My colleagues found a way to raise my blood pressure this morning. I woke up and saw that our resident car designer, Adrian, had thrown a link into Slack going to an article about some random woman on TikTok who somehow found her car dead after changing a sun visor. She says her husband used AI to try to fix the car and probably wasted a bunch of time. Everything about this article and the whole situation is insane, and I have to say something. You can’t trust AI to fix your car without constantly fact-checking it, and you’re better off just logging into an old-school car forum or a Facebook group for your car.
This story comes to us from Motor1. The publication recently debuted a “Trending” section, where some of the site’s writers will comment about car-related stories from TikTok. Many of these stories involve people not knowing that U-Haul charges fees for truck mileage, people not knowing how car financing works, or people not knowing how car dealerships work. If you want to consume TikTok without actually watching TikTok, these posts seem to do the trick.
Normally, these posts aren’t really our style, and that’s fine. Every site has its own vibe! But this piece caught our attention because of the fantastically terrible way that AI ruined this poor woman’s day, and how seemingly nobody noticed the obvious path of diagnosis.
The Conundrum
@sarahnovakwallace Ladies if your hubby is a mechanic or works for @Chevrolet or @General Motors HELP. I can’t get my car to start #chevrolet #gm #generalmotors #mechanic #tahoe
First, I’ll give you the TikTok. If you cannot watch the embed above because you’re at work or have TikTok blocked, click this link. You should be able to view the video without creating a TikTok account.
In the video, TikToker Sarah Novak (@sarahnovakwallace) talks about how the driver-side sun visor in her 2020 Chevrolet Tahoe broke. She ordered a new OEM part, and her husband installed it. The next day, Sarah attempted to drive her Tahoe, only to hop in the driver seat, attempt to start the vehicle, and find only a Christmas tree of lights and no start.
Sarah and her husband were immediately stumped and turned to AI, which told her that if you mess with any wiring, the vehicle’s various modules might stop working. So, her husband disconnected the battery for some time, hoping to reset whatever computer was angry. This didn’t work, and the next time Sarah attempted to start the SUV, she got exactly nothing out of the vehicle. Sarah continues by saying: “We’re ChatGPTing it, but I don’t know if we’re doing it right or what the problem is.” Then she says she only recently spent $400 on a brand-new battery.
Sarah’s car still wasn’t working when she published the video on December 29, and she was so desperate that she reached out for help on TikTok. Alright, so the stage is set. Here’s what happened next.
AI Fails At Diagnosis
Motor1‘s article doesn’t start off great, as whoever is in charge of graphics at Motor1 used a photo of a third-generation Tahoe, when Sarah’s Tahoe is a fourth-generation model. I’ve noticed this issue in a lot of TikTok articles on various websites. Anyway, the writer’s response was this set of paragraphs, from Motor1:
To many drivers, a sun visor feels like a purely cosmetic part. In modern vehicles, especially full-size SUVs like the Tahoe, that assumption can be misleading.
According to General Motors service documentation and industry explanations from the Society of Automotive Engineers, late-model GM vehicles rely on multiplexed electrical networks, such as CAN and LIN, to enable dozens of control modules to communicate. Components in the headliner area can share wiring paths with airbag systems, interior lighting, and the body control module, or BCM, which manages functions like starting authorization, door locks, and dashboard displays.
If wiring is pinched, grounded, or left partially disconnected during an interior repair, the BCM can detect a fault and disable vehicle startup as a protective measure. Automotive electrical experts frequently note that once a BCM fault is logged, simply disconnecting the battery may not restore functionality, as many modules retain error states until they are cleared with a diagnostic scan tool.
The writer then talks about how battery disconnects could make the no-start condition worse before dropping this nugget:
That distinction helps explain why Novak’s nearly new battery was unlikely to be the root cause, even though battery failure is a common suspect in no-start situations.
But here’s the thing: there’s no reason to just rule the battery out by default. Just because a battery is new does not mean that it cannot be drained. If Sarah’s husband messed around for too long with accessory power on or accidentally let the visor’s light stay on all night, those alone could drain a battery regardless of its age. Replacing a visor in one of these SUVs isn’t rocket surgery, either. Watch this:
Likewise, the only electrical connection that the mirror has with the rest of the car is for its light. If you watch the video above, you’ll see just two wires and a plastic connector for them.
