed and discussed in clinical supervision. The supervisor verifies that the trainee can defend their analysis, identify clinically significant findings, and verbally articulate a well-reasoned hypothesis.
Stage 3: Introduce AI as a drafting tool
With verified pre-work in hand and prior proof that they can draft a clinical impression independently, trainees use a clinical AI platform like BastionGPT to assist in developing a draft report. The AI receives organized, question-driven input and produces draft output that reflects the trainee’s clinical reasoning. The trainee reviews and refines the draft, cross-referencing it against the source data. The choice of platform matters here: a HIPAA-compliant tool built for clinical language produces output that aligns with diagnostic conventions in ways that general-purpose AI tools do not. My team and I have tested multiple platforms using de-identified patient information and find that BastionGPT’s clinical knowledge base outperforms general-purpose AI tools.
Stage 4: Supervise the full workflow
The supervising psychologist reviews the completed RIOT matrix alongside the draft report. Discrepancies between the pre-work and the output flag problems. A report that discusses findings absent from the matrix, or a matrix that highlights significant scores the report ignores, signals that the trainee either did not verify and revise the AI output carefully or did not understand the data in the first place.
The Faculty Skill Gap: Why Clinical Supervisors Resist AI
AI deskilling is a training problem, and training problems require skilled trainers. This is where a second, less discussed gap emerges: many clinical supervisors and faculty members lack the frameworks to teach responsible AI use in their training programs.
I encounter this resistance among my peers in both clinical and academic settings. Some universities have responded to AI by banning it outright. Some have reverted to requiring handwritten papers in class using blue examination booklets, retreating from word processing entirely to eliminate the possibility of AI-generated submissions. I do not think it is that they do not want to grow. I think it is that they do not have the skills to jump into this next phase.
Faculty members who built their careers in a pre-AI environment have well-developed skills for evaluating student work produced by traditional methods. They know what a competent paper looks like when it was written by a human being. They know how to spot weak reasoning, poor organization, and insufficient evidence in that context.
What they lack is a framework for evaluating work produced with AI assistance. They cannot easily distinguish between a student who used AI to polish genuinely strong analysis and a student who used AI to generate a veneer over shallow understanding. Moreover, the AI-generated text verification tools they currently rely on are increasingly unreliable for detecting AI use. And the alternative methods (structured pre-work requirements, matrix-based verification, supervision) require new skills and new workflows that faculty and clinical supervisors may not have been trained to implement.
The result is a predictable pattern. Faculty and supervisors who feel unequipped to supervise AI-assisted work default to prohibiting AI altogether. The prohibition delays the reckoning to the next stage of their training or career, where the tools will be available, and no one will be checking.
The scaffolding needs to start with the educators. Before faculty and supervisors can teach students to show their work, they need their own framework for checking it.
Beyond Psychology: AI Deskilling Across Healthcare and Professional Training
The AI deskilling dynamics I have described in psychology training are present in every professional field where AI can generate polished output.
A 2025 study published in The Lancet Gastroenterology & Hepatology found that endoscopists who routinely used AI assistance during colonoscopies saw their adenoma detection rates drop from 28.4% to 22.4% when AI support was removed, even after only three months of AI-assisted practice. A Communications of the ACM analysis of the broader deskilling phenomenon noted that knowledge workers using generative AI reported tasks feeling cognitively easier, while researchers observed them ceding problem-solving expertise to the system. Senior professionals benefited; junior ones lost ground. The pattern I observe in my training program mirrors what researchers are documenting across medicine, law, education, and technology.
The pattern repeats across professions. A junior lawyer drafts an AI-generated motion without understanding the legal standard it invokes. A medical resident produces an AI-generated differential diagnosis without understanding the pathophysiology. A financial analyst submits an AI-written investment memo without doing the underlying valuation work. In each case, fluent output creates the appearance of competence, and the appearance holds until a supervisor asks probing questions, a case deviates from the standard pattern, or a client needs genuine expertise.
The solution is also the same across fields. Teach the foundational skill. Require evidence that the foundational skill has been learned. Then introduce AI as a tool that accelerates work that the professional already knows how to do.
The specifics will differ by field: case analyses before AI-assisted briefs in law, verbal diagnostic reasoning before AI-generated notes or diagnoses in medicine, manual financial models before AI-generated presentations in consulting. The underlying principle is constant: the structured pre-work is where the learning happens, and the AI-generated output is where the efficiency happens. Both are essential, and each stage requires its own rigor.
Responsible AI Use in Training Programs: What to Do Now
My experience suggests a set of practical steps that any training program can begin implementing.
