hoakley May 15, 2026 Macs, Technology

Chinese whispers in PDF metadata

Chinese whispers is an old children’s game where everyone sits in a circle, and one child whispers into the ear of the next on their right a sentence like Send reinforcements, we’re going to advance. That child then whispers the message they heard to the child on their right, until it reaches the one who started it, who says out loud what they heard, classically Send three-and-fourpence, we’re going to a dance, as a demonstration of how messages can so easily become corrupted. What this has to do with China remains one of childhood’s mysteries. I should also explain that three-and-fourpence was idiomatic British English in the days before our currency was ‘decimalised’, and meant three shillings and four (old) pence, about 17 (new) pence, sufficient at one time to enjoy a good night out.

In this article I’m going to do much the same with metadata for a PDF document, tracing what gets indexed by Spotlight, so becoming discoverable by search, and what is displayed in the Finder. This relies on several of my utilities, most of which are available from this page.

Source PDF

I prepared a completely unrelated PDF using my favourite PDF editor, PDF Expert, by adding metadata to be saved in the file’s data. As you might expect, there are several ways that could be stored in the PDF format, including XMP metadata, but in this case for simplicity they were added in the document information dictionary.

I inspected that in a source view in Podofyllin, which found the following fields:
/Author (Author name in pdf) /Creator (Pages) /Keywords (keyword1 pdf) /Subject (Subject in pdf) /Title (0PDFtest1accessdefault)

When rendered in macOS, those are ‘flattened’ by its Quartz PDF engine, to
/Author (Author name in pdf) /Creator (Pages) /Keywords (keyword1 pdf) /AAPL:Keywords [(keyword1 pdf)] /Subject (Subject in pdf) /Title (0PDFtest1accessdefault)
Note the copying of keywords into a new attribute AAPL:Keywords.

Extended attributes

I then added seven extended attributes using Metamer, with names such as com.apple.metadata:kMDItemAuthors, as shown below in xattred.

Spotlight import

I then inspected the file in SpotTest’s new Drop Window, which reported the following attributes found by mdimport:
":EA:kMDItemAuthors" = "author in xattr"; ":EA:kMDItemComment" = "xattr comment"; ":EA:kMDItemCreator" = "xattr creator"; ":EA:kMDItemDescription" = "xattr description"; ":EA:kMDItemKeywords" = "keyword1,xattr"; ":EA:kMDItemSubject" = "xattr subject"; ":EA:kMDItemTitle" = "xattr title";
all from the extended attributes, while those derived from the PDF data were
kMDItemAuthors = (Pages); kMDItemCreator = Pages; kMDItemDescription = "Subject in pdf"; kMDItemKeywords = ("keyword1 pdf"); kMDItemTitle = 0PDFtest1accessdefault;

Those attributes have already changed, with PDF Subject becoming kMDItemDescription, Creator being duplicated into kMDItemAuthors, and the loss of PDF Author.

Spotlight indexes

Attributes reported by mdls changed again to
kMDItemAuthors = (Pages) kMDItemComment = "xattr comment" kMDItemCreator = "Pages" kMDItemDescription = "Subject in pdf" kMDItemKeywords = ("keyword1,xattr") kMDItemSubject = "xattr subject" kMDItemTitle = "0PDFtest1accessdefault"

This has lost the xattr attributes kMDItemAuthors, kMDItemCreator, kMDItemDescription and kMDItemTitle, and the PDF kMDItemKeywords. That list of 7 attributes should then be searchable using Spotlight.

The Finder

The final step was to discover which of those could be displayed in the Finder, either in its Get Info dialog, or in the Preview panel of a Finder window.

Only 5 of those attributes survived in the Finder, and were given as
Authors: Pages Content Creator: Pages Description: Subject in pdf Keywords: keyword1,xattr Title: 0PDFtest1accessdefault

Of those, 4 are taken from the metadata in the PDF file, and only the Keywords were taken from its extended attribute. The attribute named as Authors contains a duplicate of what had originally been in the PDF Creator field, but neither of the PDF Author or xattr kMDItemAuthors fields. Those paths are traced in the diagram below.

Conclusions

Of the total of 12 distinct metadata attributes added in the PDF data and extended attributes, only 6 different items were indexed by Spotlight, and 4 were displayed in the Finder (allowing for the duplication of Authors and Content Creator).

Before relying on metadata for search and access in the Finder, it’s essential to verify that the attributes you intend using are successfully indexed and displayed. Choose the wrong attributes and you’ll never find anything.

