The ChatGPT Prompt That Finds Errors in Your Hospital Bill (And What It Misses)

A family fought a $195,628 hospital bill down to about $33,000 using a chatbot and a six-page letter. That account has been widely reported and I cannot verify it, so read it as the ceiling rather than the expectation. An 83% reduction is not what happens to most people who try this.

Here is what does happen. In 2024 the Commonwealth Fund found that 45% of insured working-age adults received a bill for care they believed was covered, and fewer than half of them disputed it. Of the people who did negotiate a medical bill, 93% got it reduced or eliminated, according to a LendingTree survey of 2,016 adults in November 2024. The gap between those two numbers is the whole opportunity, and it has nothing to do with AI. AI only lowers the cost of the first step.

That first step is reading the bill. A hospital statement is written in CPT, HCPCS, and revenue codes, which is why most people pay it without looking. Decoding those codes is the thing a language model does well and cheaply. Judging the prices attached to them is the thing it cannot do, and the second half of this piece is about that line.

The prompt, and what to strip out first

Paste this after you have the itemized statement, not the summary you got in the mail.

I am reviewing a hospital bill. Use only the line items I paste
below. Do not estimate what anything should cost and do not
compare prices to any benchmark.

For every line, give me a table with:
1. The billed code exactly as written
2. What that code is, in plain English
3. Quantity and unit price as billed
4. A flag: DUPLICATE, QUANTITY, BUNDLED (a service normally
   included in another line on this bill), or CLEAR

Then, separately:
5. List any two lines that are components of the same panel or
   procedure and should not both appear.
6. Add up the line amounts yourself and compare your total to
   the total printed on the bill.
7. List every code whose description you are not certain about
   and say so. Do not guess.
8. Write the three questions I should ask the billing
   department, citing the line numbers.

Here is the bill:

Two instructions in there are doing most of the work. Item 6 forces the model to add the column itself instead of trusting the printed total, and item 7 gives it permission to be uncertain, which is the difference between a useful flag and a confident invention. The ban on price comparison is there for a reason I will come back to.

Strip your name, date of birth, medical record number, account number, and insurance ID before pasting. Keep the dates of service, the codes, the units, and the amounts. The analysis needs none of the identifiers.

Does pasting a hospital bill into ChatGPT violate HIPAA?

No, and the question comes up so often that it is worth answering directly. HIPAA binds covered entities: providers, health plans, clearinghouses, and their business associates. You are not one of them with respect to your own records, so you cannot violate the law by handling your own information.

The risk is retention, not privacy law. Chat logs live on a company server, and during the New York Times litigation a court ordered OpenAI to preserve output logs, then in January 2026 ordered roughly 200 million de-identified conversations produced. Preservation orders do not care about your intentions. Redact the identifiers, use a temporary chat, and treat anything you paste as something that might sit in a database for years.

What it caught on a bill with six planted errors

I built a sample bill so I could grade the answer instead of admiring it. The structure follows a four-hour emergency department visit billed at level 5, and the line items use real codes with charges in the range hospitals actually post. Six problems are planted in it. This is a constructed document, not a real patient's bill, and I am saying so up front because a reconstructed example presented as evidence would make the rest of this piece worthless.

The planted six: the duplicate chest X-ray, the basic metabolic panel that already lives inside the comprehensive one, the "initial" IV push billed three times, pulse oximetry billed next to the visit that includes it, a supply kit billed on top of the room, and ibuprofen at $23 a tablet.

I ran the prompt through the Codex CLI on gpt-5.6-sol at medium reasoning effort on July 28, 2026, which is a GPT model in a terminal rather than the chat window, so the output is a plain transcript rather than a screenshot. It opened by noting that it was using CMS and NCCI rules as coding guidance and not as price benchmarks.

Five of six, and the sixth is the interesting one. Every code description was correct when I checked them, which matters because a wrong plain-English translation would send you into a billing office arguing about the wrong service.

It also found something I had not planted. HCPCS code J2405 is defined as ondansetron per 1 mg, so a 4 mg dose should bill as four units, and my line 10 said one. I wrote that line carelessly and the model caught me. It pushed back on my line 2 as well, since modifier 26 splits a professional component off a service that has a technical component, and an emergency visit code does not work that way. My construction was sloppy there too.

What it misses, and the mistake it made silently

Line 11 came back CLEAR. Two ibuprofen tablets at $23 each, which is the kind of markup that ends up in news stories, and the model had nothing to say about it because I had told it not to compare prices. Fair enough. So I ran the same bill again with the opposite instruction and asked what each line should cost.

