Why a General-Purpose Chatbot Is the Wrong Place to Finish a Radiology Report

A radiologist trying RadPro told me how he finishes reports now. He dictates into his usual PowerScribe templates, pastes the text into a chatbot, asks it to clean the report up and give an opinion, and pastes the answer back. He also said RadPro felt slow next to that. Fair point, and I will come back to speed. But I would not finish a report that way, and here is why.

Chatbots make things up, and they sound sure when they do

Every large language model can hallucinate: write a sentence that sounds right and is not supported by what you gave it. In a careful study of clinical note summaries, about 1.5% of sentences contained a hallucination and 3.5% left something out, and the serious errors only came down after the authors refined their prompts and checked the output step by step. A chat window gives you none of that engineering: one prompt you typed between cases, and no check on the answer.

In a report, a hallucination rarely looks like nonsense. It is a merged pair of findings, a rounded measurement, a missing “no”, or “possible” turned into something firmer. When researchers compared model-written impressions with the radiologists' own, the model's scored lower on coherence, comprehensiveness and factual consistency. The result reads well, which is exactly why it gets signed.

The opinion is no safer. Ask for a differential and the chatbot answers from memory, often with references attached. In one study of chatbot-written medical papers, 47% of the references were fabricated and another 46% were real but cited wrongly. Newer models do better, but an answer you cannot trace is an answer you cannot check between cases.

The medical-legal side

Your signature covers every word in the report, whoever drafted it. “The software changed it” is not a defense if a finding went missing or a side flipped between your dictation and the signed text. It is a worse position still if the software was a consumer chatbot your hospital's policy told you not to use.

Then there is the patient information. A report lifted out of PowerScribe is rarely just findings: the exam line, an indication with the referrer's name, “compared with the CT of March 14”, sometimes an accession number. A company that receives patient information on your behalf is a business associate and needs a business associate agreement. The consumer and individual plans of the popular chatbots are not covered by one; their makers keep those agreements for enterprise and developer products that a hospital has to buy and set up. Pasting identifiers into one can be a reportable breach of unsecured health information, your conversation sits on the vendor's servers under its retention terms, and nothing in the chat box warns you about an MRN before you press enter. This is general information, not legal advice, but your privacy office will tell you the same.

If you practice in Canada

Canada has no business associate agreement, but the rules land in the same place. In Ontario, a radiologist reading for a hospital usually handles patient information as an agent of the health information custodian under PHIPA, and may use or disclose it only as the custodian allows. A personal chatbot account is not something your hospital has allowed. Under PIPEDA, an organization that sends personal information to another company for processing, including one outside Canada, stays accountable for it and must use a contract to give it comparable protection and tell people it may be processed abroad. A click-through consumer account between you and a chatbot company does neither.

Some provinces go further. Nova Scotia requires hospitals and other public bodies, and the companies that serve them, to keep personal information in Canada, and Québec requires a privacy impact assessment before personal information leaves the province. A consumer chatbot account comes with no contract, no assessment and no say over where your conversation is kept. These rules apply to every tool your hospital approves, this one included, so check them with your privacy office before you run real cases.

The manual loop costs more than it looks

To be fair to the chat window: its answer starts in a second. RadPro runs several steps (structure, impression, quality check), and on a large case the whole report takes longer than a chat reply. But count the full loop, on every case. Select the text, switch windows, paste, type or paste your instructions again, wait, read, copy the answer, switch back, paste, strip the bold headings and bullets PowerScribe does not want, then proofread everything against your dictation because anything could have changed. Nothing else moves while you do it.

Most of that loop is gone in RadPro. RadPro Hub, the Windows helper on the download page, moves the report from PowerScribe into RadPro with one key and pastes the finished report back with another. The impression streams as it is written, so you start reading within a couple of seconds, and the next case can run in another tab while you read this one. If your template already holds clean findings and you only want the impression, put “impression only” on the first line and the other steps are skipped. Fast Mode trades a little depth for speed on routine studies. And a single-finding study needs none of this: RadPro earns its time on complex cases, not on a normal chest film.

RadPro takes the compliant route

RadPro was built for clinical text from the start. Identifiers are detected and highlighted before any text reaches an AI model, so you see what was found and confirm it. Report steps run on AI and hosting providers that have signed business associate agreements, and identifiers are replaced with placeholders before any step that leaves them, such as the literature search. Report text is encrypted in transit and at rest, every action is written to an audit log you can export, multi-factor authentication is required on paid plans, and groups can sign a business associate agreement with us directly.

RadPro's servers and database are in Toronto, so your reports, your templates and the audit log are stored in Canada.

The pipeline is built against the errors above, too. The structuring step is told not to add or remove findings and to keep your normal statements. A quality check reads the finished report for wrong sides, contradictions and findings missing from the impression, and gives you a short list rather than a rewrite. Advanced DDX searches the current literature during the run and cites what it found. And with My Templates on Premium, your own templates stay in your own wording: Generate Template adapts yours line for line, and the impression you keep in a template becomes your default impression when the study is normal. None of this replaces your read; it makes the read faster to check.