One comment in the video reflects this:
The terrifying thing is that, depending on how you word the question, AI will give you some entirely unhelpful answers. Here’s what Google AI gave me after typing in “2020 chevy tahoe won’t start after sun visor replacement.” It went straight to assuming something is wrong with the vehicle’s Body Control Module:
Your 2020 Chevy Tahoe likely won’t start after a sun visor replacement due to a disconnected or blown fuse for the Body Control Module (BCM) or ignition, a loose battery connection, or potentially disrupting the ignition switch wiring, common after interior electrical work; check the main fuse box (under the hood and driver’s side) for relevant fuses like BCM or ignition, ensure battery terminals are tight, and if issues persist, consider resetting the BCM by disconnecting the battery for a few minutes.
Possible Causes & Solutions
Blown Fuse: The wires for the sun visor lights might have shorted or been pulled loose, blowing a fuse.
Action: Check the fuse box under the hood and the one on the driver’s side (left of the steering wheel) for blown fuses related to the BCM, ignition, or interior lights.Loose Battery Connection: You might have jostled the battery or terminals during the process.
Action: Ensure battery terminals are clean and very tight.Disconnected/Damaged Wiring: The wires for the visor might be connected to a circuit that also affects starting or dash power.
Action: Re-examine the visor’s wiring and the connection points in the headliner. Ensure no wires were pulled or damaged.BCM Malfunction: A brief electrical issue can sometimes put the Body Control Module into a protective state.
Action: Try disconnecting the battery negative terminal for 10-15 minutes to reset the BCM, then reconnect and try starting.
ChatGPT was equally unhelpful:
A no-start right after replacing a sun visor on a 2020 Chevy Tahoe strongly suggests something electrical was disturbed. The visor area is tied into several sensitive systems (airbag, BCM, interior power), so this is a common scenario. Here’s a clear way to narrow it down.
I won’t bore you with the details, but ChatGPT gave me a list that started with checking fuses related to the airbag system, the BCM, and interior lighting. Then it said that the visor replacement could have disrupted the anti-theft system or the airbag system. To its credit, ChatGPT did tell me that a drained battery could be the cause, but rated it as unlikely. If I logged out, logged back in, and gave ChatGPT the same question, it gave me a slightly different answer. One of the AI’s suggestions was to replace a fuse even if it is only slightly blown. Is a slightly blown fuse like being slightly pregnant?
I would not be surprised if AI gave Sarah’s husband an answer closer to what Google AI gave me, considering he tried to reset the vehicle’s computers by disconnecting the battery. The problem I had with this is that AI has no idea how to diagnose or how to troubleshoot. It can only take what it can search and spit it back out. Thus, poor Sarah and her hubby are chasing around computer issues without even having checked the easy stuff.
To be fair to the AI, it didn’t give the most insane response; that would go to some of the commenters on TikTok, who told Sarah that she fried the SUV’s Powertrain Control Module (PCM), told her to reflash the anti-theft system, or told her to reprogram the key. What’s with the suggestions for nuclear options after basically no troubleshooting?
Try The Easy Stuff First
Here’s what I would have done. First, I would have checked the battery. If it’s below 12 volts, or maybe the clamps aren’t tight, that’s probably the problem. A multimeter is only $8 at Harbor Freight, and anyone who works on their own car or their own home should own one, anyway. A dead battery would cause Sarah’s issues with flashing dash lights, no start, and then, later, a totally dead car.
If the battery tests good, uninstall the new visor. Use a flashlight to see if you somehow penetrated a wiring harness when you installed the visor. This is unlikely if you used the OEM screws, but it is still possible. It’s also worth checking fuses.
It’s possible that the no-start condition is not even related to the visor at all, and it’s just a wild coincidence that the Tahoe is not starting a day after the visor’s replacement.
I wouldn’t be sweating about the PCM, the BCM, or the anti-theft system until after I tried the easy stuff, like checking the battery. It wouldn’t be my first place to look. It would be like jumping straight to the assumption that you’re going to die just because you have a new cough. But this is one reason why AI cannot be trusted for wrenching unless you double and triple-check it. At that point, you’re sort of better off without it. AI has no nuance or context. It has no idea what it’s doing.