I downloaded the file it cited, all 115 MB of it, and checked. The discounted cash price for an ER level 5 visit is $1,347.90. The complete blood count is $41.21, the comprehensive metabolic panel $154.78, the chest X-ray $229.78. The file's header names Mercy General Hospital at 4001 J Street, Sacramento, last updated February 28, 2026, exactly as quoted. Every figure was right to the cent.

They are also the prices of a hospital nobody told it about. I invented the name at the top of my sample bill, the model matched it to a real facility, and it priced thirteen lines against that facility's file without once saying it had guessed. If your statement says Mercy or St. Something, and hundreds do, you would walk into a billing office holding numbers that are accurate for a building you have never been inside. The arithmetic was flawless and the premise was fabricated, which is a harder failure to catch than a wrong number.

So the working rule is narrower than the headline version. The model reads the bill. You decide which benchmark applies to it, because that decision depends on a fact about your life that is nowhere in the document.

Where to get a benchmark yourself

Four free sources cover most codes. Medicare Procedure Price Lookup carries about 3,900 procedures. The CMS Physician Fee Schedule lookup gives the Medicare allowed amount per code, which is the strongest anchor in a dispute because it is what the government pays for the same work: for CPT 99285, the 2026 professional allowable is roughly $171.35, against the $1,120 on my sample line. FAIR Health Consumer gives regional ranges built from claims data. And the hospital's own machine-readable file, required under 45 CFR Part 180, is the one that changed recently. Since January 1, 2026 those files must encode actual dollar amounts and carry an attestation that the data is true, accurate, and complete, with enforcement of the new requirements deferred to April 1, 2026. The Mercy file I opened carries exactly that attestation text. Hospital-published prices used to be unusable. They are becoming evidence.

Get the itemized bill first, and know what you can demand

None of this runs on the statement that arrives in the mail, which is usually a few summary lines and a balance. You need the itemized version.

[Date]

Patient name: [name]
Date(s) of service: [dates]
Account or statement number: [number]

To the billing department,

I am requesting a fully itemized statement for the services
above. Please list every line with its CPT, HCPCS, or revenue
code, the units billed, the date of service, and the charge per
unit. Please also send a copy of the explanation of benefits
your office received from my insurer for these services.

Please send both to [address or email]. I am making this
request in writing so that we have a shared record of it, and I
will follow up in 14 days if I have not received a response.

[Signature, printed name, phone number]

Now the correction that most articles on this topic get wrong. There is no general federal law entitling every patient to an itemized bill from every hospital. The federal hook is Section 1806(b) of the Social Security Act, which requires an itemized statement within 30 days of a written request and allows a civil penalty of up to $100 for a violation, and it applies to items and services paid under Medicare. Everything else comes from state law, which is strong in California, New York, and Texas among others, plus hospital policy and ordinary practice. Most billing departments will send it without a fight. Just do not build your leverage on a federal right that does not cover you.

What your rights actually cover

The No Surprises Act, effective January 1, 2022, is the strongest protection here, and it is narrower than its name suggests. It covers emergency care and out-of-network providers working inside in-network facilities, which is the anesthesiologist or radiologist you never chose. Ground ambulance rides are explicitly excluded, and only about 22 states have passed their own laws to fill that hole. Mistagged out-of-network charges are worth checking first, because under this law most of them are illegal rather than merely expensive.

If you are uninsured or paying cash, you have a different tool. You are entitled to a good faith estimate before scheduled care, and if the final bill exceeds it by $400 or more you can open a patient-provider dispute resolution case within 120 days of the bill date. There is a small administrative fee, and CMS publishes the current amount on its dispute page.

Insurance denials deserve their own line in your plan. KFF found that HealthCare.gov insurers denied 19% of in-network claims in the 2024 plan year, with individual insurers ranging from 3% to 36%. Fewer than 1% of denials were appealed. Of the appeals that were filed, 44% were overturned. Those three numbers together describe a system that mostly wins by default.

One thing you may have read is no longer true. The CFPB rule that would have removed medical debt from credit reports was vacated by a Texas federal court on July 11, 2025, so unpaid medical bills can still reach your credit file. The free federal help desk for surprise billing questions is 1-800-985-3059.

Where "80% of medical bills have errors" comes from

It comes from an advocacy group in the late 1990s, and nobody has replicated it. Medical Billing Advocates of America and The Access Project are the origin, and the sample was bills that patients already suspected were wrong and sent to a paid auditing service. Bills that looked fine were never in the pool. The methodology was never published, and no peer-reviewed prevalence study has taken its place, which is a striking absence for a statistic this widely quoted.

