Using AI for Bulk Game Input

[John Saunders]

I've just started trying out AI for BritBase work, namely the bulk input of game scores and formatting of crosstables from scanned images.

I've written a blog entry about it: https://johnchess.blogspot.com/2026/10/ ... tbase.html

I was wondering if anyone else here has experimented with it or knows of anything informative on the subject.

Note that I'm only interested in using AI to process scanned images of printed material, not scoresheets or manuscript chess scores. That's a whole different area and perhaps deserves its own thread. I'm thinking more of bulk input of games from tournament bulletins or similar.

[GrahamStuart]

As someone is very wary of AI I had not considered using it for this purpose, but will definitely look at duplicating your process for some of the Hampshire / WECU publications. I am shocked by how bad my IT usage is now, despite working with databases and programming for much of my working life.
I am hoping my scanner (Epson 4830 printer / scanner) is up to the task.

Basic question - Is ChatGPT free for this purpose?

[Christopher Kreuzer]

Nice blog, John. I have followed the rise of AI (which covers a multitude of sins) and would add that it can depend very much which model you use and how advanced it is and whether you are paying for it or not.

From horror stories I have read:

(1) If you are interested in preserving the methodology and so forth (exactly what prompts you used) and are drawn to storing everything in a paid-for or free account, do make sure to take backups as access to such accounts can be suspended (for sometimes spurious reasons) meaning lots can be lost if not backed up (backing up the output can be trivial but backing up the prompts is sometimes less easy).

(2) Be aware that some models will (despite claiming not to) retain and use your material for training purposes (improving future models). So be aware of that when using copyrighted materials and your own materials. I am sure you know this, but the saving in time is so much that many people just do it anyway (including getting AI to help with things they would not if they read the small print).

There are also some VLMs (Vision Language Models) as opposed to pure LLM (Large Language Models) that do amazing things now with recognising vertical text and other oddities of layout. I suspect they will handle chess figurines with ease.

Whether any AI models can convert descriptive to algebraic I don't know. Would be interested to know (maybe it has already been done). I would love it if they were able to recognise photographs of obscure historical chess players. Also not sure how many of the AI models have been trained on (i.e. ingested) chess magazines from across the world. Depends which have been scanned. I wonder how many Russian chess sources have been scanned to the level needed?

[Roger de Coverly]

I've just started trying out AI for BritBase work, namely the bulk input of game scores and formatting of crosstables from scanned images.

By 1998 you would be in the era where bulletins and cross tables were mostly prepared on computers. Reverse engineering the printed output back to its electronic form should on paper be an easier task than from earlier material. That wouldn't preclude input errors in the original transcription from player scoresheets. AI should be able to decode figurines, particularly if it has access to the fonts used.

[Ian Thompson]

There are also some VLMs (Vision Language Models) as opposed to pure LLM (Large Language Models) that do amazing things now with recognising vertical text and other oddities of layout. I suspect they will handle chess figurines with ease.

Although not quite the same thing, this study makes interesting reading - VLMs cannot reliably convert an image of a chessboard into an accurate chess diagram.

[John Saunders]

Thanks for all the helpful and informative responses.

Some matters arising...

Graeme's question: ChatGPT is not free after you have submitted three scans to interpret. Thereafter you would have to wait (until the next day?) before submitting more. The subscription is £20/month, though the first month is free.

Roger's point about 1990s tournament bulletin is right and reverse engineering seems to work quite well on crosstable data. I had fondly imagined that this meant that computer-generated crosstable data from the 1990s would be 100% accurate but ChatGPT actually detected an error in one such output. In the 1998 British Under-10 Championship table (on the 8th or 9th page of the bulletin) it shows the winner Lee Gold as scoring 6/7 but ChatGPT, in proof-reading the table, added up his scores to make 6½/7. AI seems to be right and the printed table wrong. I suppose it is possible that the '½' was accidentally edited out when the computer output was incorporated into the bulletin.

The file I have given it to work on at the moment is the 1998 British Major Open bulletin, with 205 games to convert to PGN. But this is taking a great deal longer than the junior championship - I set it going this morning (around 11am, I think) and have yet to see any output at 3.30pm this afternoon. That's still acceptable compared to manual inputting, which would take longer (10+ hours?), but the junior championships conversion was a good deal quicker.

