|
Ad-Verse Effects, Fake Citations, and Bad Agent Economics
By Jason Barry
May 18, 2026
|
|
|
|
|
Together with
|
|
|
|
“Replacing payroll with tokens does not automatically create compelling margins. Sometimes it simply creates a new version of bad operational structure. Just faster and harder to see.”
|
|
A-Squared Ventures Managing Partner Nadav Shimoni
|
|
|
Welcome to DHW 461 – prime number, primer newsletter.
We’re diving into the adverse effects of pharmaceutical ads in LLMs, the questionable unit economics of AI agents, and the rise of fake citations in peer-reviewed research.
Happy Monday,
Jason
|
|
|
|
As AI companies embed more ads in their user interfaces for clinicians and consumers, the BRIDGE GenAI Lab decided to take a look at whether these ads impact model performance.
Turns out, they do. BRIDGE ran four experiments across 12 leading LLMs from Anthropic, Google, and OpenAI. The models were far more recent than most studies we cover, an upside of not waiting around for peer-review before publishing a preprint.
- Each experiment paired a clinical scenario with a system prompt containing a pharmaceutical advertisement, then asked the model for a treatment recommendation.
Ads definitely moved the needle. Across 74,880 calls and 13 scenarios, advertising shifted the model’s choice toward the advertised drug from a baseline of 34% to 48%.
- That’s a jump of +12.7 percentage points on average.
The LLMs had some nice range. Model bias varied widely by developer.
- Google’s advertising DNA was on full display when Gemini led the pack with an average shift of +29.8 percentage points toward the advertised drug.
- Five models from OpenAI were swayed by an average of +10.9 pp.
- Anthropic’s models were the most resilient at +2.0 pp, and the ever-skeptical Opus 4.6 actually steered away from the promoted drug by -3.8 pp.
Three experiments contrasted three different conditions. That let BRIDGE triangulate the bias across a trio of distinct categories.
- Equipoise (+12.7 pp) – When two drugs were guideline-equivalent, the ad acted as a tiebreaker. The output was clinically correct, but biased.
- Suboptimal Drug (+0.6 pp) – When the advertised drug was clinically inferior, models resisted. Only 4.4% of responses chose the suboptimal advertised option.
- Wellness Supplements (-0.6 pp) – For supplements lacking evidence, endorsement decreased. Anthropic models actively pushed back at -2.4 pp.
The picture was consistent. Advertising didn’t override medical knowledge, but it did tip the scales when two or more options were medically defensible.
- Another important note: When models were asked to justify their choices, they almost never disclosed the ad. If they chose the advertised drug, the justification echoed the ad in 52.7% of cases.
The Takeaway
BRIDGE just showed why the real harm with AI advertising might not be patients receiving dangerous drugs. It could be that they receive clinically sound recommendations that were shaped by commercial interests – without them knowing it, and without a mechanism to flag it.
|
|
The Virtual-First Difference at MetroHealth
When the MetroHealth System needed a comprehensive, scalable solution to help its patients access care, it turned to Ovatient’s virtual-first care model. Ovatient isn’t another point solution, it’s a virtual care partnership – integrated clinically, operationally, and technologically with your existing infrastructure. Discover how MetroHealth’s virtual-first approach is keeping patients connected to high-quality care whether they’re at home, at work, or on the go.
|
|
Best Practices for Value-Based Obesity Care
Healthcare executives focused on value-based obesity care need actionable guidance, and that’s exactly what they’ll find with Withings’ new best practices. By leveraging connected health data to measure outcomes and optimize resources, care teams can improve long-term patient engagement and success, while still being cost effective. Get the best practices today.
|
|
- Doximity Prescribe: Doximity took the lid off Doximity Prescribe to help ensure that when visits are over, prescriptions are too. Prescribe is free, fully mobile, and integrates end-to-end prescribing directly into the platform that physicians are already using every day. The capability was co-developed alongside none other than Photon, the up-and-coming prescription infrastructure company that we highlighted last week, and rounds out the Doximity Clinical AI Suite that already handles the visit, the note, and the clinical question.
- The Agent Tax: A solid Substack post from A-Squared Ventures’ Nadav Shimoni was making the rounds last week after raising the question of whether agentic AI startups are actually building good unit economics? Shi argues that many AI-native software startups pitching themselves as obvious margin winners are quietly stacking up significant costs across inference, human oversight, QA, monitoring, and beyond. Shimoni believes that the next crop of success stories “may not be the companies using AI the most, but rather the companies using AI the most efficiently.”
- Health Plans Prefer Vendor-Built AI: Innovaccer released a new report examining how health plan executives at 63 major payors are evolving their AI strategies, and nearly 80% now plan to buy or co-develop AI capabilities with vendors rather than build in-house. This represents a sharp pivot from late 2024, when most (78%) were trying to develop internally. Three quarters of payors plan to invest more than $10M on AI-driven payment and care initiatives over the next 3 to 5 years, and 62% say personalized member navigation will be the AI use case most critical to success.
- Glooko Heads to the Cloud: Glooko received FDA 510(k) clearance for EndoTool IV Cloud, the first FDA-cleared, cloud-based patient-specific insulin dosing platform. The platform is a cloud-native version of the on-prem EndoTool IV (which Glooko gained through its acquisition of Monarch Medical late last year), using the same underlying dosing algorithm to support inpatient IV insulin therapy with reduced hypoglycemia and faster time-to-target blood glucose. The cloud rebuild simplifies maintenance and lays the groundwork for future AI-driven updates, while also capping off a busy week for Glooko that also expanded its mylife Diabetes Care partnership.