18Comments

Add yours

1

Enzo Vincenzo on May 15, 2026 at 7:09 am
Reply

Thank you so much for opening up all these new perspectives on digital computing for us, Howard!
Everyone thinks they know that nothing is more precise than a computer, and this belief — deeply rooted since the dawn of computing — is beginning to take hold in every field.
Yet I realise that your masterful explanation today can be applied everywhere, including medicine and surgery.
This makes the use of computers ‘operator-dependent’, and if there isn’t a vigilant and professional mind managing the entire process, the implications and final outcomes can be exactly as you described in your clear and perfect anecdote.
But not only because of final naive or simple-minded operators… but because – unfortunately – even at the very start of the design of operating systems and software, there are simple-minded operators who rely on chance and lack the great intuitive and supervisory skills that – for example – Steve Jobs, Johnny Ive and you possess.

LikeLiked by 2 people
- 2
  
  hoakley on May 15, 2026 at 8:05 am
  Reply
  
  Thank you, Enzo.
  Howard.
  
  LikeLiked by 1 person
3

joethewalrus on May 15, 2026 at 8:26 am
Reply

That game is known as Telephone in the USA, I believe coast-to-coast, and across generations.
There is also a (made in China??) commercially available spin on it where it is combined with drawing, named Telestrations: Player one receives the secret word, and draws it. Player two views the drawing and writes what they think they are looking at. Player three reads #2’s word (without seeing the drawing) and draws. Player four writes what they think they see. Player five sees 4’s drawing without seeing any of the previous drawings or words…you get the point.
Either it’s great fun, or I’m easily amused.

Taking time to meticulously catalog files with metadata only to have that messed and not searchable — now that is not fun.

LikeLiked by 3 people
- 4
  
  Simon on May 15, 2026 at 1:31 pm
  Reply
  
  Just FYI, this used to be called Chinese Whispers here in the US too. At least here in California in the 70s. The reason it’s referred to as Telephone these days is that Chinese Whispers at some point by some self-declared moral authority was declared racist. I wonder when Chinatown in San Francisco will be renamed — after all, wouldn’t want to be racist while we wage war on Iran. 🤦‍♂️😂
  
  LikeLiked by 3 people
  - 5
    
    hoakley on May 15, 2026 at 10:02 pm
    Reply
    
    Thank you. I’ve tried to discover the origin of the Chinese part of the name, but no one seems to know. So it can hardly be claimed to be discriminatory.
    Howard.
    
    LikeLike
- 6
  
  Duncan on May 15, 2026 at 2:53 pm
  Reply
  
  “Telestrations”
  
  Believe it or not, I may have invented that game back in 2003 (or at least arrived at it independently). Not the commercial version sold today, but as a simple bar game for a social group where we’d go to a different bar each week and partake in silly activities that often worked on the bar’s theme. In this case, being based on the game ‘Telephone’, we played in a San Francisco bar called ‘The Phone Booth’ in the Mission, which is located directly across the street from a local central office and where telecom workers would go after their shift.
  
  To play we’d alternate text and illustrations using standard paper which was folded over after each entry so that all the previous ones were hidden except for the last phrase or drawing, which served as the starting point for the next player.
  
  After a few drinks the train of thought would often get completely derailed.
  
  LikeLiked by 3 people
  - 7
    
    hoakley on May 15, 2026 at 10:07 pm
    Reply
    
    You should have claimed IP rights on it, Duncan. It looks like the commercial game wasn’t published until about 2010. Think of all the money you could have made – in the UK it sells for around £25. Even 10% of that would have been nice.
    Howard.
    
    LikeLike
  - 8
    
    Duncan on May 16, 2026 at 12:01 am
    Reply
    
    I’m surprised that someone was able to make a physical product from the idea and find a buying audience. We did it all on sheets of paper and simply folded them over.
    
    I guess you could sell a packaged version of Tic-Tac-Toe and then charge a subscription for new pads to today’s public. 🤷🏻‍♀️
    
    (Many of the activities we came up with could be performed using simple office supplies, which we encouraged people to liberate from their places of work. One involved seeing how many rubber bands you could stretch onto a beer bottle. It’s not something I’m particularly proud of, but it kept us off the streets for a few hours.)
    