Define the foundational skills explicitly
Every profession has a core set of analytical and reasoning skills that AI cannot replace. Name them. In clinical psychology, it is the ability to extract clinically relevant information from patient observation, patient interviews, direct examination, medical and educational records, and input from collateral sources, connect that data to the referral question(s), and synthesize findings to inform diagnosis and treatment or intervention planning. In your field, it will be something different. Define it clearly enough that you can assess it.
Build a “show your work” requirement
Create a structured artifact that trainees must complete before AI touches the output. The RIOT matrix is one version. An annotated outline, a data synthesis worksheet, a problem formulation template, a decision tree with cited evidence: the format matters less than the function. The artifact must demonstrate that the trainee has done the cognitive work.
Gate AI access to verified readiness
Treat AI access like calculator access in a math classroom. Trainees earn it by demonstrating foundational competence. This may mean completing an initial set of assignments entirely without AI, passing a skills assessment, or receiving supervisor sign-off on their structured pre-work before they begin using AI tools for drafting.
Train the trainers
Faculty and supervisors need their own professional development around AI-assisted workflows. This includes learning how to evaluate AI-assisted output, how to distinguish between strong AI use and hollow AI use, and how to design verification systems for their specific domain. The supervisor skill gap is as urgent as the trainee skill gap.
Enforce the process, not the prohibition
Requiring structured use of AI builds the skill that blanket prohibitions leave undeveloped. When a trainee submits a draft report without a completed pre-work artifact, the response should be the same as it would be for a math student who submits an answer without showing their work: send it back. The pre-work is the deliverable. The report is the product of the pre-work.
Scaffolding AI Adoption: The Path Forward
Too many articles now are talking about the dangers and the ethics. But that is a how-not-to. The toothpaste is out of the tube. AI is here and already embedded in many aspects of daily life. It is a matter of how you make this an effective part of your workflow. The training programs that will produce the strongest professionals in the next decade are the ones that figure out the scaffolding. They will teach the fundamentals with the same rigor they always have. They will add structured verification layers that did not exist before AI made them necessary. And they will introduce AI as a tool that rewards the work their trainees have already done.
AI amplifies what is already there. A well-trained professional with AI is faster, sharper, and more productive. A poorly trained professional with AI is a liability wrapped in fluent prose. The difference comes down to whether someone learned to add before they got the calculator.
My training program demonstrates what responsible AI adoption looks like when the scaffolding is in place. My strongest trainees complete the RIOT matrix, verify and explain their reasoning in clinical supervision, and then use BastionGPT to develop reports grounded in their own clinical analysis. The platform’s HIPAA compliance means data stays protected throughout the workflow. Its clinical knowledge base means the output reflects diagnostic conventions and terminology that general-purpose AI tools frequently miss. And the structured pre-work means the trainee, the supervisor, and the AI are all working from the same evidentiary foundation.
The question facing training programs is whether they will build the scaffolding that makes AI a force multiplier for their trainees or leave their trainees to figure it out on their own. I encourage you to establish your domain-specific framework, build the scaffolding, and give the proverbial calculator to trainees who have already learned to add.
Frequently Asked Questions
Should students use AI from their first day in a training program?
It depends. With the right guardrails, BastionGPT can add value from day one when trainees use it for brainstorming, exploring ideas, or pressure-testing their own thinking. Where training directors should set clear boundaries is around final work product: require demonstrated competence in foundational skills (data interpretation, clinical reasoning, writing structure) before trainees use AI to draft deliverables that will be reviewed, signed, or shared with patients and families.
How can I tell if a student used AI without doing the analysis first?
Look for output that summarizes data without connecting it to the clinical question. A well-analyzed report links specific findings to specific diagnostic criteria. A hollow, AI-dependent report restates what was in the source data without clinical interpretation. Requiring structured pre-work artifacts (like a RIOT matrix or data synthesis outline) makes the gap visible before the trainee’s draft report is submitted.
What is a RIOT matrix?
RIOT stands for Records, Interviews, Observations, and Tests. It is a graphic organizer used in forensic work and psychological evaluation training. Trainees enter relevant data from each source category with a specific referral question in mind, then synthesize their findings into a hypothesis. The completed matrix serves as both a learning tool and a verification artifact that proves the trainee completed the clinical reasoning before using AI to draft. Training programs in other fields can create equivalent structured organizers tailored to their domain.
Can AI replace clinical training?
No. AI can produce polished text, format tables, organize data, and draft narratives. It cannot develop clinical judgment, interpret ambiguous findings, or understand the human context of a patient’s presentation. Training programs that treat AI as a replacement for skill development will produce graduates who can generate professional-looking documents without the clinical competence to stand behind them, the skill to prescribe appropriate interventions, and the ability to respond to patient questions in a helpful way.
How do I introduce AI to faculty or supervisors who are resistant?
