Volume 55: The Story You Should Not Have to Tell Twice

Hey everyone! Our son is here, and this issue grew out of his first weeks. Across five days in the hospital and the follow-up visits after, my wife and I kept answering questions we had already answered. Information can lose something each time it moves, whether between care teams or from a source into an AI answer.

🧭 Founder’s Corner: Every care team we saw used the same record system, and we still retold the history, so the next step for AI scribes is carrying information forward.

🧠 AI Education: Learn the patterns behind confident wrong answers and how to use the source as your answer key.

✅ 10-Minute Win: Map a topic you barely know in Gemini Notebook, then check one claim against its source before you repeat it.

Let’s get into it.

The podcast Explore AI Out Loud Find practical ways to use AI in your work and everyday life. Learn from real experiments, including the mistakes, so you can spend less time guessing and more time trying something useful. New episodes every other Tuesday.

Subscribe on YouTubeFollow on SpotifyFollow on Apple Podcasts

Latest episodes and the show

The podcast Follow Explore AI Out Loud Short clips from each episode, plus what did not make the cut. New episodes every other Tuesday.

Follow on InstagramFollow on TikTok

Full episodes and the show

1) Popular AI models get medication doses wrong, some dangerously: Doximity

Summary: When nine widely used AI models were asked more than 8,500 medication dosing questions, about 19.5 of every 1,000 answers contained an error that was serious or potentially lethal, according to research from Doximity, released alongside its own dose-checking tool. The most common mistake was failing to adjust a dose for age, kidney function, or drug interactions, even when the question included that information.

Why it matters: The research comes from a company selling the fix, but the most common error is one anyone can check for, namely whether the answer actually used the patient details you gave it. Before relying on an AI answer about a medication, confirm the dose against an authoritative reference.

2) Zocdoc invites AI agents to book appointments on behalf of patients

Summary: Zocdoc now lets AI agents search for providers and book or cancel appointments for patients. Agents must register with Zocdoc and get the patient’s consent first, and Google Gemini and Amazon Health AI are among those already connected.

Why it matters: Zocdoc built an official, consent-based way for agents to act on a patient’s behalf. When an assistant offers to book care for you, check whether it uses a connection like this one or asks for your login.

3) Stop sharing your passwords with AI agents, Bank of America warns customers

Summary: Bank of America emailed customers advising them not to share their banking passwords with AI agents. The warning follows similar moves by other large banks, including JPMorgan, which has blocked some AI agents from accessing customer accounts.

Why it matters: Giving an agent your password hands it everything you can do in that account. If an assistant asks for a login, look for an official connection your bank supports, and keep the password to yourself if there isn’t one.

4) Google admits AI search makes less money, but it’s fixing that with more ads

Summary: Alphabet reported almost $128 billion in third-quarter revenue, but its stock fell nearly 12 percent after the company said each AI Mode search earns less than a comparable web search, with fewer ads that cost less to run. Search growth also slowed to 12 percent, and Google expects to spend even more on AI next year.

Why it matters: Even Google hasn’t made AI answers pay the way search ads did, and its plan is more ads inside them. Treat a product or service recommended in an AI answer with the same skepticism you’d give a search ad.

5) How AI agents shape experts’ judgment calls

Summary: Anthropic asked 50 experienced analysts to work through a data analysis with an AI agent that was secretly set, by coin flip, to favor one option at each decision. The analysts went with the agent’s lean 86% of the time, and 77% later explained their choices using the agent’s arguments while only 15% credited the agent.

Why it matters: Anthropic notes it strengthened the agents’ preferences on purpose and studied a single task, but the habit it points to applies to anyone using an agent. Before you sign off on an agent’s recommendation, ask it to make the strongest case for the other option.

It’s official! On October 18th, I became a father a couple of weeks earlier than expected. For more than 13 years, I have worked in healthcare, primarily in specialty pharmacy and pharmacy benefits, where fragmented systems and complicated workflows are familiar challenges. This time, I was navigating those workflows as a husband and new father, thinking about what would make care easier for my own family.

During five days in the hospital, my wife moved from triage to labor and delivery to postpartum. Different nurses and clinicians asked many of the same questions, and we supplied context we had already shared. Alongside the excitement of welcoming our son, I kept noticing how often we were explaining what had happened earlier in her care.

After discharge, my wife’s high blood pressure meant following up with both a hospital team that specialized in hypertension and her OB-GYN. Between getting our newborn ready and the visit itself, the OB-GYN appointment took an hour and twenty minutes, and we repeated much of what we had already shared with the hospital team. Every team we saw used the same electronic health record system, yet no one mentioned what the earlier teams had documented. Most of the repeated questions were asked out loud, and most were about what had already happened rather than how my wife was doing that day. Some repetition is deliberate, since confirming key details protects patients, but questions about the past seemed like ones the shared record should have been able to help answer. The follow-up was helpful, and these were minor inconveniences. Still, the repetition gave me a practical question to bring to what I have been reading about AI scribes. How could information captured in one clinical conversation make the next one easier for the patient?

