Mobile Fidelity Sound Lab is giving Rush’s classic 1974 to 1985 studio run the audiophile treatment with a multi-year reissue campaign on UltraDisc One-Step vinyl, 45RPM 2LP sets, and Hybrid SACD, beginning this summer with Fly By Night and A Farewell to Kings. I have not stopped grinning since seeing Grace Under Pressure, Moving Pictures, and Power Windows on the schedule. Take my money, MoFi. Take it now.

Rush fans have spent decades arguing over pressing variations, bass pedals, drum fills, Ayn Rand references, and whether Signals still gets treated unfairly because the keyboards started ordering bottle service. Mobile Fidelity Sound Lab has now handed them fresh material to debate for the next 30 years.

By the time the arguments over these new editions are finally settled, Geddy and Lerxst may well be reunited with the Professor somewhere beyond the drum riser — or, more appropriately, in some old-school basement in Willowdale, Ontario, where the amps are too loud, the time signatures are suspicious, and nobody has any interest in playing it safe.

This is not the entire Rush catalog, and it does not include the band’s later Atlantic-era albums. The series focuses on the first decade-plus of Rush’s recorded evolution, from the bluesy hard rock of Rush and Fly By Night to the widescreen progressive architecture of 2112, A Farewell to Kings, Hemispheres, Permanent Waves, Moving Pictures, Signals, Grace Under Pressure, and Power Windows.

According to the announcement, most titles in the series will be released as UltraDisc One-Step vinyl sets, while every title in the campaign will receive a numbered Hybrid SACD edition. The first wave begins in summer 2026 with Fly By Night and A Farewell to Kings, followed by Rush in fall 2026. The remaining titles are scheduled to roll out through 2028.

The timing is not exactly subtle, and that is the obvious part. Rush are back in the public conversation in a way many fans probably never expected after the final R40 show at the Forum in Los Angeles on August 1, 2015, and Neil Peart’s death in January 2020. The Fifty Something tour is not Rush trying to pretend nothing changed. It is Geddy Lee and Alex Lifeson stepping back onto the stage together after 11 years, carrying the music forward while honoring the friend, lyricist, drummer, and impossible standard who made the whole thing feel larger than rock music.

The tour is presently on a short interruption after Geddy Lee was diagnosed with laryngitis and bronchitis, forcing the postponement of two Fort Worth shows originally scheduled for June 30 and July 2. Those concerts have been rescheduled for July 11 and July 13, with the tour currently listed to continue in Chicago on July 16.

That makes the MoFi announcement feel less like a random catalog exercise and more like a very calculated moment. Rush are selling tickets, reintroducing their music to arena audiences, honoring Neil Peart’s legacy, and bringing Anika Nilles into one of the most intimidating drum chairs in rock history. At the same time, MoFi is preparing premium physical editions for the exact period of the catalog most likely to send Rush collectors into pre-order behavior normally associated with emergency storm preparation.

Pricing and Formats

MoFi’s current product listings confirm the pricing for the initial releases.

Fly By Night is listed as a numbered 180g 45RPM 2LP set priced at $59.99, with the numbered Hybrid SACD priced at $34.99.

A Farewell to Kings is getting the deluxe treatment first, with MoFi listing it as a limited UltraDisc One-Step 180g 45RPM 2LP box set priced at $125. The Hybrid SACD edition is listed at $34.99.

The 1974 debut Rush, scheduled for fall 2026, is listed as a numbered 180g 45RPM 2LP set at $59.99, with the Hybrid SACD priced at $34.99.

That gives buyers a fairly clear pricing structure for the first wave: standard 45RPM 2LP editions at $59.99, Hybrid SACDs at $34.99, and UltraDisc One-Step titles at $125 when the album receives that treatment. Future title pricing has not yet been fully confirmed by MoFi, so assuming every later One-Step title will remain at the same price would be a little too enthusiastic, even for Rush fans who can count in 7/8 before coffee.