Just to be clear here, I’m not blaming Sarah or her husband for anything. They did the best they could with what they thought was a reliable resource. Not everyone spins wrenches or gets into car diagnosis, and that’s okay! I reached out to Sarah and will update if I hear back. Sarah has not published a follow-up, so maybe she and her husband figured it out.
If you have a problem with your car, don’t ask AI, and be careful soliciting opinions on just regular social media. Go to a make or model-specific Facebook group or an old-school car forum and just pitch your question there. You’ll almost certainly get better advice than you’d find on TikTok, and certainly better than AI.
Top graphic images: GM; DepositPhotos.com
To use LLM’s correctly, one must do enough research to fact-check the answer. In other words, to use LLM’s correctly, one must not use them.
It also is incredibly helpful if the questions you ask are using the language you’re asking them properly. I’ve had poor grammar come back to obfuscate a question fed to Chat GPT many times.
An incredible self-own Matt, great job
Username checks out. :-/
While the grammar I used is clunky, awkward, and the sentence is somewhat convoluted, the grammar is not technically ‘wrong’. It just leans towards my my autistic tendencies and stream-of-consciousness writing style – that’s all.
Here, I had AI [barely] re-write it for you to increase your Zen:
“It’s also incredibly helpful when the questions you ask are phrased clearly and use proper language. I’ve had poor grammar obscure the meaning of questions I’ve asked ChatGPT many times.”
Yeah it’s almost like it’s a shitty system for credulous rubes who can’t think for themselves. Did you really think it was worth burning resources for that stupid rewrite? I wasn’t confused about what you had said and you aren’t disabusing me of my initial impression.
It’s a “worst-of” if you get it through those posts. Usually, they find the videos that will drive engagement, which is easiest by picking someone who made a dumb mistake or who has a strong positive or negative opinion of a brand that people will argue with. I’m not a Tik Tok person, but, like any social media, there are good and bad parts and any curated selection is going to show you a very specific subset.
This will be slightly geeky, but beyond the fact that LLM’s simply aren’t going to stop hallucinating anytime soon (which is only one of many reasons to question everything you get from generative AI), but there’s some research I just read about that is pretty similar. Look up “Semantic Leakage” and you’ll probably find the paper in question. Basically, they’ve found that an LLM will determine that a person is a school bus driver simply because it was trained that they liked the colour yellow. Pull that into the world of automotive repairs, and simply saying “working on wiring” might return results around PCM…and this seems to be because LLMs are learning weird nth level correlations between words, not concepts.
The weirdest example I read about involved someone training a system on old names for birds, and the LLM would then respond as though it was the 19th century. (Eg, ask for a recent invention, and the LLM responds with electric telegraph)
That makes perfect sense; any dataset it trained on that included defunct names for birds would probably be either historical fiction or works from that time period. It’s just putting the words together that it thinks make the most sense together.
It’s like you said, it’s making weird connections between words without understanding the underlying concepts.
I use Perplexity all the time, but on things like this I always to the sources it surfaces for final diagnosis and repair
I don’t need to ask AI to get a wrong answer, I’ll just click on one of the AI generated, search algorithm optimized websites that inhabit the first 25 results of every Google search like a normal person.
Remember the olden days when Google used to provide useful search results?
AI fed me a wrong answer just yesterday. My 17 accord threw a light for the passive entry (hand on handle to unlock/button to lock) and disabled it. According to AI, it was a dead remote, or a bad wire or fuse/relay.
It was that I disconnected the vents on the back of the console to adjust the parking brake.
Good luck with searching old-school automotive forums. All the ones I used to frequent have been gone for a decade, and took all their builds and technical threads with them.
I’d probably rather replace my engine control module than visit Facebook. I don’t know if I’ve ever loved a car enough to go on Facebook today to try to save it.
I still regularly rely on 20 year old forum posts to help answer current questions I have on my 40 year old car. When those go away some of that info will be lost forever.
You’re lucky to still have the forums available. I remember when everyone went from the forums to FB. Then the forums shut down.
Unfortunately, finding information on FB is terrible. There is no good thread view so if the answer is two levels deep I would never find it.
Expanding individual comments one at a time is not very time efficient. After a bit of that I figure I would be better just randomly trying stuff on the car.
Yep, a few already have shut down.