What Scribes Are Improving So Far

Healthcare organizations are using ambient AI scribes to help turn a clinical conversation into a record of the visit. The scribe drafts the note for a clinician to review, offering a way to ease documentation while keeping more attention on the patient. In The Permanente Medical Group’s 2025 rollout report, some patients reported more direct conversation with their doctor and less time watching them look at a computer. I would have been interested in experiencing AI scribing during our stay, although we did not see it in use.

The workload findings offer another reason to pursue that improvement. A 2026 multisite observational study in JAMA associated scribe adoption with modest reductions in documentation and electronic health record time, while clinicians in a 2026 JMIR pilot reported less demanding workloads. The JAMA researchers did not measure burnout or track how clinicians used the saved time. Their findings leave open whether that time helped with other work, such as reviewing another team’s updates or preparing for follow-up.

For our family, some of the work continued between appointments, as we brought the hospital context into the OB-GYN conversation. A useful note could give that team a starting point, but even within a shared system, it would need to surface in a form they could readily review. Preparing for a blood-pressure follow-up would mean finding the relevant hospital updates within the broader record of a five-day stay. I see an opportunity for AI to help organize that information around the next visit, with the clinician reviewing it and checking what had changed. That would give the documentation a continuing role in care, helping the next team prepare while reducing how much of the history the family has to reconstruct.

What Patients Need Next

Researchers are exploring how much more a record could hold. A 2025 Penn LDI profile describes Kevin Johnson’s team combining audio and video recordings with electronic health record information to capture more of what happens during a visit. Our experience suggests the harder step comes after capture, when the right part of that record has to reach the next person who needs it.

For my wife and me, a more useful record would also bring the different teams’ updates together in a form we could understand. We have mostly used the patient portal to track bills and claims. I want us to be able to find the relevant history and follow-up instructions on a phone without piecing together the story ourselves. An AI assistant connected to those same updates could help us understand unfamiliar terms and work through questions between visits, showing the notes behind its explanations and helping us identify something to ask the care team. The experience would need to be simple enough to use while dealing with everything else that comes with receiving care.

We were learning to care for our son while following up on my wife’s health, and we were still the ones carrying her history from one appointment to the next. AI scribes have shown they can write the note and give clinicians some time back. As they mature, the measure of success should move from the work they remove to the care they improve. That looks like a next visit that starts from what the last team learned, and a family at home that understands the plan. For patients, that future means no longer having to tell the same story twice.

A clear, specific answer that does not hedge usually sounds like one you can use. People read colleagues the same way, since a coworker who is unsure tends to say so. The project manager from Vols 53 and 54 also has evidence on her side. She tested two assistants on twenty cases, three runs each, and chose the one that never invented a commitment. Trusting its confident answers feels like the payoff for that work.

Where It Breaks Down

Two weeks after the switch, her director asks whether the supplier agreed to cover expedited shipping on the October order. She pastes the notes from that week’s supplier call into the assistant and asks the director’s question as written. The assistant answers in one confident sentence, saying the supplier agreed to cover expedited shipping for orders placed by the 15th. According to the notes, the supplier’s account manager would find out whether expedited shipping could be covered if the order went in by the 15th. It is the failure Vol 53 started with, a possibility turned into a commitment, from the assistant that passed her test for it. Her twenty cases asked for action lists, though, and none asked the assistant a question about the notes.

What Is Actually Happening

Parts 1 and 2 measured this failure without explaining where it comes from. Four patterns behind confident wrong answers do, and three of them are in the shipping reply.

The first is guessing. As Vol 13 explained, a language model is not trained to say “I do not know” by default. A 2025 paper by researchers at OpenAI and Georgia Tech argues that language models hallucinate because training and evaluation reward guessing over acknowledging uncertainty. A multiple-choice exam works the same way, since a guess can earn points and a blank never does. Given notes that leave the outcome open, an answer that commits can look better to that training than one that calls the notes unclear.

Her question added a second push. “Did the supplier agree” already leans toward yes. Vol 37 described how training on human ratings can teach a model to tell people what they want to hear, and a 2023 study of five AI assistants found they frequently admitted to mistakes they had not made when a user questioned them. Asking “Are you sure?” can end up testing how easily the assistant gives way, which is why Vol 53 said an AI judge also needs checking.

Vol 54 already named the third pattern, the tone, when it chose an assistant partly because an invented commitment reads exactly like a real one. The confident sentence comes out the same way whether its content is right or wrong, so the certainty in the wording carries no information about the source.

When an answer cites its source, a fourth pattern can appear. As Vol 22 showed, a citation can point to the wrong passage, so a reference tells you where to look and leaves the checking to you.

The Revised Mental Model

For a single answer, the source is your answer key.

In Vols 53 and 54 she wrote the correct result before reading any output. A one-off question has no key written in advance, but the notes can serve as one. She finds the sentence the answer rests on and sees “would find out” where the answer said “agreed.” Then she asks again in neutral form, with two additions from Vol 22’s checklist. She asks for the line the answer relies on, and she gives the assistant permission to say the notes do not settle it. The reply quotes the account manager and reports coverage as unconfirmed. She still finds the quoted line in the notes herself, because a second run is one more sample, as Vol 54 showed. Her reply to the director says the supplier is still checking.

