Joseph R. Farah: Google's AI overview got cosmic rays all wrong

See also this Astraveo video:

A few days ago, I uploaded a video explaining that the Mars jellyfish anomaly was very likely to be a cosmic ray artifact due to the presence of a row of clipped pixels with no antialiasing, a sure sign of an imaging artifact--real objects, even alien spacecraft, don't produce that. I explained the mechanism for how a cosmic ray can create an artifact on an image sensor, and explicitly showed how if it impacts during a calibration frame, the subtraction can produce a black artifact on the final image. I then provided an independent confirmation of this in the form of the Mars Science Laboratory camera engineering lead explicitly describing this exact process producing a black artifact in an image from a few years ago.

Despite this, I saw a surprising number of comments that confidently and incorrectly insisted that cosmic rays can only produce WHITE artifacts, never black. Here's a few:

"As I understand it, and if the supposed experts are really experts who know what they are talking about, cosmic rays are never black."
"You may be an astrophysicist but you definitely aren't a photographer. Cosmic ray hitting a camera sensor would not leave a black spot of multiple pixels. It would excite the sensor leaving a bright white spot and it would be permanent."
"I have never seen a BLACK cosmic ray artifact. It is always WHITE."

I responded to each of these and more, reiterating the dark frame explanation from the video. But it did get me wondering--why are there so many people repeating this incorrect thing so confidently? I scratched my head about it for a while, but I didn't get a clue until a friend texted me something alarming--he had Googled a question about cosmic rays and Google gave him an answer that conflicted with my own!

I was frankly shocked by this. Cosmic rays creating black artifacts due to subtraction of calibration images is a very well-known and understood effect. It was not only confirmed by the MSL lead camera engineer but its even reproduced by amateur astronomers, like this thread on Cloudy Nights where AlphaTriplePlus says, "If you use a master dark for calibration that has a cosmic ray hit, the calibrated lights will probably exhibit a wiggly black streak as the hot pixel streak in the master dark is subtracted from every light in the live stack. Happened to me last night!"

So why did Google's AI overview get it wrong? I decided to investigate a little. I tried my friend's query and was able to reproduce what he saw. Asking "do cosmic rays cause dark artifacts?" caused Google's AI overview to confidently say, "No, cosmic rays do not cause dark artifacts", even bolding and highlighting the wrong answer! I even tried asking a few different ways--asking specifically whether cosmic rays can cause dark artifacts when impacting on a bias frame (wrong answer still!) or even going as specific as asking "if a cosmic rays strikes during calibration image acquisition, can the final image have a black artifact?" (wrong answer again!)

But after a few prompts, I did notice something interesting. Even though it always highlights and bolds the WRONG answer, sometimes, buried down deep in the answer, is the RIGHT answer, which it for some reason didn't feel was important enough to point out. For example, in the very first prompt--"can cosmic rays cause dark artifacts"--while it highlights and bolds its "NO" answer, it has a section below titled "What causes dark artifacts?" where it explicitly says that "subtracting a master dark frame can create dark "holes" or negative spots if a hot pixel was not active during the actual exposure." That's literally the explanation I gave in the original video, where the hot pixel is caused by a cosmic ray, and the correct answer to the question! But its buried down in the answer for some reason.

However, seeing this right answer pop up in the response, just buried, gave me an idea about how to get it to give the right answer. If you ask it specifically, "what artifact does a cosmic ray in a dark frame create after dark subtraction?" it will give the right answer--a negative pixel or a black hole/spot in the resulting calibrated image. So if you used AI to validate your comment by just googling and reading the first line in the Google AI overview, I hope this clears that up--its just a logical error from Google's AI overview. And I hope this makes it clear that AI makes mistakes very frequently and it is always better to try and dig up the answer yourself, or talk to an expert vs blindly accepting an AI's answer at face value, especially if the question is in a field that isn't your own. You have to be careful about interpreting responses from any AI, as exemplified by this problem we and you may have encountered trying to use AI to understand a subject as complex as image analysis of an optical camera in a very harsh planetary environment.

For the record, I tested this problem on ChatGPT and Claude as well, on the weakest models to give AI overview a fair shake. ChatGPT made a similar error but put the right answer right near the top, so I would consider that a success. If you don't read one sentence down that's kind of your fault. Claude just got the right answer out of the box. This seems to be an issue with Gemini relying too much on the fact that "cosmic rays" are most often associated with "bright artifact", but its a perfect example of how LLMs which may assume the correct answer is the one most repeated in the training data can be deeply flawed if not paired with careful reasoning capabilities and primary sources.