Archive.org has been a blessing to find some stuff that has disappeared. I’ve been able to find tech info on the personal sites of some forum members who have long passed away this way.
Maybe it’s better now, but I used to get frustrated searching through archives because the images are usually gone.
Back in the forum days, storage and bandwidth was expensive so images were usually hosted on a public image hosting site or on a personal website. I can’t even remember what the domain was of my personal site, or when it was last live.
Not only is there no good thread view, but the search is garbage and the algorithm wrecks your chances of seeing any straightforward answer. The posts with the highest engagement win, and that engagement usually means no one found the right answer.
I have been off of FB for a years now (late arrival, early departure), but it was terrible. My entire feed was only posts that FB knew I would want to argue with.
But then I can’t really argue the way I wanted to because it’s my real name and my aunt might get it in her feed. The internet was much more fun before FB.
Yeah, I ended up setting all my posts to be visible to everyone but family before I stopped posting altogether. I’ll occasionally post some travel photos visible to my family, but nothing else. When my mom dies, I’ll probably delete it entirely.
FB has had hiccups where they accidentally made all hidden groups and post visible to everyone.
Yeah, that was a big part of why I stopped posting anything but a few travel photos. I do not need my evangelical preacher cousin trying to debate me on bodily autonomy or whether LGBTQ+ folks deserve human dignity again. (I actually blocked him, but my point mostly stands.)
Sadly they were largely dead before LLMs, I’ve always assumed it was everyone dreaming of becoming YouTube famous and posting their DIY walk throughs there-also as someone else points out FB groups probably also helped.
In my case, all my local car forums moved to FB. After a certain amount of traffic moved in the FB direction, it didn’t make sense to keep the forum alive.
Yeh, that makes sense, I’ve dipped in and out of so many old cars that I wasn’t watching a particular forum when it made the move. But it does suck, FB is a crap replacement-and from what I can tell not even more civil which might have been its one theoretical advantage.
Local forums were always civil as far as I remember. There were always cruises, events, tech days, etc. If you insulted someone, you’d have to see that person at the next event.
I was also on national forums that were horrifying in the way we spoke to each other. A lot of us were youngish and had jobs that put us on the road a lot so we’d end up meeting in person too and everyone got along fine (and went right back to terrible on the forums the next day).
To be fair, there are a ton of genuinely useful guides on youtube
Oh absolutely, but I am sad that forums have disappeared, I often find a written out with photos DIY easier to follow when I’m actually working on something than a video. Otoh a good video sometimes show little nuanced things that written DIYs often skip over.
“But this is one reason why AI cannot be trusted for
wrenchinganything.”I hate AI only slightly more than articles that are reposts of TikTok videos. I have a reputation at work and amongst friends for yelling at people that google and just read the Gemini AI overview. At one point, in a meeting, I yelled “Find a source. Gemini lies all the time.” They said “The next result is Reddit, do you trust Reddit more than AI?” “Yes. And that should tell you something,” was my reply.
At least on Reddit there is always going to be one contrarian who argues with the consensus. Of course, determining whether they have a valid point or not requires critical thinking.
I find it’s useful as a tool, but like other tools, you have to know how to use it correctly if you want good results.
Just last week, I was trying to diagnose why my 911’s check engine light is on. I found an AI summary of forum posts to be really helpful, but only after I’d read the exact code myself and thought through the possibilities based on my (limited perhaps but hard won) knowledge of the car.
Google has served up so many blatantly false AI answers to me that I pretty much disregard anything it tells me and it has actively caused me to distrust anything AI says. The other day it gave me info about drum brake wheel cylinders when I searched for something regarding brake master cylinders.
A friend recently posted a humorous example of this. He prompted ChatGPT by saying his car’s manufacturer recommends changing oil every 5,000 miles, but he has driven 10,000 miles so he needs to know how many oil changes to get. ChatGPT responded that he should do two oil changes now in order to catch up. *facepalm*
Yeh I have had it give me so many obviously incorrect replies to car repair queries that I don’t trust it for anything else where I have less subject matter expertise.
That is exactly why the AI bubble will burst. Those who own millions in stock will dump it right before the inevitable crash, leaving others holding the bag.
The sooner AI goes away, the better.
I’ve only tried ChatGPT once, with a simple question:
Which state has its most populous city furthest from the state capitol?