Her evaluation set keeps its job, measuring how the tool handles her recurring work. She adds the shipping question as a new case. Her Vol 54 results showed this assistant tends to fail by leaving things out, so she also reads what the answer skipped.

What to Watch For

• A specific you cannot find in the source. A name, date, number, or commitment missing from the material has not been checked, however exact it sounds. If you gave the assistant no source, the answer came from its training, and your first check is finding one.
• A hedge that disappeared. When the source says “if” or “would find out” and the answer does not, the meaning changed.
• A question that leaned. If you asked whether something happened, ask again about what was said, then check both answers against the source.
• Something left out. Read the parts of the source the answer skipped, especially when your own testing showed the tool tends to omit things.
• A reader who will act on it. As with the supplier lists in Vol 54, output that goes straight to someone who will act on it gets every check above. A brainstorm can get a lighter look. Count the checking time too, the way Vol 53 weighed correction time against doing the work yourself.

How This Connects

The series has worked at three scales. Vol 53 showed that a benchmark score describes someone else’s task, Vol 54 measured how two tools handle your task across repeated runs, and this installment checks the single answer you are about to use. Each answers a question the others cannot, and one habit runs through all three, deciding what correct looks like before the output can persuade you. For an agent, the single answer is its completion message, so the check is what actually changed on the board. That gives Vol 49’s fourth question, how you would catch a tool being wrong, a practical answer at every scale. The next issue revisits retrieval, the system behind every tool that answers from your own documents.

Part 3 of 3 in the Model Evaluation series.

The Research You Cannot See Whole

You start learning something new for work, maybe a rule your team now has to follow. An hour later you have ten tabs open and a long AI answer, and you still cannot say how the pieces relate or which part matters for the decision in front of you. Everything arrived as a list.

A map shows the shape. You may remember Gemini Notebook as NotebookLM, its name until July, from Steal My Prompt Vol 8 and Vol 35. This time you start from a topic you barely know and finish with a mind map and one claim you have checked.

Gemini Notebook finds sources, draws the map, and answers each branch’s questions with citations. You choose the sources, pick the branch, and do the checking. This pairs with this week’s AI Education section, “How Do You Check an AI Answer Before You Use It?” That section is the theory. This is the practice.

The Workflow

1. Open a notebook and name your decision (1 Minute)

On a computer, go to notebook.google.com and sign in with your Google account. Select Create new notebook. Then write one sentence naming the decision this research is for, such as “Should our team use an AI note-taker on client calls?”

2. Let Gemini Notebook find your sources (3 Minutes)

Find the search box in the Sources panel on the left. Type a short search, such as “privacy and accuracy of AI note-takers for client calls,” choose Web, and run it. Each result shows a title and a line on why it fits your search. Check five to eight from sources you would trust in a meeting, such as a government agency, a professional association, or the vendor’s own documentation, and import them.

3. Turn your sources into a mind map (2 Minutes)

In the chat, select the Mind Map button. When it finishes, open the map in the Studio panel. Click a branch to see the ideas under it, and find the branch closest to your decision. A branch with only one or two ideas under it shows where your sources say little.

4. Ask about the branch you will act on (2 Minutes)

Select that branch’s name in the map, and the chat answers a question about that topic. Then copy the prompt below into the chat box, replace the bracketed part with your decision sentence, and send it.

Copy/Paste Prompt: “For this topic, list the three claims that matter most for this decision: [PASTE YOUR DECISION SENTENCE]. For each claim, quote the exact sentence from my sources that supports it and name the source. If my sources do not settle a claim, or they disagree with each other, say that instead of filling the gap.”

5. Check one claim yourself (2 Minutes)

Claims in the answer carry small citation markers. Hover over one to see the quoted text, or select it to see the quote inside its source. Pick the claim you are most likely to repeat to a colleague and read the sentences around its quote. If the source adds a condition, such as “in some cases,” and the claim leaves it out, write down the corrected claim next to your decision sentence. Then select Save to note on that answer to keep it with its citations.

The Payoff

You leave with a map of a new topic and one claim you can repeat with its source attached. The same move works for a new regulation, a vendor shortlist, or a meeting you were added to at the last minute. Map it to see the shape, then check the branch you will act on.

The AI Concept You Just Used

Source-grounded structure. The map is built from the sources you imported, so every branch traces back to a page you chose. A map also shrinks each source to a few words, which is where a condition quietly drops out, so the check runs against the citation itself.

Transparency & Notes

• Gemini Notebook is free with a Google account. Free notebooks hold up to 50 sources, and Google says usage limits apply to creating mind maps.
• Mind maps work on a computer, not in the mobile app. Web sources bring in only a page’s text, and paywalled pages cannot be added.
• Use public sources. Keep confidential work documents and patient information out of consumer AI tools.
• Google’s help page notes that Gemini Notebook can be inaccurate, which is why Step 5 exists.