This also leads into something else I want to discuss, which was another theme in the comments. If AI can't be trusted--and to be clear, the answer is, it cannot--what can you trust? More importantly, I realize that I'm essentially pitting myself against Google here, which has vastly more knowledge than I have. Why should you trust me over Google Gemini, a powerful frontier AI model?

I saw this question raised MANY times in the comments. One common theme was the idea that astrophysicists don't know anything about imaging, or insinuating that astrophysics and imaging/sensing/photography are somehow disjoint mutually exclusive fields. If you think this, I can of course understand why you might be hesitant to trust my explanation when Google is very vehemently telling you the opposite. So I wanted to dive into this a bit.

First, you don't need to trust me OR Google at all. All the analysis in the previous video is analysis you can do yourselves. You can check the pixel values and see they are either (0, 0, 0) or (1, 1, 1) in the region of the artifact--that's classic clipping. You can see it goes from pitch black to sky-white with zero antialiasing, a sure sign of a sensor glitch. And if you think that I, as an astrophysicist, am not qualified to weigh in here, I showed an email from the lead engineer on the MSL camera explaining away a similar black artifact as a "textbook cosmic ray". And I assure you, the lead engineer for the MSL camera is indeed an expert on imaging.

But I would also like to speak up in defense a bit of astrophysicists everywhere. Reading comments like this, I feel like there is a severe misconception about what astrophysicists do and where our expertise lies. The vast majority of astrophysicists are observational or experimental astrophysicists, and while many of us including myself spend a lot of time working with theory and with theorists, most of our work is on data captured via IMAGING. How do you think we study the stars? We can't exactly go to them. We mostly just take pictures--sometimes fancy pictures, like what a spectrograph does, or pictures in non-optical wavelengths, like what radio telescopes do--but they are just pictures, and we analyze them and extract data from them. Even if you don't specialize in imaging, you almost certainly know fundamental principles of imaging and the details of the instrument you work on specifically, and these fundamental principles and many instrument details are shared across almost all imaging instruments. Many of us work at observatories directly and have our hands on the raw data flowing in from the telescopes. Some of us, in this case not me, but like my Astraveo colleague and good friend Dr. David James, have built the telescopes and cameras on these mountaintops with our bare hands. While I mostly work with observational data and how we can use that data to test our various theories, David literally built this enormous telescope on top of Mauna Kea in 2010! Here is David to say a few words about his thoughts on this anomaly.

I am also happy to point to my own credentials here, which I think serves as a good example of how even an astrophysicist who does not work directly on the instrument hardware must learn the details and nuances of imaging in order to do any level of data processing. In my PhD thesis alone there are countless examples of this--almost every paper I ever wrote features some discussion of handling image artifacts, and almost every paper I wrote features discussion of image artifacts specific to cosmic rays! Here are a few instances.

In any paper I wrote on imaging supernovae, we had to analyze the data with pipelines that explicitly include cosmic ray flagging and rejection. This includes recent papers from 2025 and 2026, which explicitly discuss this pipeline, and even have specific discussions of a feature in one of our spectra caused by a large, narrow cosmic-ray artifact, which is something we had to work around. In another paper, we use similar cosmic ray detection and flagging in photometry collected from many different instruments across UV optical and near-infrared wavelengths.

In one paper, we found a very early data point from ATLAS, and we weren't sure whether it was imaging noise or actually data from the supernova. It matched our model but that could have been a coincidence, and the statistical significance was only 2sigma. Using extensive analysis of the science imaging as well as control curves and the processing pipeline, we were able to moderately increase the significance to almost 3sigma but ultimately left the result inconclusive. This is a good example of the high standards we have for considering whether something is data or noise--even though we found that there was less than 1% chance that the data point was purely noise, we still determined it to be inconclusive, because with millions of data points, a less than 1 in a 100 chance isn't as unlikely as it sounds. So when you see us being cautious about artifacts or looking carefully at data, we're not just kowtowing to the "establishment", "accepting NASA's explanation", or "bending to peer pressure". We're trying to be rigorous and hold things to a high standard of discovery, because that's the only way actual science gets done.

My point with this is to say that even I, an astrophysicist who doesn't work directly on instrumentation, has had to become deeply involved and experienced with imaging algorithms and artifacts just to get my work done. So the separation that's being drawn between astrophysicists and imaging is not a very good one because that line is VERY blurry. So if you're wondering who to trust and your choices are between astrophysicists or AI--particularly the Google AI overview--I hope this gives a bit more info for you to make your decision.

Previously on this blog:

Did NASA find a jellyfish on Mars?