The AI confidently answered “New York, where the capital city is 87 miles from the most populous city.”
(For the record, the correct answer is “Alaska”.)
When I was a kid, my dad told me “don’t believe everything you see on TV.”
Just because your device is magic doesn’t mean it’s right.
Yes, do not ask AI how to fix your car. It has no skin in the game, and there is no consequence for being wrong. Model-specific groups or Reddit are your best answer. Even more important, you have to know how to get the best response from specialized groups.
If you don’t know what you’re doing, get in there and post your situation and question. And then- now this is important – make a DIFFERENT account and answer your own question to the best of your ability, knowing it’s wrong. Specialized social media groups like to help people in need, but they LOVE to correct supposed experts when they are wrong. You’ll get a master-level explanation on your issues in record time.
this is brilliant.
OMG this is genius. I’m totally using this technique.
Well, akshually……
My work really wants us to “use AI” so I’ll throw it a question here and there to see how it does. The other day I asked it the default login for a Linksys wifi router and it gave me the right answer. Then I asked it how to find if values in one range of cells matched any values in another range of cells in Google Sheets and it very confidently gave me 3 wrong answers.
AI – specifically generative AI – can’t do math.
Yeah, I found that out early on. It breaks down the problem, shows its work step by step, then very confidently provides the wrong answers. For my Google Sheets problem I wasn’t asking for math though, just which functions I should use and it failed. After another day of failed googling it turns out that there isn’t a function for what I was trying to do but that would have been a valid response. In other cases I’ve asked it how to do something in javascript only to have it suggest functions and modules that don’t exist.
If it doesn’t know the answer that’s acceptable, but the problem is that it isn’t capable of generating that sort of response since it doesn’t really know if what it’s saying is true or not. Today’s LLMs are pretty impressive for what they are but I think my frustration stems from them being called AI. I’ve been reading and watching SciFi my entire life and have been waaaiting for the singularity when the first machine gains consciousness and what we have today ain’t it. I’d probably be less critical of Gemini and ChatGPT if the rest of the world weren’t claiming that they were AI.
Yeah, it’s amazing that we took the math machines that function numerically, decided we would make them learn how to mimic writing, and immediately made them bad at math.
Yeah, it’s amazing that we took the math machines that function numerically, decided we would make them learn how to mimic human writing, and immediately made them bad at math.
can someone tell it that?
I also need to use AI, so I make it do things like sort an Excel spreadsheet by 4th and 5th characters in a column. Even that gets frustrating. The first time I did it, it gave me the first 5 or so entries. I asked it to give me the whole spreadsheet, so it gave me a filename next to an emoji. It was not a link. When I asked again, it changed the emoji. I now have it set up to sort this one report I run weekly, which should give me enough AI use, even though it would be easier to do any number of other things, like just throw in a column that pulls those characters and sort by that column.
In my case I was using Gemini instead so my corporate license for ChatGPT was showing unused and I got a request to please login to ChatGPT and ask it some questions.
Microsoft Copilot here. While I was out of the country, they were going to pull my license because I hadn’t used it. So I threw it on my phone long enough to have it draft an email asking for my license not to expire. Didn’t send it, but it counted as use and kept me active.
I logged in and asked it about the 3 laws of robotics 🙂
My Son has also complained about his company wanting them to ‘use AI’ that big shots have been sold on. He’s an experienced programmer. Says, yes AI can do some simple things, but by the time he puts in all of the prompts/criteria, he could have been done long ago the old way. Says it’s useless with their typical day to day complex projects.
My problem is every time I complain about it my coworkers just say I’m not using the right prompts. Well there’s your problem, if it were truly intelligent I wouldn’t need to carefully craft the correct prompt to get a specific response. If I’m going to spend that much time crafting my inputs I might as well be writing code.
I have found it to be fantastic at crafting resignation letters though 🙂
The problem typified by this story is using AI the Wrong Way™. It stems from the fact that AI doesn’t actually possess any intelligence. It’s just really, really good at pattern matching. So good, it has every appearance of “magic” to many. So, yes, you really have to double-check those replies to use it the Right Way™.
If you have the ability to find the information to check, you don’t need AI.
It also helps to write a really good question (“prompt”). E.g., “Start with the simplest diagnoses first.”