The Mastering Chain Matters

There is another important detail here: MoFi is being specific about the mastering chain on the listed vinyl titles. The Fly By Night 45RPM vinyl page lists the source chain as “1/4” / 15 IPS analog master to DSD 256 to analog console to lathe.” The Rush debut vinyl listing specifies “1/4” / 15 IPS Dolby A analog master to DSD 256 to analog console to lathe.”

That means these are not AAA vinyl reissues. They are sourced from analog master tapes with a DSD 256 step in the chain before the analog console and lacquer cutting. For some collectors, that will be a deal-breaker. For others, it will be less important than the quality of the tapes, mastering choices, pressing quality, and whether Geddy Lee’s bass finally gets the scale and articulation it deserves without turning the upper midrange into a dental procedure.

MoFi also says A Farewell to Kings was mastered at its California studio from tape boxes marked with both “master” and “safety copy,” pressed at Fidelity Record Pressing, and limited to 5,000 numbered copies in the UltraDisc One-Step 45RPM 2LP box set configuration.

Why These Albums Matter

The 1974 to 1985 window is the logical Rush era for this kind of treatment. It captures the band’s most dramatic transformation: the raw bar-band thunder of Rush, the arrival of Neil Peart on Fly By Night, the growing ambition of Caress of Steel, the breakthrough scale of 2112, the fantasy and architecture of A Farewell to Kings and Hemispheres, the leaner precision of Permanent Waves, and the commercial and sonic peak of Moving Pictures.

Then things get more complicated, which is usually where Rush became more interesting. Signals, Grace Under Pressure, and Power Windows pushed the band deeper into synthesizers, sequencers, colder textures, and more layered production. Some listeners still want to pretend that was some kind of betrayal. They are wrong, but at least they are persistent.

For MoFi, those albums present different challenges. The earlier titles need weight, separation, and better organization without sanding off the urgency. The later albums need clarity and space without making the production sound sterile. Rush records are not supposed to sound polite. They need precision, but they also need velocity.

2026 Rush Audiophile Reissue Schedule

• Fly By Night (1975): 45RPM 2LP and Hybrid SACD, coming summer 2026
• A Farewell to Kings (1977): UltraDisc One-Step 2LP and Hybrid SACD, coming summer 2026
• Rush (1974): 45RPM 2LP and Hybrid SACD, coming fall 2026

2027 to 2028 Rush Audiophile Reissues

• Caress of Steel (1975): 45RPM 2LP and Hybrid SACD
• 2112 (1976): UltraDisc One-Step and Hybrid SACD
• Hemispheres (1978): UltraDisc One-Step and Hybrid SACD
• Permanent Waves (1980): UltraDisc One-Step and Hybrid SACD
• Moving Pictures (1981): UltraDisc One-Step and Hybrid SACD
• Signals (1982): UltraDisc One-Step and Hybrid SACD
• Grace Under Pressure (1984): 45RPM 2LP and Hybrid SACD
• Power Windows (1985): 45RPM 2LP and Hybrid SACD

The Bottom Line

The MoFi Rush series arrives at exactly the right moment. Rush are back on the road, the Fifty Something tour has reignited interest in the band’s catalog, and the presence of Anika Nilles has added a new chapter to a story many fans assumed had ended with Neil Peart’s final performance.

The audiophile angle also makes sense. Rush records are built on musicianship, density, movement, and tiny details that disappear quickly when the mastering gets too compressed, too bright, or too polite. If MoFi gets this right, the series could become one of the most important rock reissue campaigns of the next few years.

It will not be inexpensive. The SACDs are the most accessible path at $34.99 each, while the 45RPM 2LP sets are priced at $59.99 and the One-Step titles begin at $125. But this is Rush, not a forgotten soft-rock side quest from 1978. The audience is there, the tour has put the band back in the spotlight, and the catalog deserves serious treatment.

Just remember: all titles are getting Hybrid SACD editions, but not every album is getting the UltraDisc One-Step treatment. That difference matters. So does the DSD 256 step in the vinyl mastering chain. Rush fans notice details. Our review copies are expected shortly.

