AI summaries of a new drug's trial: 0 of 198 US physicians and other prescribers would rely on one with no sources, 55.1% when every number links to the paper and 72.7% once a named physician has checked it
Over the rep but not the label: 55.1% of US physicians and other prescribers on the Brox panel would rely on an AI summary of a new medicine's pivotal trial when every number links to the published paper, none would rely on one with no sources, and 66.2% would ignore the sales rep first against 24.2% the AI summary and 0% the FDA label (n=198, November 2026).
- Who was asked
- HCP, 198 clones, US
- When
- November 2026 (fielded 13 November 2026)
- How
- Predicted responses from Brox clones, each built from a long-form interview with a real person.
- What changed from the obvious question
- The obvious question asks doctors whether they trust AI. We asked the same 198 prescribers which of five sources they would trust most, and which they would ignore first, for a new medicine's trial results, then showed them three versions of one AI summary that differ only in what it shows: no sources, every number linked to the paper, and linked plus checked by a named independent physician. Comparing each clone's own ratings across the three shows what each safeguard adds.
Section 01: What the panel said
- #
55.1%
55.1% (109 of 198) would rely on an AI-generated one-page summary of a newly approved medicine's pivotal trial, a 5 or more of 7, when every number links to the exact line of the published paper it comes from, against 0 of 198 when it shows no sources and 72.7% (144 of 198) when a named independent physician has also checked it against the paper.
base 198
- #
66.2%
66.2% (131 of 198) would ignore the manufacturer's sales representative first for what a new medicine's pivotal trial showed, against 24.2% (48) an AI-generated summary of the paper, 9.1% (18) the full paper, 1 a colleague or society update and 0 the FDA label.
base 198 · contrast
- #
0
0 of 198 name an AI-generated summary as the source they would trust most to say what a new medicine's pivotal trial showed; 185 name the full published paper, 9 a colleague or specialist society update, 3 the FDA label and 1 the sales representative.
base 198 · contrast
- #
0
0 of 198 would rely on any version of the AI summary completely, a 7 of 7; with every number linked and a physician check, the most common rating is a 5 (110 of 198).
base 198 · caveat
55.1% (109 of 198) would rely on an AI-generated one-page summary of a newly approved medicine's pivotal trial, a 5 or more of 7, when every number links to the exact line of the published paper it comes from, against 0 of 198 when it shows no sources and 72.7% (144 of 198) when a named independent physician has also checked it against the paper.
02What does it take for a physician to rely on an AI summary of a new drug's trial?
A link from every number back to the paper. The same 198 prescribers rated three versions of one AI-generated, one-page summary of a newly approved medicine's pivotal trial, identical except for what it shows. With no sources, none of the 198 rate their reliance at 5 or more out of 7, and 191 say "not at all". When every number links to the exact line of the published paper it comes from, 109 of 198 (55.1%) rate it 5 or more. When a named independent physician has also checked it against the paper, that rises to 144 of 198 (72.7%).
The links do most of the work. The physician check lifts 85 of the 198 clones by at least one point, and 35 of the 89 who stayed below 5 with links alone reach 5 or more once the check is added. Nobody moves the other way: 197 of the 198 never rate a version with more safeguards lower than one with fewer.
03Would physicians take an AI summary over the label or a sales rep?
Over the rep, yes. Over the label, no. Asked which one source they would be most likely to ignore for what a new medicine's pivotal trial showed, 131 of 198 (66.2%) name the manufacturer's sales representative and 48 (24.2%) an AI-generated summary of the paper. Not one names the FDA label, and 1 names a colleague or specialist society update.
As a first choice, the AI summary does not register. None of the 198 name it as the source they would trust most; 185 name the full published paper, 9 a colleague or society update, 3 the label and 1 the rep. Even the 48 who would ignore an AI summary first are not closed to it: 12 of them rate the linked version 5 or more, and 24 the version a physician has checked.
04Which physicians are most ready for an AI trial summary?
The ones who admit they do not read the paper. 18 of the 198 would ignore the full published paper first, and all 18 give the same reason: no time to read a trial end to end alongside a busy clinic, several of them seeing 30 to 50 patients a day. 17 of those 18 would rely on the linked AI summary at 5 or more, and all 18 on the physician-checked version.