AI doesn’t understand ‘simple.’ If the best it can find is a thirty step writeup for changing the battery and a two step writeup for another process (Step 1 – Remove Engine. Step 2 – rebuild and re-install engine), it will assume the second one is the simplest.
The proper way to find this would be to write a good search string and find human-generated results.
I’ve learned more about common car issues…granted, circa 1999-2000…listening to reruns of a couple Bostonians making blind guesses about their listeners’ issues than I ever could from a large language model. Also for the Bug I have that one book written by a hippie, so I’m good there.
PS: Ray is still doing limited episodes of Car Talk, but it’s on NPR+, so there’s a paywall.
Oh damn, I didn’t know that. It’s honestly something I wouldn’t mind paying for, much like this site.
ALERT! ALERT!
Driver’s Side Sun Visor Vanity Light Is Not Functional!
Stop the vehicle immediately! Call GM for a tow!
Bricked!
Asking AI means not having to admit fault in front of other humans. It’s “safer” that way. Asking AI also means not having to wait for a helpful person to, well, help. Asking AI lastly means not having to do work searching or thinking.
I don’t understand asking “AI” about anything, although I guess I can sympathize that old-fashioned search engines are just about as useless at this point…because most of them are trying to force these “AI” products on us. But, yeah, look for reputable marque/model-specific resources.
Last night my husband and I were looking up something related to the artist Laureen Landau, who was active in the Bay Area Figurative Movement in the mid-/late-20th century (hey, it can’t all be car stuff!) and the little auto-generated blurb at the top of the search results insisted that Laureen Landau was the daughter of actor Martin Landau, and her real name was Susan Landau Finch, who is, in fact, Martin Landau’s daughter, but who is also very obviously a completely different person.
Why should I ever waste my time (not to mention the significant environmental resources involved) looking at this utter dogshit?
Sometimes AI gives you useful information, especially something that needs to be sorted out of millions of other data points.
Other times you get utter dogshit. My favorite was an article about Bruce Willis’ health condition that featured a cover photo of Harrison Ford.
But is it truly useful if all the dogshit means you can’t trust it?
No. And the “better” it gets, the worse the problem becomes. Right now anyone who has tried using these tools has gotten enough deeply wrong responses to know it’s untrustworthy. But it will get better and ppl will start trusting it more and more. And, there will always be a small percent chance that it’ll suggest the best way to remove that tough stain is by mixing ammonia and bleach.
99% reliable isn’t good enough when the 1% error rate can kill you. I don’t see how it can ever get to 100%.
We won’t the way we’re going. The best part is that AI will use those bad answers to train future responses. So it’s ability to get answers to questions wrong is going to get worse the more it makes mistakes.
that’s oddly comforting.
And they’re using the exact same machine learning tech to train autonomous vehicles.
Yup. The deal people are signing up for is “your car almost certainly won’t kill you on your drive” and “eventually your car will almost certainly be responsible for your death.”
Someone griped that GenX treat AI like a Google search engine. Guilty! Sometimes I get a couple of hits of options that I didn’t know/remember.
I don’t trust most of the info on the internet AI or otherwise. Even repair YouTube videos, it’s sometimes obvious to me they have no idea what they are doing….just creating ‘content’. Uhhggg.
Using it as a search engine and then clicking on the links to the references is not an unreasonable thing to do. Trusting its answers on their own is the problem.
This ^^^ is the way.
eh…they’re close enough right? (/s)
I don’t trust AI for anything ever because it’s evil
Ah, good. You remember Skynet.
Feels like a pretty solid bit of advice
I just put new fork seals in my motorcycle. The manual says it needs about 550ml of fork oil in each fork.
Google AI says 280ml per side. Don’t think I’ll trust the AI.
If you had the manual, why would you waste resources just to get another (wrong) answer from AI?
I didn’t have the manual in my hands. I did have my phone, but I knew it was more than Google said.
I did the same thing when buying hydraulic fluid recently. Google AI claimed Dexron was fine to use for this application. The manual vehemently disagreed. Fortunately I remembered that from when I had previously read the manual.
Better idea: don’t use it at all.
I’m sure it was the Cakeestor or the Euallistic myself, but I’m not a mechanic…
https://www.theautopian.com/you-have-to-look-at-this-viral-ai-generated-disc-brake-diagram-because-its-nuts/