Where to pre-order: mofi.com

Related Reading:

If a new estimate is correct, and we suspect it is, then Apple is paying dramatically more to produce each iPhone 18 Pro Max compared to its predecessor, and that will mean price rises of $200 or more.

Apple left all iPhone models untouched in its recent and unprecedented price rise, but it was already likely that the forthcoming iPhone 18 Pro range would be affected. Now according to estimates from Counterpoint Research, the bill of materials (BOM) that Apple will pay for the iPhone 18 Pro Max is $300 more than for the current model.

Component costs are always increasing, which is why the iPhone 15 Pro Max cost 12% more than its predecessor, and the iPhone 16 models also had a slight rise. But because of the global chip shortage, Apple is having to pay significantly more for certain components.

It’s specifically the memory that will account for the higher costs for Apple. Counterpoint Research does not give detailed estimates, but on a 12GB iPhone 18 Pro Max with 1TB storage, it is expected that compared to the previous model:

• NAND will cost four times more
• DRAM will cost almost four times as much
• Processor costs may remain the same
• Camera costs will rise slightly
• Display costs may decrease slightly

Estimated component costs for the iPhone 18 Pro Max – image credit: Counterpoint Research

Apple has previously tried to hide price increases with a strategy of removing lower-cost configurations. With the iPhone 15 Pro Max, for example, Apple dropped the 128GB model and had its starting price be the same as the previous model’s 256GB configuration.

The new report expects that Apple will make similar changes for the iPhone 18 Pro range. It predicts an average $200 price rise for consumers, although even that is said to mean Apple will see reduced profits.

All of these estimates are reportedly based on information regarding each separate component that is expected in the iPhone 18 Pro Max. However, the bill of materials is only useful as a comparison to the BOM of previous models.

It takes little account of Apple’s buying power meaning that it can sometimes negotiate better deals than its rivals. There’s also no calculation of development and marketing costs, which won’t have risen because of the chip shortage, but will certainly have gone up.

Consequently, Apple is juggling much more than the BOM with this iPhone launch, and we won’t know the outcome until the phones are released in September.

This does, though, seem like the worst year for Apple to launch its potentially exorbitantly-priced iPhone Fold.

The internet loves a good usage dashboard, and today Anthropic is releasing one for Claude designed to help you “reflect” on the time you’ve spent with its chatbot. Think of the new feature as something of a cross between Spotify Wrapped and Apple’s Screen Time tool. You can generate your first report by opening Claude’s settings menu and navigating to the new “Reflect” tab. For now, the dashboard is only available through the Claude web client and desktop app.

At the top, you’ll see a paragraph-long summary of your recent conversations with the chatbot. By default, Claude will collate your last month of interactions, but you can also see your last three, six or 12 months of usage. Under that, the interface lists your most active day, peak hour and total chats over your selected time period, with a visual representation just below. In the future, you’ll also be able to see how much time you’ve spent chatting with Claude, but for now that metric isn’t available. If you continue scrolling, there’s a toggle to configure break reminders and time limits. You can independently set these of the Reflect interface by navigating to the “Time and focus” tab, and dismiss the nudges if you’re in the middle of something.

Anthropic

Further down, there’s a breakdown of topics you’ve discussed with Claude, with a percentage assigned to the ones you bring up most often. If you’ve been following along with the screenshots, all of this should feel broadly familiar. The penultimate section does things a bit differently, offering a set of AI “fluency” recommendations designed to streamline your usage of Claude, which are grouped around guidelines Anthropic co-created with a group of academics. For example, if Claude finds you frequently re-establish the same or similar context when you go to write a question or request, it will recommend you use its Projects feature to group your prompts together, so that you don’t need to repeat yourself so often.  