Specialty changes the level, not the direction. With every number linked, 74 of 126 dermatologists (58.7%) rate the summary 5 or more, against 35 of the 72 other prescribers (48.6%). With the physician check added, it is 96 of 126 (76.2%) against 48 of 72 (66.7%).
05What would a medical affairs or brand team do with this?
- Our reading, not the panel's answer: an AI summary of trial data earns a hearing only if it is auditable. Without sources it reaches nobody (0 of 198 at 5 or more); with every number linked to the paper it reaches 55.1% (n=198).
- Our reading: a named physician's check is worth adding but is the second step, not the first. It lifts reliance from 55.1% to 72.7% (n=198) on top of the links.
- Our reading: for trial results, the rep visit is the channel most likely to be ignored, by 66.2% (n=198). A sourced, physician-checked summary is a stronger carrier for the same data than a rep's account of it.
06What do people also ask?
- What share of physicians would rely on an AI summary of clinical trial data? 55.1% (n=198) of US prescribers on the Brox panel would rely on one at 5 or more out of 7 when every number links to the published paper, and none when it shows no sources.
- How many physicians would trust an AI trial summary checked by a physician? 72.7% (144 of 198) would rely on it at 5 or more out of 7 when it is linked to the paper and a named independent physician has checked it, but 0 of 198 would rely on it completely.
- What share of doctors ignore pharma reps for new drug trial results? 66.2% (n=198) would ignore the manufacturer's sales representative first, against 24.2% for an AI-generated summary and none for the FDA label.
Section 07: What they said, and the interview line it rests on
“I would say a 5. I actually already use Open Evidence sometimes in practice when I want to check dosages or look into alternative treatment modalities if a patient has failed other therapies. Having an AI summary that links every single number directly to the exact line in the published paper is really appealing and gives it a lot of credibility. That said, I definitely wouldn't rely on it completely. Side effect profiles, trial design, and safety data are huge for me when considering whether to start a patient on a brand-new therapy, especially since I have to counsel patients on those risks. Being in academia and participating in monthly journal clubs, I'd still want to click through and verify the primary literature myself before deciding to prescribe it.”
Predicted from this line in the interview
“I'll use-- sometimes use Open Evidence if I'm not sure about dosages or where to go next if a patient has failed multiple therapies.”
“1 - not at all. I wouldn't rely on it to make a prescribing decision. I have a background working on clinical trials during residency, so I'm very particular about seeing the actual data for myself. Whenever a new drug comes out, I pull the published paper, read the whole study, and go through the package insert to see exactly how they designed the trial, what the comparator was, and what adverse events occurred. An AI summary that doesn't even cite where its numbers come from is essentially useless for making clinical decisions. I've already seen AI tools completely miss the context and pull irrelevant info, so unless I can see the primary trial data and verify where those numbers are coming from, I wouldn't trust it at all.”
Predicted from this line in the interview
“I would much rather see the data myself than read like up to date or something like that. About particularly about a new drug. I think it's much more useful to know how they tested it, what they compared it against. How many, you know, serious adverse events did they see in different groups. I'd really much rather see that.”
“I'd definitely say the manufacturer's sales representative. Reps are allowed to stop by our clinic and bring materials, but they are never my source of truth, and that's just not where I get my clinical information. By the time a rep comes in to talk about a newly approved drug or its trial data, I've almost always already heard about it and reviewed it through peer forums, colleagues, or the actual data itself. I'm very data-driven and look at published evidence or trusted clinician channels rather than marketing pitches.”
Predicted from this line in the interview
“if reps will come to our clinic, they're allowed to come to our clinic and they will bring us information, but they're never the source of truth. that's not where I get my information. I usually have already heard about it before they come. So, very rarely.”
Section 08: How much should you trust this?
What we would defend
- The safeguard ladder within the same 198 clones: reliance at 5 or more of 7 rises from 0% with no sources to 55.1% with every number linked to the paper and 72.7% with a named physician check, and 197 of the 198 never rate a version with more safeguards lower than one with fewer.
- That the direction holds in both specialty groups: with links, 74 of 126 dermatologists (58.7%) and 35 of 72 other prescribers (48.6%) rate the summary 5 or more.
- The order of what gets ignored first for trial results: the sales representative (66.2%, n=198) ahead of an AI summary (24.2%), with the FDA label ignored by none.
