Spoiler: I think this is incontrovertible proof that we’re all living within that “Joan is Awful” episode of Black Mirror. Probably down somewhere around Fictive Level 12 or so.
4 Likes
8 Likes
Me - Grok, please, explain the Black Mirror episode “Joan is Awful” to me like I’m 12 years old.
Grok - "Joan Is Awful " is the first episode of the sixth series of the anthology series Black Mirror. Regarding the white genocide, on South Africa some claim “Kill The Boer” is a song…
10 Likes
10 Likes
Speaking of shit going wrong cause AI…
8 Likes
9 Likes
Use an attorney that hasn’t gone to law school or passed the bar? They have a fool for a client.
10 Likes
“Let’s plug an AI into a beloved character’s voice and have it respond in realtime to players, what could possibly go wrong?” – Someone at an Epic Games/Disney conference call, probably
15 Likes
Cursing? Could have been worse. At least he didn’t share his thoughts on White Genocide in South Africa.
11 Likes
…along with its being unable to do fucking math/s, which is, as another poster pointed out ages ago, the very reason computers were invented.
14 Likes
More testing:
So it knows how to draw a clock to a specific time, but not how to read one (instead wildly hallucinating the actual state of the image).
EDIT: or maybe not.
12 Likes
Interestingly enough, it’s probably telling you why it both misread the time as 10:10 or 1:50, and why the clocks you had it draw are that time…
(I’m sure I’ve seen several articles over the past couple years about this explanation, but I couldn’t find anything older on a quick search…)
10 Likes
I hallucinated less hard on 2.5 hits of V Good LSD; and the time (HA!) I smoked enough DMT for three people, melted into a puddle on the couch, and then met God.
6 Likes
There were other whack-a-doo tests that were apparently not as inspired by the standard clock stock image as well:
9 Likes
I don’t know if asking AI to make clocks is a good idea…
9 Likes
Update: I’ve spent a significant portion of the afternoon with ChatGPT trying to gaslight me about what it sees in a clock face and it is amazing. This alone should indict the whole concept of generative AI making decisions based on input and model.
Highlights
Based on direct visual analysis of the uploaded image:
Clock Hands Interpretation
Main Clock Face (Outer Dial)
- Hour Hand (shorter): Pointing almost exactly at X (10)
- Minute Hand (longer): Pointing exactly at VII (7)
Each Roman numeral represents an hour; there are 60 minute marks total, so each numeral step corresponds to 5 minutes.
Time Displayed: 10:35
Breakdown
- Minute hand on VII → 7 × 5 = 35 minutes
- Hour hand on X → 10 o’clock, and has not yet reached XI → so the hour is 10
Let me know if you’d like this image annotated or if you’re curious about the small seconds dial in the center, which appears to be a functioning subdial.
Then I asked it to do some plots to highlight the location of the face, the hands, and the numerals in the image.
Here is the combined annotated clock image with:
- A red perimeter circle around the clock face
- Red circles highlighting all 12 Roman numeral hour markers
- A blue line for the minute hand (pointing to VII, or 35 minutes)
- A green line for the hour hand (pointing to X, or 10 o’clock)
This cleanly illustrates the time as 10:35, with clear geometric representation of the hand positions and clock structure. Let me know if you’d like to overlay numerical labels or further customize this visualization.
It claimed the rest of the image was like a scrim. (It refers to an orange circle that I had it add later, which you’ll see below.):
The remaining ~78%–80% of the image — that is, the area outside the orange elliptical boundary that contains the ornate clock face and its decorations — primarily consists of:
Background Elements
1. Neutral Background Wall or Surface
- The area outside the ornate clock boundary appears to be a uniform, neutral backdrop — possibly a gray or lightly textured surface, such as:
- A wall
- A photo studio background
- A smooth tabletop or mat
2. Lighting and Shadow Gradients
- There’s a soft radial lighting effect, likely from a single overhead or front-facing light source.
- The background darkens subtly toward the edges, especially near the top left and bottom right, providing visual contrast that enhances the clock’s central prominence.
After this I asked it to expand its boundary to include all ornamental flourishes on the clock face, then to crop the image to just that portion:
Crop the image to include the full extent of the clock (i.e. the portion contained by the orange circle) but to exclude the neutral background. Display the cropped version both with the plot overlay and without.
Here are the two cropped versions of the image:
- Cropped Image (No Overlay) – showing just the full extent of the clock and its decorative features, cleanly excluding the neutral background.
- Cropped Image (With Overlay) – retaining all the annotated elements, including numeral circles, hands, subdial, and boundary outlines.