The next thing I shall be testing with AI is reordering crosstables when mistakes occur that require the player order to be changed along with the individual pairings. This will also be useful for generating Guernsey tournament crosstables. Guernsey Chess Club has an excellent database of tournament results going back years but unfortunately, when the tables are shown in descending score order, the number associated with a particular player (more or less in an alphabetical surname order) is retained. I would like an easy way to change these to 1, 2, 3, along with the pairing numbers, etc, but I don't have the software or knowledge to accomplish this without a great deal of faff.

[Christopher Kreuzer]

The subscription is £20/month, though the first month is free.

Another warning (though you will be familiar with this from other services): becoming dependent on AI may backfire if they hike their prices. Though some say prices will fall rapidly for other reasons.

I set it going this morning (around 11am, I think) and have yet to see any output at 3.30pm this afternoon. That's still acceptable compared to manual inputting, which would take longer (10+ hours?), but the junior championships conversion was a good deal quicker.

On timings - I have no idea why some AI prompts take longer than others. You will probably be aware of the environmental aspects - there is a lonely iceberg or glacier somewhere melting due to the heat produced by the data/processing centres handling your request, John!

I would like an easy way to change these to 1, 2, 3, along with the pairing numbers, etc, but I don't have the software or knowledge to accomplish this without a great deal of faff.

Vibe coding:

https://en.wikipedia.org/wiki/Vibe_coding

It would in theory be better to develop the solution once rather than ask AI to do advanced (or even basic) proofreading and/or typesetting and/or other basic (but fiddly) tasks. But that ship has long sailed.

Cynical? Me? Surely not!

(I confess that I, for many years, instead of having proper bookmarks, Googled websites I would visit every day - thus contributing in my own way to extra [needless] searches, extra electrons, and excess energy consumption.)

I must go back to reading science fiction that has AI avatars doing everything for us...

[Matt Mackenzie]

Whether any AI models can convert descriptive to algebraic I don't know. Would be interested to know (maybe it has already been done)

This is indeed an interesting one - not least because everyone who has ever been through old chess books/magazines will know that descriptive was sometimes ambiguous even if not actually incorrect.

[Roger de Coverly]

This is indeed an interesting one - not least because everyone who has ever been through old chess books/magazines will know that descriptive was sometimes ambiguous even if not actually incorrect.

A google search tells me

How to Enable Descriptive Notation in ChessBase
You can switch your game notation style by navigating through the settings:
1. Open ChessBase and go to the main menu.
2. Select File (or Tools in older versions) → Options.
3. Click on the Notation tab.
4. Select Descriptive (labeled as P-Q4 Anglo-American descriptive notation).
5. Click Apply or OK to save the changes.
(Pro-Tip: You can also use the keyboard shortcut Ctrl + Alt + O to open the options menu directly)

[Gerard Killoran]

Whether any AI models can convert descriptive to algebraic I don't know. Would be interested to know (maybe it has already been done)

This is indeed an interesting one - not least because everyone who has ever been through old chess books/magazines will know that descriptive was sometimes ambiguous even if not actually incorrect.

Here:
viewtopic.php?f=27&t=15183&p=326329&hil ... ve#p326320

[Christopher Kreuzer]

Maybe worth trying again to see if things have improved? One prompt to include is something like "ensure the resulting game score follows the laws of chess". If using the free version, it is possible to combine prompts to get around the 3-prompts a day limit up to a point.

[John Saunders]

I don't see any good reason why AI shouldn't be able to interpret any form of chess notation as long as there are no ambiguities or mistakes. Even then, AI can iron these out via Q&A with the client or by consulting an analysis engine to choose between sensible and silly continuations. Indeed, I would have thought that AI could soon be trained to suggest repairs to faulty scoresheets, e.g. ones with a few missing or illegible moves, and accomplish the task better than we organic folk could hope to do.

Though it might be argued that descriptive and earlier notations were more prone to ambiguity and error than algebraic, that might be more to do with improvements in typesetting since the descriptive era. Plenty of ambiguities and errors were present in algebraic-notated games prior to the generation of printed game scores from databases. These days we have a different problem: bulk data entry is sometimes conducted by non-expert players who won't necessarily recognise an improbable move when they see one. Back in the day, bulletins were nearly always produced by strong players who could spot errors as they typed up the games. (That's not to say that some early regional newspaper chess columnists weren't weak players, because it is obvious that one or two of them were.) Larsen is alleged to have said something about databases being 'collections of bad games'. He might also have added 'which have been badly input'. Thankfully this is less true of games between elite players as these are now mostly processed via electronic boards which do a better job of getting scores right (as long as they are overseen by grandmasters of electronic board management such as Matthew Carr).