    LikeLiked by 2 people
  - 9
    
    Tristan Hubsch on May 16, 2026 at 12:32 am
    Reply
    
    As Howard says, had you claimed IP rights…… In turn, various variations of that game were “old hat” where I grew up, in Southeastern Europe, taught to my cousins and myself by our grandmother. All we needed was a strip of paper (to be folded over, just as you say) and a pencil. And, ‘fore you ask: in my grandmother’s version, the libations were 100% alcohol free—but I get the twist added.
    
    LikeLiked by 1 person
- 10
  
  hoakley on May 15, 2026 at 9:58 pm
  Reply
  
  Thank you, Joe. It’s definitely not fun when your Authors attribute gets overwritten all the time by the Creator – that had long puzzled me.
  Howard.
  
  LikeLiked by 1 person
11

Tristan Hubsch on May 15, 2026 at 1:06 pm
Reply

Hilarious! Back where I come from, in Southeast Europe, a country that …ahem, dissolved in the 1990s, the game was called “broken telephones,” emphasis on the broken technology. (“Great minds think alike?” 😇 And, why did the Brits have to scapegoat the Chinese?🤷) But, the inability (of this increasingly more opaque black box of a bitten fruity logogram) to find something you know is “in there” pales in comparison to the not too distant possibility of that same increasingly more opaque black box of a bitten fruity logogram starting to hAIllucinAIte about what’s “in there.”😨

LikeLiked by 3 people
- 12
  
  hoakley on May 15, 2026 at 10:00 pm
  Reply
  
  Apparently, there is an AI (or, rather, LLM) equivalent of Chinese whispers, in which LLMs mutate the output of other LLMs until they produce something that makes a ‘hallucination’ look sane.
  Howard.
  
  LikeLiked by 1 person
  - 13
    
    Tristan Hubsch on May 15, 2026 at 11:44 pm
    Reply
    
    Indeed! We’ve disagreed in the past on whether LLMs “think” or not, but I’m glad you see that AI-to-AI “feature.” In some grand (millennial) sense, not all that different from toddlers babbling to each other until the gr’ups in the room give them candy.
    
    LikeLiked by 1 person
    - 14
      
      hoakley on May 16, 2026 at 8:57 pm
      
      I don’t recall any disagreement. LLMs can’t think, and can’t reason. All they can do is manipulate language, as that’s what their models are.
      Chinese whispers in AI/LLMs is more than just their slop echo chamber. It’s where you can trace any initial human statement and follow it through different model iterations as it changes progressively until it emerges as total nonsense. A recent example is the number of T1 ‘Trump’ mobile phones pre-ordered, where an initial speculative report was published in the The Times of India based on an unsubstantiated social media post, and now all the AIs are claiming the figure is 600,000, according to many different sources that when you check them never actually quoted a number.
      Howard.
      
      LikeLiked by 1 person
    - 15
      
      Tristan Hubsch on May 16, 2026 at 9:38 pm
      
      @hoakley, 8:57 pm: far be it from me to disagree with my host (says he with an ear-to-ear bright-eyed grin)… But, turning that into a definition of what “thinking” isn’t, the number of actually thinking carbon-based, flesh-and-blood bipeds just plummeted.
      
      LikeLiked by 1 person
  - 16
    
    Duncan on May 16, 2026 at 12:10 am
    Reply
    
    “Apparently, there is an AI (or, rather, LLM) equivalent of Chinese whispers, in which LLMs mutate the output of other LLMs…”
    
    Isn’t that going on all the time now with today’s web-scrapers? With more and more web sites filling up with AI slop (some appear to be generated on-the-fly just to “answer” a query, and then fill the page with ads) I’d think the Garbage-In, Garbage-Out treadmill is becoming neck-deep in bull manure. Meanwhile, all the older human-generated web pages seem to be getting harder and harder to find.
    
    (Or do the better AI training models simply limit their scraping to whatever’s available on Wikipedia and Internet Archive to avoid this vicious feedback spiral?)
    
    LikeLiked by 2 people
    - 17
      
      hoakley on May 16, 2026 at 9:00 pm
      
      As I’ve written to Tristan, the Chinese whispers in AI/LLMs is more specific. The example of how many T1 ‘Trump’ phones have been preordered has been fairly well investigated, and can be traced from a human origin to many different LLMs now making the same claim and citing bogus sources.
      Howard.
      
      LikeLiked by 1 person
    - 18
      
      Tristan Hubsch on May 16, 2026 at 9:53 pm
      
      @hoakley, 9:00 pm: Isn’t that how rumor mills work? And jokes? And urban legends? And tall tales? And………😇
      
      LikeLiked by 1 person