For a more specific example, I’ve been working on a story about inference costs — the amount of AI labs like Anthropic pay for their trained models to process data — and I’ve turned to Claude a few times to track down statements from executives like Dario Amodei and Sam Altman. Based on that usage, the Reflect dashboard recommended creating a custom fact-checking skill for Claude. When I went ahead with that suggestion, the chatbot devised a text template to ensure it would always list its source for a claim, as well as its confidence in the information it had found alongside any caveats. I’ll admit I found the template helpful, and it wasn’t something I would have thought to ask Claude to do on its own.

Anthropic

According to Ryn Linthicum, Anthropic’s head of wellbeing policy, Reflect came out of a study the company’s Societal Impacts team recently completed where participants expressed a mixture of optimism and anxiety around Claude and other AI products. “We were really intentional about building [the dashboard] with an eye toward how we can upskill people’s usage of Claude, not in a way that encourages them to spend more time with it, but instead enables them to get more efficient at meeting their goals, and hopefully get off of Claude or preserve the things that they want to think about,” Linthicum told Engadget.

Linthicum says part of the reason the dashboard doesn’t currently display the exact amount of time you’ve used Claude is because it wasn’t a metric Anthropic had been tracking internally, due to it being something the product team “didn’t want to maximize.” Looking forward, they note Anthropic plans to surface that information for users so they can use it alongside the usage management settings. In the future, the dashboard will also be available mobile, and reflect your usage there.

Anthropic is releasing Reflect in beta today. Notably, the company is making the feature available to Free accounts, in addition to Pro and Max subscribers.

from the pay-no-attention-to-the-formaldehyde-plumes dept

Civil rights groups like the Southern Environmental Law Center (SELC) have noted how Elon Musk’s Colossus xAI data centers in Memphis disproportionately pollute the air in minority neighborhoods. A joint lawsuit by SELC, Earthjustice, and the NAACP filed last April argued that Musk and friends didn’t bother to get the necessary permits to run the 57 gas turbines powering the system.

As is so often the case, Musk’s “innovation” is quite often just a combination of media manipulation, opportunism, crony capitalism, and openly ignoring public safety. The more you look, the more of a theme it becomes across Musk’s entire post-IPO delusionverse.

Anyway, the lawsuit points out that his Memphis-area data centers are violating the Clean Air Act, funneling all manner of pollutants into minority neighborhoods that already see a disproportionately high number of pollution-caused childhood illnesses:

“xAI’s power plant in Southaven has the potential to emit more than 1,700 tons of smog-forming nitrogen oxides (NOx) each year. The staggering emissions numbers likely make the facility the largest industrial source of NOx in the greater Memphis area, an area already failing to meet national smog standards. The illegal turbines also have the potential to release up to 180 tons of fine particulate matter, 500 tons of carbon monoxide, and 19 tons of formaldehyde—a toxic, cancer-causing chemical—each year.”

Musk also promised to build a next-gen water filtration system so that the xAI data center doesn’t imperil the local water supply, but he simply decided to apparently not do that. Instead, Musk ran crying to the Trump administration, whose DOJ is trying to have the pollution case-dismissed on national security grounds, because the administration sometimes uses Musk’s shoddy fifth-place AI services.

Hoping to quell growing anger in Memphis at the fact he’s openly violating environmental law, Musk is leaning on his other company, Starlink, to try and calm down locals. Starlink announced last week they’d be offering Memphis-area locals half-off the typically expensive Starlink monthly rate for an unspecified amount of time:

“The unique capabilities of the Colossus datacenters could not be accomplished without the partnership and support from the local Memphis community.” SpaceX’s vice president of Starlink engineering, Michael Nicolls, wrote on X on Tuesday.

“Happy to bring affordable and great @SpaceX @Starlink connectivity to our neighbors,” Nicolls added.

I’m sure a temporary (these sorts of discounted rates never last) internet access discount will definitely help the kids with formaldehyde-driven asthma. There are a few other layers of irony: one being that data has repeatedly shown that Starlink is routinely too congested to be of help in more densely populated areas like Memphis. Hidden “congestion” fees also ensure the service isn’t really affordable.