Two better-sourced numbers point lower. The AMA National Health Insurer Report Card put the commercial claim error rate at 9.5% in 2013, improved from 19.3% in 2011. CMS reported a 6.55% improper payment rate for Medicare fee-for-service in FY2025, about $28.83 billion, down from 7.66% the prior year, and "improper" there mostly means insufficient documentation rather than a patient being overcharged.

My read: the 80% figure is doing rhetorical work for people who sell bill review, and repeating it makes the case weaker rather than stronger, because a single skeptical billing manager can dismantle it. The 45% from the Commonwealth Fund is the number to carry into the conversation. It is recent, it is survey-based, and it describes exactly the situation you are in when a bill you expected to be covered shows up anyway.

FAQ

Can ChatGPT find errors in my hospital bill?

It finds structural errors, which are the ones you can check without medical training. I ran a prompt on a test bill carrying six planted problems and the model flagged five of them: a chest X-ray billed twice, a basic metabolic panel billed alongside the comprehensive panel that already contains it, an "initial" IV push billed three times, pulse oximetry billed next to the emergency visit that normally includes it, and a supply kit billed on top of the room. It also caught a units problem I had not planted, since the ondansetron code J2405 is defined per 1 mg and a 4 mg dose was billed as one unit. What it passed over was the price: two ibuprofen tablets at $23 each came back marked clear. Duplicates, quantities, and bundling are visible inside the bill. Whether a price is reasonable is not.

Is it safe to paste my medical bill into ChatGPT, and does it violate HIPAA?

You cannot violate HIPAA with your own medical information. The law binds covered entities such as providers, health plans, and their business associates, and you are none of those when you handle your own records. The real exposure is retention. Chat logs sit on a company server and can be swept into litigation, which is what happened when a court ordered OpenAI to preserve output logs during the New York Times case, and in January 2026 ordered roughly 200 million de-identified conversations turned over. So treat the paste as publishing. Strip your name, date of birth, medical record number, account number, and insurance ID before pasting, since none of them help the model read a CPT code. Keep the dates of service and the codes, which is all the analysis needs, and use a temporary chat that does not feed training or long-term history.

How do I request an itemized hospital bill, and is the hospital required to give me one?

Ask in writing, and know that the federal right is narrower than most articles claim. Section 1806(b) of the Social Security Act requires an itemized statement within 30 days of a written request, with a civil penalty of up to $100 for a violation, but it applies to items and services paid under Medicare. Outside that, your leverage comes from state law, which is strong in California, New York, and Texas among others, plus the hospital policy and the practical reality that billing departments send them routinely. Send a dated written request with the patient name, dates of service, and account number, and ask for every line with its CPT, HCPCS, or revenue code, units, and charge per unit. Ask for the explanation of benefits in the same letter. The first statement you receive is usually a summary, and a summary cannot be audited.

How do I know if a hospital charge is too high?

Compare it to a published number, because the charge on the bill is not evidence of anything. Medicare Procedure Price Lookup covers about 3,900 procedures and the CMS Physician Fee Schedule lookup gives the Medicare allowed amount per code, which is the cleanest anchor in a dispute. The 2026 Medicare professional allowable for a level 5 emergency visit, CPT 99285, is about $171.35. FAIR Health Consumer gives regional ranges built from claims data. And every hospital now publishes a machine-readable file of its own rates under 45 CFR Part 180, which as of January 1, 2026 must encode actual dollar amounts with an attestation that they are true and complete. Enforcement of the new encoding rules was deferred to April 1, 2026. Those files are large, but they are the hospital admitting on paper what it accepts from other payers.

Do 80% of medical bills really contain errors?

Nobody has shown that, and the number keeps getting repeated anyway. It traces to Medical Billing Advocates of America and The Access Project in the late 1990s, and the sample was bills that patients already suspected were wrong and had sent to a paid auditing service. That is selection bias, the methodology was never published, and no peer-reviewed prevalence study has replaced it. The counterweight is the AMA National Health Insurer Report Card, which put the commercial claim error rate at 9.5% in 2013, down from 19.3% in 2011. CMS reported a 6.55% improper payment rate for Medicare fee-for-service in FY2025, about $28.83 billion, and improper mostly means documentation problems rather than money you were overcharged. The useful number is different: 45% of insured working-age adults got a bill for care they thought was covered, per the Commonwealth Fund in 2024.