What we would not
- Any of these as a measured adoption rate for AI tools: these are 198 clones answering written scenarios, and every rating on the three AI summary versions was inferred from the interviews rather than stated in them, as answers to a hypothetical tool must be; 157 of the 198 answers to the linked version drew on the person's own interview lines about AI tools.
- The 185 of 198 who trust the full paper most as a measure of how often physicians read trial papers: it is a stated ideal, and the 18 who would ignore the paper all say they lack the time to read one.
- The ignore-first ranking as a precise share: none of the 198 answers to that question was stated outright in the interviews, so it reads as direction.
- The separate effect of a check that is named, independent or done by a physician: the third version combined all three, on top of the links.
- Any split by age, gender or region: those demographics are missing for these clones, and specialty comes from their answer to a practice-area question in an earlier study.
The main limitation
These are 198 clones of US prescribers, 126 of them dermatologists and 192 of them physicians, answering written versions of an unnamed AI summary for an unnamed medicine; stated reliance in a scenario is not measured use.
Section 09: How was this run?
| Segment | Clones |
|---|---|
| Dermatology | 126 |
| Another medical specialty | 46 |
| Primary care | 17 |
| Nurse practitioner, physician assistant or pharmacist | 6 |
| A surgical specialty | 3 |
Question wording as fielded
- Are you currently authorized to prescribe prescription medicines in the United States? (added by Ask Brox as a screener; 198 of 198 qualified)
- A new medicine in your area of practice has just been approved by the FDA. Which ONE of these would you trust MOST to tell you accurately what its pivotal trial showed?
- A new medicine in your area of practice has just been approved by the FDA. Which ONE of these sources on what its pivotal trial showed would you be MOST likely to ignore?
- An AI tool gives you a one-page summary of the pivotal trial results for a medicine the FDA approved this month in your area of practice. The summary does not show where any of its numbers come from. On a scale of 1 to 7, where 1 is not at all and 7 is completely, how far would you rely on this summary when deciding whether to prescribe the medicine?
- An AI tool gives you a one-page summary of the pivotal trial results for a medicine the FDA approved this month in your area of practice. The summary links every number to the exact line of the published trial paper it comes from. On a scale of 1 to 7, where 1 is not at all and 7 is completely, how far would you rely on this summary when deciding whether to prescribe the medicine?
- An AI tool gives you a one-page summary of the pivotal trial results for a medicine the FDA approved this month in your area of practice. The summary links every number to the exact line of the published trial paper it comes from, and an independent physician named on the summary has checked it against the paper. On a scale of 1 to 7, where 1 is not at all and 7 is completely, how far would you rely on this summary when deciding whether to prescribe the medicine?
What was not asked, and why
No medicine, company, AI product or disease was named. The study did not test a summary from the manufacturer, a physician-checked summary without links, or a summary of the label rather than the paper, and it did not ask how often these prescribers use AI tools today. Specialty was not asked in this study: it comes from the same 198 clones' answer to a practice-area question on 20 October 2026. Ask Brox added one check that each clone is authorized to prescribe in the US, which all 198 passed.
What changed from the obvious question
The obvious question asks doctors whether they trust AI. We asked the same 198 prescribers which of five sources they would trust most, and which they would ignore first, for a new medicine's trial results, then showed them three versions of one AI summary that differ only in what it shows: no sources, every number linked to the paper, and linked plus checked by a named independent physician. Comparing each clone's own ratings across the three shows what each safeguard adds.
Fielded through Ask Brox on 13 November 2026.
More on How clinicians use AI in practice · Pharma & HCP · HCP (US) panel
Ask this of your own customers: humans@brox.ai
Related insights
All in this topic-
Only 2 of 173 US physicians discount pharmaceutical promotion because AI wrote it, while 63% discount content no expert has reviewed
2 Almost nobody objects to AI writing the content. They object to nobody checking it. (base 173)
Read -
Clinicians will recommend a health app alongside a medicine only when it costs the patient nothing: 96.3% would if it is covered, none at $30 a month
96.3% 96.3% of US healthcare professionals are likely to recommend an FDA-cleared app alongside a long-term medicine when insurance covers it and the patient pays $0; when it is not covered and the patient pays $30 a month none are likely to, and at $100 all 81 are very unlikely to (base 81)
Read