Meanwhile, you have guys like Mark Andreessen pretending to be confused why civil rights groups like the NAACP wouldn’t be big fans of discriminatory environmental pollution. As with all (bipartisan) opposition to AI data centers, the very legitimate complaints going on outside of Memphis are being portrayed as unreasonable attacks on innovation by radicals:

These people are, in case you’ve not been paying attention, just foundationally not good human beings, who simply love to engage in phony surprise at the width and breadth of the public backlash to AI.

Filed Under: ai, ai data centers, automation, broadband, elon musk, memphis, naacp, pollution, southern environmental law center

The $140, 1,252-piece set goes on sale on August 1.

Kylian Mbappe, Erling Haaland and Lionel Messi are among the stars aiming to fire their countries one step closer to glory in the FIFA World Cup 2026 quarter-finals — and you can live stream all four games around the world for free.

France face Morocco, Spain play Belgium, Norway take on England and Argentina go up against Switzerland in four days of football action.

A new phishing-as-a-service (PhaaS) operation called Forg365 focuses on stealing Microsoft 365 accounts by combining adversary-in-the-middle (AiTM) and device code methods with AI-assisted lure generation.

The platform also provides a browser extension for continued access to the Microsoft services linked to the compromised accounts without the need to re-authenticate.

Researchers at ZeroBEC email security company say that many of the features in Forg365 are present in other infamous PhaaS platforms such as Kali365 and Sneaky2FA, although they could not establish a connection.

Their investigation began by analyzing phishing emails that posed as business documents, carefully crafted to mimic a trusted service.

“The observed sender domain used Amazon SES delivery, while the message body included SendGrid-hosted image or tracking resources,” ZeroBEC says in a report today.

This combination of legitimate services and phishing infrastructure indicates a mature PhaaS operation capable of blending its messages into regular email traffic.

The platform features device-code phishing, adversary-in-the-Middle (AiTM) phishing, AI-assisted email content generation, token and cookie management, and post-compromise operations.

Digging deeper, the researchers obtained access to the Forg365 dashboard, which allows creating new phishing campaigns, manage phishing links, configure OAuth apps and SMTP profiles, manage tokens, and generate phishing emails with the help of AI.

Although the use of AI in crafting custom phishing lures is not new, the researchers highlight that the feature is directly integrated in Forg365’s panel, allowing the operator to create the malicious emails, prepare the text, and refine the messages from the same dashboard used to control the post-compromise activity.

According to the researchers, this integration is strategic, as “AI reduces the cost of developing custom phishing content, but it also reduces the cost of building custom PhaaS platforms.”

The panel also includes an account intelligence dashboard and a keyword monitoring feature that scans compromised mailboxes for predefined terms, alerting operators whenever a match is detected.

Operators are provided a browser extension called ForgCookie that is compatible with Google Chrome, Microsoft Edge, and Brave, and is specifically designed for automatically refreshing Microsoft SSO cookies.

The extension works by requesting account data from the Forg365 backend, clearing session cookies, and triggering a silent OAuth flow to capture the fresh cookies.

This provides the attacker with persistent access to the Microsoft services associated with the victim’s account.

According to ZeroBEC, Forg365 supports two primary attack paths: the trending device-code phishing and the more traditional AiTM phishing.

In the first case, victims are shown a Microsoft-style verification code page and instructed to complete authentication using Microsoft’s device-code flow designed for any endpoint with input constraints (e.g. smartTV, IoT appliances, tools without a browser).

Rather than targeting the victim’s password directly, the victim is tricked into authorizing an attacker-controlled gadget through the legitimate OAuth 2.0 device code flow authentication method.

For AiTM phishing, the platform uses a proxy for the authentication requests and data exchanged between the Microsoft infrastructure and the target account, capturing session cookies in the process.

To prevent researchers from accessing the administration panel, Forg365 has an AntiBot feature that boasts “AES-encrypted redirectors, bot detection, debugger traps, sandbox checks, and polymorphic code.”

Additionally, when a VPN connection is detected, the platform redirects to innocuous content instead of exposing the phishing pages.

