It is remarkable that in the field of AI and psychology, so many professionals are suddenly very concerned about data security and protecting PHI. This is strange, because there is little that is technologically unique about the data-breach threat when using AI. The same issues that apply to uploading to ChatGPT, Gemini, Claude or a customised platform apply equally to your OneDrive or Google Drive backups, to everything in your Gmail account, to everything in your WhatsApp, and certainly to everything in your clinic management software. That is, unless you forget to take the one simple step — ticking the box that says you do not want your chats used to train future models. My claim to expertise is not rigorous training in cybersecurity or privacy law. But as a software developer — some of whose software uses AI models — I have come to understand how these technologies actually work. Perhaps because I am not an IT professional, I can explain things in layman's terms. I believe much of the anxiety stems from the unknown: to non-IT-literate psychological professionals, this is a black box, or even a Pandora's box. Psychoeducation requires mapping the beast — understanding it. A further driver of the anxiety, I think, was timing. Just as AI start-ups were beginning to hit the market, the newly upgraded, punishment-heavy amendments to the Privacy Protection Law arrived — and the prospect of fines running to hundreds of thousands, even millions, of shekels dampened many appetites. A final factor is the lack of clear guidelines and regulation in Israel. Contrast this with the USA, where HIPAA — the Health Insurance Portability and Accountability Act — sets out standards for keeping sensitive PHI safe, enforced by a dedicated regulator. In Israel there is no equivalent statutory technical standard. Instead we have a highly lucrative industry — not unlike the unpopular, expensive Kashrut certification enterprises — trusted to assess software and stamp it "kosher," declaring that it is indeed up to standard. In this series of psychoeducation articles I want to clarify two things: (1) how software works (dumbed down), and (2) how an individual psychologist needs to behave so that, should there ever be a data breach, they can confidently say they did their due diligence. This post focuses on how software works. The three pieces of any software The frontend. All software has a frontend, which overlaps with the user interface. This is code that runs inside your browser. When you load a webpage, there are several things your browser can do on its own — for example, when a field auto-fills, that uses code embedded in the page running in your browser. The backend. Most software also has a backend, otherwise known as a server. This is another computer located somewhere — sometimes privately owned by a tech company, but most often part of the vast server farms owned by giants like Google Cloud Platform and Amazon Web Services. This computer, just like yours, runs code — specifically, the code unique to the software, which the company does not want to share with you. For example, if a program auto-scores your standard-score data, it's unlikely the scoring logic and norming data will ever be sent to your browser — because with a little know-how, or a little AI assistance, anyone can read code that reaches their browser. The database. The final component is a database — also a computer drive — which stores the information you access each time you log in. Since the server is like working memory, it can't hold everything at once. Instead it parks information in a database — the long-term memory store — and calls it up whenever you ask. In practice: say you log in to the Electricity Company's website to pay your bill. You use your browser (the frontend UI) to enter your details and choose to pay. This sends a messenger — a specific piece of code — to the server, which pulls your information out of the database. Applied to your Google Drive: the frontend displays your file names and types. When you click a file, Google's server pulls it from the database and sends it to your browser, where it opens. Although it looks as though the file was on your computer all along, that is not the case. And now, AI And now, AI — otherwise known as an LLM. In your browser you open a frontend where you type prompts and upload files. This frontend may belong to the AI company itself — the standard ChatGPT, Gemini or Claude chat interface. Or it may belong to a private company that built its own interface, such as the dozens of AI-and-psychology start-ups, each offering its own beautifully designed — usually by AI (!) — wrapper. When you click send or upload, the content travels to a backend server. If you're using the ChatGPT/Gemini/Claude interface, it goes straight to their server. If you're using third-party software, there's one extra server in the middle: the data goes from your browser to the company's server, where it's packaged and forwarded to the AI company's server. (Note: a few start-ups have built their own LLMs, run on their own local servers, but most sit on top of the big three providers.) The LLM lives on yet another supercomputer belonging to the AI company. When it finishes processing your upload, it sends the output back to the server, and from there it journeys back to your browser along the same chain it came in on. Where the real vulnerabilities are So where are the real data-privacy vulnerabilities? Two places: in transit and in databases. Let's take them one at a time, and look at the solutions. In transit means while data travels from your computer to a server — along cables that, unfortunately, run under the ocean and can be tapped. Everything you send could, in principle, be intercepted and read by someone else. In databases, there is often an enormous amount of sensitive information sitting in one place. If someone physically got hold of the drive — or a rogue employee gave in to the criminal urge to scrape it — vast amounts of data could fall into the wrong hands. The answer to both is encryption. Whenever data travels between two computers, it is scrambled into incomprehensible code; a second piece of code, the key, is required to unscramble it. That protects data in transit. And whenever data sits in a database, it should be stored encrypted too — so that even if someone scraped the drive, what they got would be meaningless. Ever noticed whether a web address begins with http or https? The s is what matters: https means the traffic between your browser and that server is encrypted. Plain http is not — so the "s" is worth looking for. As for the database side, any HIPAA-standard software ensures that data at rest — parked in a database — is encrypted as well. In practice, this means the only realistic way for someone to read your data is by logging in to your account, or an admin account with access to it. Part of that responsibility rests on you, the psychologist. Use a strong password and protect it. Wherever it's offered, put your sensitive data behind MFA — multi-factor authentication — so that cracking a single password isn't enough to reach your data. The gold standard: client-side storage and zero retention High-security platforms go one step further: they don't use a central database for sensitive data at all. Instead they store everything client-side — the actual PHI stays with you, in your browser or linked to a store you control, such as Google Drive or OneDrive. This makes it effectively impossible for the company to read your data, because two separate systems are needed to reassemble it: your side holds the data (and half the key), and the provider holds the other half. Neither can read it alone. That is genuinely high-security territory. Zero-retention AI is the gold standard for protecting your clients' data, and it works on this same principle. It means your data isn't parked in Google's or OpenAI's databases for the long term: the moment the server finishes processing it, it's wiped. It's the kind of practice HIPAA-conscious vendors adopt, and the rationale is clear. You shouldn't have your client's data sitting anywhere without good reason, even encrypted. It's reasonable to send it to a third-party server to be processed. It's reasonable to back up or organise client records in a secure third-party database. But there's no good reason for your client's data to be sitting idle in some random AI company's database. The bottom line: if you use HIPAA-standard software that encrypts data in transit and at rest — and especially where the PHI never lands in a database at all — your clients' data is very safe. The legal side I'll cover in a future post. Say a breach happens anyway — how do you protect yourself? Which contracts count as due diligence: between the software company and its server providers, between you and the software company, and between you and the client? The kind that let you say, if you're ever sued, "I behaved sensibly, and this is not my fault." That's for next time. But remember: none of this is unique to AI. It applies every single time you put your client's data into any software with an online component. Disclaimer: This is not professional legal advice. It was assembled via human research and written by human hand, then fact-checked, lightly edited and translated using AI. If you identify an error, please reach out so it can be corrected for the benefit of others.