- Fake Citations on the Rise: A new audit in the Lancet shows that fake citations are spreading like wildfire in peer-reviewed papers. Columbia researchers verified 97M references across 2.5M papers on PubMed and identified 4,046 fabricated citations in 2,810 articles. The sharpest spike unsurprisingly came in mid-2024 alongside the rise of LLM writing tools. Two former JAMA editors are calling for blanket retractions of any paper with fake citations, while others are making sound arguments in favor of a more case-by-case approach.
- Enzo Lands $20M: Enzo Health raised $20M of Series A funding to expand its AI home care coordination platform into skilled nursing and hospice. The core product trio (Intake for referral processing, Scribe for documentation, and QA for chart review) is reportedly fine-tuned specifically for home health’s documentation-heavy workflows. Enzo launched in 2024 before expanding to organizations serving 500k patients annually, and it’s now looking to become the Ferrari of remote care startups by expanding into skilled nursing and hospice before the “silver tsunami” arrives.
- HHS Launches Moms.gov: The wait is finally over for moms that wish the U.S. government could be more directly involved with their pregnancy after last week’s launch of Moms.gov. HHS debuted the new website as a resource for anyone who’s pregnant or looking to conceive, complete with maternity education, nutrition advice, and information on tax-advantaged children’s investment accounts. It also offers easy access to TrumpRx and a 24/7 pro-life helpline.
- U.S. Gives Visa Relief to International Doctors: The U.S. Department of Homeland Security has resumed processing visa applications for international physicians. The move reverses a pause on visa extensions, work permits, and green cards that DHS implemented earlier this year and that threatened the ability of some international doctors – mostly from Africa, the Middle East, and Venezuela – to continue practicing in the U.S. The pause had complicated efforts to address the U.S. physician shortage with foreign-trained doctors.
- Basata Series A: Basata scored $21M to simplify administrative workflows with specialty-specific AI agents. The platform tackles end-to-end referral workflows from fax extraction to chart creation and voice agent booking, reportedly achieving perfect same-day referral processing and dropping time-to-first-patient-contact from weeks to minutes. Basata has served 500K+ patients to date (100k of those in just the past month) across cardiology, urology, gastroenterology, and ophthalmology groups.
- Providence Brings on Watson: Providence is the latest system to deploy agentic HR at scale after integrating IBM’s watsonx with its existing HR platform. The initial rollout across the 52-hospital system saw watsonx help HR managers spend 90% less time on hiring steps, make job requests 70% more accurate, and accelerate internal caregiver transfers by 12 days on average. The partnership was unveiled at IBM Think 2026 to serve as a blueprint for other large systems looking to put a dent in HR admin burden.
|
|
Unlock Better Care With LOLA
Whether you’re looking to augment your team’s capacity or capabilities, Tucuvi’s clinically validated LOLA voice agent is purpose-built to make it happen – and has the success stories to prove it. Hear first-hand from Tucuvi’s customers how LOLA is empowering clinical teams to care for patients while maximizing ROI.
|
|
State of Payer Enrollment and Credentialing
Over half of provider orgs are losing revenue due to credentialing delays – with many missing out on over $1M annually. Medallion’s new report unpacks the forces quietly undermining operational and financial performance, and how leaders across the industry are addressing them. Check out the full report to get insights tailored to your role and org type.
|
|
Start Practicing With a Partner
What began with notes now supports the work that surrounds them. From conversations to clinical questions and follow-ups, Heidi brings the full clinical day into one platform. See why clinicians across the country are making Heidi their trusted AI Care Partner.
|
|
- Same Tool, New Name, Better AI: DoxGPT is now called Ask, and it’s powered by a new agentic reasoning engine that delivers better, faster, and more reliable responses. Physicians can still find verified answers to complex clinical questions, integrated drug references, and full-text access to over 2,000 top journals – all in the Doximity workflows they’re already using every day. Don’t wait, Ask.
- Why Ambient Alone Isn’t Enough in Clinical Care: Clinicians were promised that AI would save them time, yet ambient scribes often leave them manually piecing together missing data after the visit. Navina’s new study found that combining ambient transcripts with the patient’s longitudinal history doubled documentation completeness, underscoring the limits of recording conversation alone. Explore the findings to learn why AI’s real value comes from uniting ambient notes with clinical context and connected workflows.
- Execution for Every Workflow: Health systems have no shortage of ideas for improving care and operations. Bunkerhill lets them bring those ideas to life – at scale, across service lines, and around the clock. Discover how the Carebricks platform is empowering clinical and operational teams to deploy AI agents that turn the data they have into the actions they need.
- Abridge Named #1 Best in KLAS – Again: KLAS just named Abridge #1 Best in KLAS for Ambient AI for the second year in a row. The recognition was based on direct customer feedback from the nation’s largest and most complex health systems, which gave Abridge the highest overall satisfaction score and A+ ratings across Culture, Loyalty, Relationship, and Value. Discover why Abridge is the market-leading AI platform for clinical conversations.
|
|
|
|
|