ZeroBec reports that the platform leverages Amazon SES for phishing email delivery and Cloudflare Pages for the landing pages. Also, the Gophish infrastructure is used for campaign delivery.

Users are recommended to restrict or disable Microsoft device-code authentication unless required and to monitor Microsoft Entra logs for device-code authentication events.

Mailbox rules, new device sign-ins, Microsoft Authentication Broker activity, and OAuth grants must also be investigated for unexpected entries.

If a compromise is suspected, all tokens and sessions must be revoked and refreshed as soon as possible.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Advertisement

Get the whitepaper

I just celebrated a big milestone I hope you never reach: I hit my health insurance plan’s $10,150 out-of-pocket maximum less than five months into 2026, thanks mostly to two major eye surgeries. That means no more co-pays or coinsurance for authorized in-network care this year, as long as I keep paying my monthly premiums.

But earlier this year, as I accumulated what seemed like an eternal fountain of medical expenses, I couldn’t help but wonder whether I was paying bills that contained errors. As a certified financial planner and a longtime personal finance writer and editor, I’m familiar with how many medical bills contain mistakes that make them more expensive. 

Occasionally, medical bills contain obvious errors, such as a charge for a treatment you explicitly declined. Otherwise, though, these mistakes are often difficult for a typical patient to spot. Finding billing errors can require clinical knowledge, along with an understanding of medical coding, revenue cycle management and the opaque American health insurance system. 

You may also have to sift through huge volumes of information. For example, I discovered that I’d had 87 insurance claims during the first four-and-a-half months of 2026 and that the contract I’d signed during open enrollment was 149 pages long.

I had no desire to get an education on medical coding or meditate over the meaning of 149 pages’ worth of insurance jargon, but I thought perhaps generative artificial intelligence would be up for the task. After all, AI excels at taking in complex information and finding irregularities in huge volumes of data.

Turns out, though, using AI to find errors buried in my stacks of medical bills wasn’t as easy as I’d hoped. Here’s how I did it — and what I learned.

How I used AI to search for medical bill errors

I expected to find a plethora of AI tools to help patients identify billing errors. Wrong.

Most AI tools aimed at improving billing accuracy are designed for providers, not patients, for obvious reasons.

The few patient-facing tools that exist often target a fairly narrow segment of billing issues. For example, Counterforce Health uses AI to analyze medical bills and records to help patients understand why their insurance claims were denied and to draft an appeal. But few AI resources for patients exist that offer a general audit of your medical bills.

So I settled on using generative AI — specifically, my $20 monthly ChatGPT Plus subscription, which had already been hugely helpful to me in crafting scripts to use with my insurer when they attempted to deny care.

My step-by-step process:

1. Narrowed my focus to claims where I’d spent at least $150 to simplify the review.
2. Retrieved my 146-page insurance contract and explanations of benefits, or EOBs, from my insurer’s website.
3. Requested itemized medical bills from my providers, which are essential for identifying costs and inaccuracies.
4. Compiled 14 itemized bills and EOBs, along with a spreadsheet summarizing all 87 of my claims.
5. Redacted all personal information — such as my name, date of birth, address and insurance ID number — from the documents before uploading them to the AI.

Then I used the following ChatGPT prompt:

Act as a medical billing expert and auditor with deep knowledge of the US healthcare system, medical billing codes, surgical billing practices and outpatient billing practices. I will provide my insurance contract, an itemized bill and an explanation of benefits. Look for incorrect charges, unusually expensive or questionable charges, mathematical errors, charges that appear inconsistent with my insurance contract and other potential inaccuracies.

Did ChatGPT find medical billing errors?

Before I’d even uploaded my itemized bills to ChatGPT, I could see an obvious flaw: How was AI supposed to know whether the bill accurately reflected the care I received?

For example, the first two itemized bills from the surgical center included 31 to 60 minutes of operating room time. But I hadn’t brought a stopwatch into surgery with me. 

