
Yes — because you keep your voice, your structure, your clinical reasoning, and your final sign-off. Soma takes the mechanical assembly off your plate, not the thinking. You read every section, you adjust what you want, and you sign it. If it doesn't sound like you, it doesn't go out the door.

I want to answer this one honestly, because it's the question I hear more than any other — and it's a fair worry.

![A professional's hands typing on a laptop keyboard in warm light](/images/article-body/pexels-gustavo-fring-6699309.jpg)

*Your voice, your structure, your final sign-off stay yours.*

## Will an AI-drafted report actually sound like me?

It will, and that's the whole point of building Soma the way we did. The psychologists I talk to don't want a generic template wearing their name. They want the report to read like *theirs* — same structure, same phrasing, same clinical judgment they'd bring on a good day with enough time.

So that's what we anchored everything to. Your report keeps your voice and your way of organizing things. The software's job is the slow, mechanical part — pulling the pieces together into a first draft you then shape. The thinking stays with you, every time.

One clinician told me their old notes were "knowledge for myself, not really what a note is supposed to be." That stuck with me. The goal was never to replace how you write. It was to get you to a finished version of *your* report faster, without losing the part that makes it yours.

![Hands typing on a laptop with handwritten notes beside it on a desk](/images/article-body/pexels-rdne-8124192.jpg)

*The software handles the assembly; you keep every clinical decision.*

## What does the software write, and what do I still write?

The software handles the assembly; you handle the judgment — and you have the final say on every word. Think of the draft as a strong starting point that already follows your structure, rather than a blank page you have to fill from nothing.

You're still the clinician. You decide what's clinically meaningful, what to emphasize, what to soften, what to cut. You review each section and edit anything that isn't right. Nothing leaves your hands until you've read it and you're satisfied it reads like your work.

For a lot of the people I speak with, a single assessment report is a 10-to-15-hour job — and the dread usually isn't the testing. It's the blank page afterward. The slow assembly of everything into prose. That's the part we wanted to take off your plate, so the time you spend is spent on the clinical thinking, not the typing.

## Won't it read like generic AI?

That worry is the most common one I hear, and honestly, it's earned. We've all seen the flat, padded, say-nothing AI writing — and clinicians can smell it from a mile away. If a report read like that, it would be useless to you, and worse than useless to a reader who's trusting your name on the page.

So that's exactly what we built against. Soma isn't trying to write *a* report — it's trying to help you finish *your* report. Because it follows your structure and your phrasing, the draft doesn't start from some anonymous house style. It starts from yours.

And the safeguard is simple: you read it. The number-one fear I hear from clinicians about AI writing is that it'll come out generic and not sound like them. The review-and-sign step is there precisely so that fear never becomes reality. If a sentence sounds off, you change it before it's ever final. You're the filter, and nothing skips you.

I'd rather be straight with you than oversell it. The promise isn't "magic prose with one click." The promise is: a real first draft, in your shape, that you finish — far faster than starting cold. (More on [whether faster reports still hold up](/articles/faster-reports-but-will-they-hold-up), and on [the report-writing problem this all comes from](/articles/report-writing-problem-private-practice).)

## Do I keep control of the clinical content?

Completely. Every clinical decision in the report is yours, and you sign every section. The software doesn't decide what's significant, doesn't form impressions, and doesn't get the last word — you do.

That matters to me for reasons beyond comfort. Soma is software that helps you draft your reports and notes. It is not a clinician, it doesn't practice, and it doesn't give care or advice. The judgment, the interpretation, the responsibility for what the report says — that all stays with you, where it belongs. We built the workflow so you're always the one reviewing and approving, never rubber-stamping something you didn't shape.

In practical terms: you look at each section, you keep what's right, you fix what isn't, and you decide when it's done. If something doesn't reflect your read of the situation, you change it. The report is yours start to finish — Soma just spares you the hours of mechanical writing in between.

There's good evidence the in-between is where the time goes. One [survey of practising psychologists](https://pubmed.ncbi.nlm.nih.gov/31728882/) found a median of around five hours a week on non-billable clinical work like documentation — and just over half called that admin load a real source of work stress. That's the weight we're trying to lift, without ever touching your clinical voice.

![An open notebook filled with handwriting beside a fountain pen on a wooden desk](/images/article-body/pexels-david-bares-42311-204511.jpg)

*However you write is yours to keep — that difference is the whole value.*

## What if my reporting style is unusual?

Then it stays unusual — in the best way. There's no single "correct" way to write a report, and we never tried to force one. Some clinicians write tight and clipped; some write rich and narrative; some have section orders and headings they've refined over years. All of that is yours to keep.

Soma is meant to follow *your* style, not flatten it into an average. If the way you write is a little different from the next clinician, that difference is the whole value — it's what makes the report read like you wrote it, because you did. Our job is to fit your way of working, not ask you to fit ours.

And if you ever look at a draft and think *that's not how I'd say it* — good. That's the moment you edit, and that's the moment it becomes fully yours. You can [see how the reports side works](/product) whenever you're ready, no pressure either way.

## So — will it sound like you?

It will, because you stay in the chair. You keep your voice, your structure, your reasoning, and the final signature on every section. Soma takes the mechanical assembly, not the thinking — and nothing goes out until you've read it and it reads like yours.

I won't pretend the worry isn't real. Generic AI writing exists, and clinicians are right to be allergic to it. We built Soma to be the opposite of that: a tool that gets you to a finished version of your own report sooner, without ever asking you to hand over the part that makes it yours. You can see how Soma [drafts a report you review and sign](/l/clinical-report-generator).