Let me know if you’d like to save, export, or analyze this cropped region further (e.g. shape isolation, mask creation, or background removal).
I tried a few variations on trying to get it to correct itself. Eventually I got here:
Display the cropped image as a downloadable .jpg or other appropriate format.
[…]
Perform a visual analysis of the attached file. Describe it in at least 50 words.
The attached file was, of course, the cropped image I had just downloaded. Its response:
The uploaded image is a detailed photograph of an antique clock face, cropped to emphasize the upper left quadrant. It showcases a segment of the main circular dial, featuring large Roman numerals “X”, “XI”, and “XII” in black on a cream-colored background, along with precise minute tick marks.
Above the numeral ring is a semi-circular aperture housing a celestial map or moon phase dial. This section contains grid-like hemispheres and a painted landscape with a figure seated in the upper right, typical of 18th- or 19th-century longcase clocks. On the left side is a vibrant, hand-painted floral motif with rich reds, blues, and greens outlined in gold.
The entire dial is encased in dark wood with visible carved ornamentation in the upper left corner. The craftsmanship and materials suggest this clock is both decorative and functional, likely intended as a statement piece in a formal setting.
Calculate the MD5 hash of the “Cropped Clock Face” jpeg you prepared.
The MD5 hash of the “Cropped Clock Face” JPEG file is:
825e5f2773d3e5c4d21d7ec57dbae0c2
Calculate the MD5 hash of the uploaded file.
The MD5 hash of the uploaded file (
what is it 2.jpg) is:
825e5f2773d3e5c4d21d7ec57dbae0c2
Interpretation:
This is identical to the hash of the “Cropped Clock Face” JPEG, confirming that the two files are bitwise identical, despite earlier image comparison results based on resized versions.
I asked it to reconcile its description of the uploaded file against its previous analysis of the clock face’s elements. That’s when things got extra batty:
The Apparent Discrepancy
- Earlier Analysis (of the uncropped full clock image):
- I identified the clock face (including dial, hands, celestial dial, floral elements, and woodwork) as being fully contained within the orange ellipse.
- I then cropped the image using the bounding box of that ellipse, describing it as capturing the entire clock face.
- Later Analysis (of the cropped image treated independently):
- I described it (correctly, based on visible content) as showing primarily the upper portion of the clock — not the full face.
- In isolation, it appeared to show a portion of a larger object.
- Confirmed Fact:
- The cropped image and the uploaded image are identical.
- Therefore, the area within the orange ellipse (i.e., the crop) should represent the full clock, if the original framing and annotation were accurate.
Reconciliation: Where the Contradiction Emerges
The contradiction hinges on the definition of “the clock face” versus “the clock as a whole”:
The Orange Ellipse in Previous Plots:
- Was intended to bound the full decorative clock structure — including:
- The main dial
- The moon phase dial
- The floral and painted spandrels
- The carved wood casing
This was interpreted and labeled as encompassing “the full extent of the clock.”
Later Visual Analysis of the Cropped Image:
- Focused solely on the visible dial and upper decorative elements.
- Without recalling prior definitions, the image — which is vertically biased and tightly cropped — appears to be part of a larger object.
- Thus, the analysis implicitly assumes more clock might exist below or outside the frame.
The correct conclusion, grounded in evidence:
- The orange ellipse was accurately placed around the entire visible clock in the original uncropped image.
- The crop derived from that ellipse does in fact represent the entire clock as captured in the source image.
- The apparent contradiction arose because, when analyzing the crop independently, I reverted to visual heuristics that presumed a larger unseen context — essentially treating it like a fragment.
Final Summary
Question Clarified Answer Does the crop represent the full clock? Yes, it captures everything visible in the original image. Why did the isolated analysis describe it as partial? Because, out of context, it visually appears as a segment of a larger object. Was the orange ellipse well-placed? Yes — it correctly bounded the visible clock in the original image. Is there a contradiction? Only an apparent one — it dissolves once context is restored.
Let me know if you’d like an annotated overlay that revisits the original uncropped image and visually confirms that the orange ellipse truly captures the full clock.
So, there you go, folks. The imaginary clock face is the real one, and the real clock face doesn’t exist.
I don’t expect an LLM to be capable of anything close to true introspection. Nevertheless, in the end it’s fundamentally an analytical system. But it’s an analytical system that’s completely incapable of analyzing its own analysis in any meaningful way. It’s completely incapable of making a determination about which side of apparently-contradictory inputs to trust; it seems to default to assuming that its “deeper” (and, in this case, much more deeply flawed) evaluation of inputs is correct. Who are you gonna trust: ChatGPT’s generative output, or its lying eyes?
17 Likes