Maybe ChatGPT would have flagged it if I’d been billed for several hours of surgical time for a procedure that usually takes a few minutes. But how would ChatGPT know if, say, I was only in the OR for 28 minutes? Or whether the 200 or so preop eyedrops I’d received were accurately reflected in the itemized surgical bill?

Instead, ChatGPT kept focusing on things such as the fact that the amount my insurer paid looked ridiculously low compared to what the surgeon, anesthesiologist and facility actually billed. Fair enough, but that’s more an indictment of the opaqueness of the American healthcare system than a sign of a billing error.

AI told me to look into the only claim marked “denied” on the spreadsheet. But the reason for the denial was that my surgeon had voluntarily withdrawn and resubmitted it before my insurer had even processed it. A few pharmacy claims had been reversed, but those also had an easy explanation: The pharmacy had automatically processed a couple of refills I hadn’t needed.

I quickly lost hope that AI would help me find potential billing errors that I hadn’t already identified. So I started asking it point-blank questions about specific claims.

There was one potential error I had already spotted: For one procedure, I had been charged both a $100 specialist co-pay and a $150 co-pay for a physician-administered drug, or $250 total. I talked with a customer service rep online who said I should only have been billed for one. So, I uploaded my live-chat conversation with the rep, asking:

This conversation with an insurance representative says that I will only owe a $100 retina specialist co-pay or a physician-administered drug co-pay of no more than $150 for anti-VEGF injections, but I was charged $250 for the visit and injection. Is this an error?

ChatGPT quickly dashed my hopes on that front, directing me to the section of my 149-page insurance contract stating that I was responsible for both co-pays. The insurance rep had clearly been wrong.

OK, but why had I paid $11,512 in co-pays and co-insurance when my maximum patient responsibility was $10,150?

ChatGPT kept insisting that I’d only paid $10,150. Then it hit me: ChatGPT showed that I’d only paid $10,150 because that was my patient responsibility, according to my EOBs.

Three weeks later, I’d had the exact same surgery on my right eye. Since I’d hit my deductible, I’d had to pay a lesser amount: $1,552, which I assumed represented 50% co-insurance. But my EOB listed my patient responsibility at $999.

Again, I asked ChatGPT about the discrepancy. This time, it pointed out something that seems obvious in retrospect.

The $1,552 I’d paid upfront was the amount I was actually responsible for after the first surgery. Since I was having the same surgery on the other eye, the facility had estimated the amount I’d owe based on the first surgery, without accounting for how my patient responsibility would change after I hit my deductible.

So ChatGPT confirmed that I’d overpaid by $1,512 for that second eye surgery, and it helped me understand why. But it didn’t actually find the $1,512 overpayment on its own. I found that by keeping careful records of every medical expense I incurred.

Just supplying ChatGPT with all the information it needed to confirm the error took a huge amount of work. In that respect, it seems as if using ChatGPT to comb through medical bills is a bit like using tax filing software: It’s only as accurate as the data you supply, and gathering all that takes a ton of work.

It’s possible that my itemized medical bills did contain additional errors. If they did, that’s a matter for my providers and my insurer to fight about. As long as I don’t have to pay more than my $10,150 out-of-pocket maximum — and I have no doubt that the amount I’m responsible for as a patient reached that amount — I honestly don’t care if they have to fight between themselves; that’s not my problem.

As of this writing, I’m still waiting for my $1,512 refund.

How to do this yourself

If you want to use AI to help you audit your own medical bills, keep these prerequisites in mind:

• Request itemized bills: You’re entitled to a breakdown of every cost incurred during a procedure. Contact your provider to request this, as it’s essential for identifying specific billing inaccuracies.
• Redact sensitive data: Before uploading any documents to an AI tool, remove all personal information, such as your name, date of birth, address and insurance ID number.
• Maintain a personal spreadsheet: AI is only as accurate as the data you provide. Keep a detailed log of every medical claim, the amount billed, the amount your insurer paid and your actual out-of-pocket payments. This manual tracking is crucial for spotting discrepancies between what you were charged and what you were actually responsible for.