Disclaimer

This article is an educational explainer, not medical, legal, financial, or billing advice, and it recommends no specific provider, insurer, advocate, or course of action for your situation. The sample bill is constructed for testing and is not any real patient's statement; its charge amounts are illustrative. The AI transcripts are excerpts of real runs on July 28, 2026 through the Codex CLI on gpt-5.6-sol at medium reasoning effort, and a different model, prompt, or day will produce different output. Coding rules, regulations, fee schedules, and hospital-published prices change, and the $195,628 to $33,000 account is a widely reported story I could not independently verify. Verify every code, price, and legal provision against the cited primary source and with your own provider, insurer, or advisor before you act.

If the bill is larger than what you have on hand, The Bank Said No: Emergency Borrowing Options Ranked compares what each source of cash actually costs. Unpaid medical bills can still reach your credit file, and Does Buy Now, Pay Later Hit Your Credit Score? walks the same report you would check. For a different bill built on someone else's number, How to Appeal Your Property Tax Assessment With AI uses the same evidence-first method.

Sources

• Centers for Medicare & Medicaid Services, No Surprises Act rights and dispute resolution (emergency and in-network facility protections; good faith estimate and patient-provider dispute resolution for uninsured and self-pay patients, $400 threshold, 120 days from the bill date; help desk 1-800-985-3059): https://www.cms.gov/medical-bill-rights/help/dispute-a-bill
• Centers for Medicare & Medicaid Services, No Surprises fact sheet (effective January 1, 2022; ground ambulance excluded): https://www.cms.gov/newsroom/fact-sheets/no-surprises-understand-your-rights-against-surprise-medical-bills
• Social Security Act, Section 1806(b), itemized statement on written request within 30 days for items and services paid under Medicare, civil money penalty up to $100: https://www.ssa.gov/OP_Home/ssact/title18/1806.htm
• Centers for Medicare & Medicaid Services, FY2025 improper payments fact sheet (Medicare fee-for-service 6.55%, about $28.83 billion, down from 7.66% in FY2024): https://www.cms.gov/newsroom/fact-sheets/fiscal-year-2025-improper-payments-fact-sheet
• Medicare Procedure Price Lookup (about 3,900 procedures): https://www.medicare.gov/procedure-price-lookup/
• FAIR Health Consumer, regional cost estimates from claims data: https://www.fairhealthconsumer.org/
• Dignity Health, Mercy General Hospital machine-readable standard charges file, version 3.0.0, last updated February 28, 2026 (discounted cash prices verified July 28, 2026: ER level 5 $1,347.90; CBC with auto diff $41.21; comprehensive metabolic panel $154.78; chest X-ray 2 views $229.78; venipuncture $11.78): https://www.commonspirit.org/patient-resources/dignity-health-sacramento-price-transparency
• HFMA, hospital price transparency guidance under 45 CFR Part 180 (actual dollar amounts and attestation required from January 1, 2026; enforcement of new requirements deferred to April 1, 2026): https://www.hfma.org/price-transparency/new-guidance-makes-hospital-price-transparency-obligations-more-exacting/
• Commonwealth Fund, 2024 research on insured working-age adults (45% received a bill for care they believed was covered; fewer than half disputed it): https://www.commonwealthfund.org/press-release/2024/new-research-insured-working-age-americans-face-widespread-medical-billing
• KFF, claims denials and appeals in ACA marketplace plans in 2024 (19% of in-network claims denied, 3% to 36% by insurer; under 1% appealed; 44% of appeals overturned): https://www.kff.org/patient-consumer-protections/claims-denials-and-appeals-in-aca-marketplace-plans-in-2024/
• LendingTree, November 2024 survey of 2,016 adults (93% of those who negotiated a medical bill had it reduced or eliminated): https://press.lendingtree.com/news/press-releases/93-of-americans-who-negotiate-a-medical-bill-get-it-eliminated-or-reduced
• Becker's Hospital Review on Medical Billing Advocates of America, origin of the "80% of bills contain errors" claim: https://www.beckershospitalreview.com/finance/medical-billing-errors-growing-says-medical-billing-advocates-of-america/
• AMA National Health Insurer Report Card via Healthcare Finance News (commercial claim error rate 9.5% in 2013, down from 19.3% in 2011): https://www.healthcarefinancenews.com/news/ama-health-insurers-cut-claims-errors-half
• OpenAI, response to New York Times data demands (preservation order on output logs; January 2026 order covering roughly 200 million de-identified conversations): https://openai.com/index/response-to-nyt-data-demands/
• Reported account of a family reducing a $195,628 hospital bill to about $33,000 with a chatbot, unverified: https://www.aol.com/articles/family-used-ai-chatbot-fight-170224852.html