Thank you for reading, and for the care you put into your work — it means a great deal to us. If you want to talk it through, I'd genuinely love to hear about your practice.

— Ian


Will an AI-Drafted Report Still Sound Like Me?



![A metal padlock latched on a weathered wooden door](/images/article-body/pexels-aleksandar-djordjevic-109766345-9933020.jpg)

*Protecting client data takes more care than simply removing a name.*

Protecting client and user privacy is paramount when processing clinical text or speech data with AI. **Anonymization and de-identification techniques** are key tools to prevent sensitive personal identifiers from being exposed. Basic steps include removing or masking names, dates, addresses, and other protected identifiers from text ([nature.com](https://www.nature.com/articles/s41598-024-81170-y?error=cookies_not_supported&code=a562f8ed-85c2-4f23-8612-a1c0ec5e9609#:~:text=limited%20participation%20in%20this%20research,In%20contrast%2C%20automated%20de)). For example, clinical guidelines like HIPAA enumerate 18 types of identifiers that should be redacted or pseudo-anonymized before sharing records ([nature.com](https://www.nature.com/articles/s41598-024-81170-y?error=cookies_not_supported&code=a562f8ed-85c2-4f23-8612-a1c0ec5e9609#:~:text=limited%20participation%20in%20this%20research,In%20contrast%2C%20automated%20de)). Automated natural language processing (NLP) de-identification systems can assist in detecting and removing protected identifiers, though they are not yet perfect and may miss certain ones. Speech data can likewise be de-identified by removing metadata and using **speaker anonymization** – altering or filtering voice characteristics so that individual identity cannot be recognized. Recent research shows that voice anonymization (using signal processing or deep learning to change vocal attributes) can conceal personal biometric identity in speech while preserving the linguistic content needed for clinical analysis ([nature.com](https://www.nature.com/articles/s43856-024-00609-5?error=cookies_not_supported&code=3edd2e36-4865-401a-95b8-41bda04efb5c#:~:text=Integration%20of%20speech%20into%20healthcare,has%20not%20been%20extensively%20examined)). This means the speech captured in a session, for instance, could be processed to obfuscate who is speaking, yet still allow AI models to detect clinical insights from the speech **without compromising client identity** ([nature.com](https://www.nature.com/articles/s43856-024-00609-5?error=cookies_not_supported&code=3edd2e36-4865-401a-95b8-41bda04efb5c#:~:text=his%20speech%20data%20due%20to,characteristics%2C%20preventing%20potential%20eavesdroppers%20from)).

![Abstract rendering of translucent blue and violet data blocks](/images/article-body/pexels-googledeepmind-17485706.jpg)

*Layered techniques — masking, distortion, and differential privacy — defend against re-identification.*

Despite these techniques, researchers caution that **traditional anonymization is not foolproof in the era of AI**. Simply stripping obvious identifiers (names, etc.) does not guarantee privacy – machine learning models can sometimes **re-identify individuals by cross-referencing anonymized data with other datasets** ([iapp.org](https://iapp.org/news/a/the-myth-of-anonymization-why-ai-needs-a-new-privacy-paradigm#:~:text=This%20perspective%20was%20underscored%20by,is%20someone%20could%20reidentify%20individuals)). In fact, an analysis of de-identified clinical text found that privacy attacks like _membership inference_ could still determine if a particular person’s records were used to train an AI model ([nature.com](https://www.nature.com/articles/s41598-024-81170-y?error=cookies_not_supported&code=a562f8ed-85c2-4f23-8612-a1c0ec5e9609#:~:text=Researchers%20have%20mostly%20focused%20on,identified%20clinical%20notes)). This implies that even if personal names are removed, unique patterns in one’s clinical narrative might be recognized by a model or linked with external information to re-identify someone. To counter these risks, the latest best practices advocate **privacy-enhancing technologies** such as _differential privacy_ and _synthetic data generation_. Differential privacy techniques inject statistical noise or use to ensure no AI model reveals information about any single individual, providing mathematically provable privacy ([jis-eurasipjournals.springeropen.com](https://jis-eurasipjournals.springeropen.com/articles/10.1186/s13635-025-00203-9#:~:text=anonymization%2C%20and%20monitor%20for%20possible,6)). Likewise, researchers are exploring generation of **synthetic clinical data** – AI-generated records that mirror real data statistically but do not correspond to actual individuals – as an avenue to enable model training without using raw personal records ([nature.com](https://www.nature.com/articles/s41598-024-81170-y?error=cookies_not_supported&code=a562f8ed-85c2-4f23-8612-a1c0ec5e9609#:~:text=One%20potential%20solution%20to%20this,3.5)). In summary, **a layered approach to anonymization** is recommended: remove identifiers, distort or mask biometric signals (like voice), and incorporate advanced privacy techniques to defend against re-identification. By doing so, software can help therapists, counsellors, and psychologists harness personal data for documentation insights while rigorously safeguarding individual privacy, with the clinician retaining judgment over every record. I've written more on [the privacy and compliance gaps in using ChatGPT for therapy notes](/articles/2025-10-06_ChatGPT_and_Therapy) and on [what clinicians tell me they worry about with AI](/articles/what-ive-learned-talking-to-psychologists-about-ai). If you want notes drafted without pasting anything into ChatGPT, here's [a clinician-first alternative](/l/chatgpt-for-therapists).


Trust & Privacy

Will an AI-Drafted Report Still Sound Like Me?

AI and Client Data Privacy: What Anonymization Actually Protects