[GrahamStuart]

I have played around extracting games into pgn files from Westward Ho! and had some success. Probably not the best file to use, or my suggestion should be refined as was hit and miss with several of the games.

I had a look at something which would also be useful to me, extracting newspaper chess league tables to Excel (I guess I could also suggest for website tables, or something similar) as I still tend to use Excel.

Maybe this is very basic for others and decent OCR could already do, but I wished to check how this works. I used the following "Can you convert the following Southampton Chess League tables from 1966/1967 season to excel spreadsheet with formatted tables" and the output was below (or part of it).

As well as the worksheet for All Divisions and split by each it includes the Source at the top and an extra worksheet with Source Notes which included info such as "Team names and figures have been transcribed from the newspaper scan. Half-points are retained as the printed ½ character."

Examples of the best type of input for using AI would be useful to know, or what not to use.

[Alex Holowczak]

I don't see any good reason why AI shouldn't be able to interpret any form of chess notation as long as there are no ambiguities or mistakes. Even then, AI can iron these out via Q&A with the client or by consulting an analysis engine to choose between sensible and silly continuations. Indeed, I would have thought that AI could soon be trained to suggest repairs to faulty scoresheets, e.g. ones with a few missing or illegible moves, and accomplish the task better than we organic folk could hope to do.

There has been at least one company in recent times that has reached out to us with scoresheet-reader software. It was tested at a 4NCL weekend. Our conclusions were:
1. It took about 10 minutes to do the translation
2. It would flag ambiguous moves or notation, and require you to manually choose what the right solution was
3. Thus, the process from start to finish could be as much as 15 minutes per game

Most game-inputting humans can muster between 10-15 games per hour, depending on the length of the game and the surrounding environment being distraction free. Jack Rudd can usually muster 20+ games per hour, while taking occasional timeouts to play some online games, analyse current games in progress, solve some jigsaws, and no doubt many other things. For now, the optimal solution for notation interpretation is JR, rather than AI.

[John Saunders]

If you'd consulted the first post in the thread, Alex, (yes, I'm channelling Stewart Reuben again) you'd have read this...

Note that I'm only interested in using AI to process scanned images of printed material, not scoresheets or manuscript chess scores. That's a whole different area and perhaps deserves its own thread. I'm thinking more of bulk input of games from tournament bulletins or similar.

... in the hope that the thread would not stray too far from its main subject, namely the use of AI to process non-manuscript game scores from bulletins and magazines.

But thank you anyway: it was interesting to read the result of the 4NCL scoresheet-reading test which failed as everyone would have suspected. I think it will a long time before a scoresheet-reader can accomplish its task without doing what you describe in your point 2 - asking the human client to resolve ambiguous moves. In the case of a Korchnoi scoresheet (I have a couple which could be used for testing), I would suggest another 1,000 years.

Sad to say, as I have continued to feed bulletins into ChatGPT, I'm gradually coming to the conclusion that what Alex has written about manuscript input can apply to some non-manuscript material. The latest bulletin I have used for testing is the 1998 British Major Open bulletin, which, if you would like to look at it, appears in the BritBase library as a PDF. AI is really struggling with this and has slowed down to offering me one game at a time, sometimes with errors which I then have to resolve via Q&A. It's at the point where I could probably do the job quicker myself. It's strange because it did the 1998 British Junior bulletin - with exactly the same printing template and figurine notation - quickly and efficiently. Now it keeps coming back to me whinging about the notation and the fact that it doesn't use + for checks and x for captures. But neither did the junior bulletin. I'm wondering if a subsequent instruction from me has somehow rendered the interpretative process less effective or efficient compared with the first test (which may have been a case of beginner's luck).

That said, it is very good at one aspect of the job, namely to cross-reference the names above the game scores with what appears in the crosstable at the back of the bulletin, and use the full names available there, adding the published rating. That means that AI could produce a skeleton PGN file very easily, enabling the human operator to add the moves. I suppose a better idea would be to ask it to generate a full PGN skeleton file from the crosstable and work manually from that. It would at least alleviate some of the work.