Acknowledge that their concerns are grounded in real problems. Students do skip steps with AI. AI output can mask skill deficits. The response to these problems is a structured framework for responsible use. Share concrete examples of what scaffolded AI adoption looks like: pre-work requirements, supervision checkpoints, progressive access. Faculty resistance often stems from not having the tools to teach and evaluate AI-assisted work, so providing those tools is the most productive first step.
Does this framework apply outside of psychology?
Yes. Any profession where AI can generate polished deliverables from raw input faces the same deskilling risk. Medical scribing, coding and billing, lesson planning, legal document drafting, patient education, care planning: the principle is the same. Build the foundational skills first, require evidence of understanding before AI enters the process, and verify that the final output reflects real professional reasoning.
What are the signs that a training program is ready to introduce AI?
Readiness depends on two factors: trainee competence and supervisor capability. Trainees should be able to demonstrate the core analytical skills of their discipline without AI assistance before they begin using AI for drafting. Supervisors should have a framework for reviewing AI-assisted work, including structured pre-work requirements and the ability to distinguish between grounded and hollow output. If either side lacks these foundations, introduce them before introducing AI.
How does HIPAA compliance factor into AI-assisted training?
Any training program that involves patient data must use AI tools that are designed to support HIPAA compliance. BastionGPT includes a Business Associate Agreement (BAA) with all plans, encrypts data in transit and at rest, and does not share data with third-party AI providers for training. These protections apply whether the user is a licensed clinician or a supervised trainee. Using consumer AI tools like standard ChatGPT for work involving protected health information introduces compliance risks that training programs should address directly in their AI policies.
What does BastionGPT offer for training programs?
BastionGPT is a HIPAA-compliant AI assistant and AI scribe built for healthcare workflows. Features relevant to training programs include document upload and analysis (up to 500 pages on Professional Plus), a saved prompts library that supervisors can share with trainees using the Share Prompts feature, healthcare-appropriate content filtering that allows discussion of sensitive clinical topics, and unlimited transcription with multi-speaker recognition. A BAA is included with all plans. Plans start at $20/user/month, with a 7-day free trial and no credit card required.
About the Author
Dr. Marta M. Shinn, Ph.D., is a licensed clinical psychologist, licensed educational psychologist, and nationally certified school psychologist. She is the founder of the diagnostic assessment practice Variations Psychology, the Training Director at the Child Guidance Center in Orange County, California, and an Assistant Clinical Professor of Pediatrics at the UC Irvine School of Medicine. She is an APA-published researcher and has trained graduate students and licensed clinicians for more than 20 years, and she has been a valued member of the BastionGPT advisory board since March 2025.
Learn more about Dr. Shinn and her practice at variationspsychology.com/dr-shinn.
References
- Abdulnour, R.E., Gin, B., & Boscardin, C.K. (2025). Educational strategies for clinical supervision of artificial intelligence use. New England Journal of Medicine, 393(8), 786-797. https://doi.org/10.1056/NEJMra2503232
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
- Budzyń, K., et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: A multicentre, observational study. The Lancet Gastroenterology & Hepatology. https://doi.org/10.1016/S2468-1253(25)00133-5
- Hass, M. R., & Leung, B. P. (2021). When you can’t R.I.O.T., R.I.O.: Tele-assessment for school psychologists. Contemporary School Psychology, 25(1), 33–39. https://doi.org/10.1007/s40688-020-00326-5
- Kittur, A., et al. (2025). The AI deskilling paradox. Communications of the ACM.
- Leung, B. (1993). Assessment is a R.I.O.T.! Communiqué, 22(3), 1–6.
- Mulligan, C. A., & Ayoub, J. L. (2023). Remote assessment: Origins, benefits, and concerns. Journal of Intelligence, 11(6), Article 114. https://doi.org/10.3390/jintelligence11060114
A Note on Compliance
BastionGPT is designed to support HIPAA-compliant workflows for healthcare professionals, including clinical training environments. All data is encrypted in transit and at rest, a BAA is included with every plan, and no data is shared with OpenAI or other third-party providers for training. AI-generated output should always be reviewed by a qualified professional. BastionGPT assists clinical work but does not replace professional judgment.
Try BastionGPT Free for 7 Days
BastionGPT is built for healthcare professionals who want the power of advanced AI with the compliance their work demands. Every plan includes HIPAA compliance with a BAA, unlimited AI scribe transcription, and a saved prompts library you can share with your team.
- Professional: $20/user/month. Core AI assistant, unlimited transcription, document reference up to 10,000 words.
- Professional Plus: $45/user/month. Document reference up to 500 pages, image and chart analysis, multi-document reference, cross-device sync.
- Enterprise: Custom pricing for organizations with 100+ users. SSO, custom branding, dedicated account management.
Start your free trial today. No credit card required. 45-day satisfaction guarantee. Start your free trial.
Questions? Email [email protected] or call (214) 619-8696.